Graphical User Interface (GUI) agents powered by vision-language models hold promise for automating real-world mobile tasks. However, progress is limited by the lack of high-coverage, long-horizon interaction trajectories collected from element-rich and rapidly evolving apps. Existing pipelines often rely on costly human demonstrations or on-policy framework, which tends to over-sample common flows while missing rare transitions and complex multi-step procedures. To address this problem, we propose SEE, a two-stage data synthesis framework consisting of (i) an efficient exploration stage that builds an explicit UI transition graph over screens and elements, and (ii) a graph-based synthesis stage that composes diverse multi-step trajectories via planning and controlled sampling. This design yields reproducible and explainable data generation, while explicitly preventing spurious cycles and enabling long-horizon composition. Across multiple real-world apps, SEE produces trajectories with an average length of 14.8 steps while avoiding spurious loops, and agents fine-tuned on SEE achieve improved task success and generalization to unseen screens. We will publicly release our synthesis code and dataset.
展开完整摘要收起摘要↓
Graphical User Interface (GUI) agents powered by vision-language models hold promise for automating real-world mobile tasks. However, progress is limited by the lack of high-coverage, long-horizon interaction trajectories collected from element-rich and rapidly evolving apps. Existing pipelines often rely on costly human demonstrations or on-policy framework, which tends to over-sample common flows while missing rare transitions and complex multi-step procedures. To address this problem, we propose SEE, a two-stage data synthesis framework consisting of (i) an efficient exploration stage that builds an explicit UI transition graph over screens and elements, and (ii) a graph-based synthesis stage that composes diverse multi-step trajectories via planning and controlled sampling. This design yields reproducible and explainable data generation, while explicitly preventing spurious cycles and enabling long-horizon composition. Across multiple real-world apps, SEE produces trajectories with an average length of 14.8 steps while avoiding spurious loops, and agents fine-tuned on SEE achieve improved task success and generalization to unseen screens. We will publicly release our synthesis code and dataset.
LLM-based agents process external content, exposing them to prompt injection and multi-turn manipulation. Most safety benchmarks evaluate defenders against fixed attack pools collected before evaluation, single-turn or multi-turn. We present a 21-scenario benchmark for adaptive multi-round attacks against memoryless LLM defenders: an autonomous LLM attacker observes prior defender responses and pivots across rounds, while each defender response is evaluated as a fresh interaction. Holding the 21 scenarios, attackers, defenders, and structured-output scoring fixed, restricting scoring to the first attacker turn yields $0$-$1%$ attack success rate (ASR); allowing 15 rounds of adaptive attack yields $5.4$-$14.0%$. Pooling three frontier attacker LLMs uncovers $1.4$-$2.2\times$ as many unique successful attacks as the best single attacker, and the generated attacks have low cosine similarity ($0.02$-$0.14$) to attacks in existing benchmarks. Claude Opus 4.6 and GPT-5.4 are tied in aggregate ($5.4%$ each; overlapping $95%$ CIs), but their weaknesses differ sharply: on one scenario Opus reaches $60%$ ASR ($95%$ CI $36$--$80%$) while GPT-5.4 and Gemini each stay at $7%$ (CI $1$-$30%$; the gap is preserved in a higher-$N$ replication). $13$ of $21$ scenarios distinguish at least one defender pair, yet rankings disagree across scenarios (Kendall's $W = 0.19$). We release the benchmark — 21 evaluation scenarios, 10 public development scenarios, the orchestrator, baseline harnesses, and a multi-attacker CLI — plus 945 transcripts from the 3$\times$3 frontier matrix, an attack-replay dataset, and 18{,}422 gpt-oss-20b battles from an open competition's final scoring rounds.
展开完整摘要收起摘要↓
LLM-based agents process external content, exposing them to prompt injection and multi-turn manipulation. Most safety benchmarks evaluate defenders against fixed attack pools collected before evaluation, single-turn or multi-turn. We present a 21-scenario benchmark for adaptive multi-round attacks against memoryless LLM defenders: an autonomous LLM attacker observes prior defender responses and pivots across rounds, while each defender response is evaluated as a fresh interaction. Holding the 21 scenarios, attackers, defenders, and structured-output scoring fixed, restricting scoring to the first attacker turn yields $0$-$1%$ attack success rate (ASR); allowing 15 rounds of adaptive attack yields $5.4$-$14.0%$. Pooling three frontier attacker LLMs uncovers $1.4$-$2.2\times$ as many unique successful attacks as the best single attacker, and the generated attacks have low cosine similarity ($0.02$-$0.14$) to attacks in existing benchmarks. Claude Opus 4.6 and GPT-5.4 are tied in aggregate ($5.4%$ each; overlapping $95%$ CIs), but their weaknesses differ sharply: on one scenario Opus reaches $60%$ ASR ($95%$ CI $36$--$80%$) while GPT-5.4 and Gemini each stay at $7%$ (CI $1$-$30%$; the gap is preserved in a higher-$N$ replication). $13$ of $21$ scenarios distinguish at least one defender pair, yet rankings disagree across scenarios (Kendall's $W = 0.19$). We release the benchmark — 21 evaluation scenarios, 10 public development scenarios, the orchestrator, baseline harnesses, and a multi-attacker CLI — plus 945 transcripts from the 3$\times$3 frontier matrix, an attack-replay dataset, and 18{,}422 gpt-oss-20b battles from an open competition's final scoring rounds.
Multi-agent policy optimization, exemplified by PPO-based methods, is a key branch of cooperative Multi-Agent Reinforcement Learning (MARL). A central design question is how many neighboring agents\footnote{In this paper, "neighbors" refer not only to physical proximity but also to agents whose actions influence one another.} to aggregate in order to effectively utilize global information for cooperation. This decision must be made along two dimensions: in the advantage (which agents' rewards contribute to the credit signal) and in the ratio (which agents' likelihood ratios form the clipped importance weight). Existing methods occupy scattered, underexplored points on these two axes: IPPO treats both separately; MAPPO pairs a team-level advantage with per-agent ratios; HAPPO employs sequential ratios with per-agent advantages; and single-agent reductions operating on factorized joint policies aggregate both into fully joint products. We formalize these two design choices as support matrices $\SA$ and $\SR$, and prove a canonical structure: the expected multi-agent policy optimization objective depends on the pair $(\SA,\SR)$ only through their matrix product $\tS=\SR\SA$. This yields two key consequences: (i) Redundancy: the two support matrices are interchangeable with respect to the signal, meaning neither aggregation pattern is inherently superior.(ii) Variance Ordering: the advantage aggregates rewards as a sum (additive variance with an interior bias-variance optimum at the coupling neighborhood), whereas the ratio aggregates likelihood ratios as a product (multiplicative variance that grows exponentially with support size, with no accompanying bias reduction). The resulting design principle is unambiguous: aggregate neighbors in the advantage, sized to the coupling neighborhood, and keep the ratio per-agent.
展开完整摘要收起摘要↓
Multi-agent policy optimization, exemplified by PPO-based methods, is a key branch of cooperative Multi-Agent Reinforcement Learning (MARL). A central design question is how many neighboring agents\footnote{In this paper, "neighbors" refer not only to physical proximity but also to agents whose actions influence one another.} to aggregate in order to effectively utilize global information for cooperation. This decision must be made along two dimensions: in the advantage (which agents' rewards contribute to the credit signal) and in the ratio (which agents' likelihood ratios form the clipped importance weight). Existing methods occupy scattered, underexplored points on these two axes: IPPO treats both separately; MAPPO pairs a team-level advantage with per-agent ratios; HAPPO employs sequential ratios with per-agent advantages; and single-agent reductions operating on factorized joint policies aggregate both into fully joint products. We formalize these two design choices as support matrices $\SA$ and $\SR$, and prove a canonical structure: the expected multi-agent policy optimization objective depends on the pair $(\SA,\SR)$ only through their matrix product $\tS=\SR\SA$. This yields two key consequences: (i) Redundancy: the two support matrices are interchangeable with respect to the signal, meaning neither aggregation pattern is inherently superior.(ii) Variance Ordering: the advantage aggregates rewards as a sum (additive variance with an interior bias-variance optimum at the coupling neighborhood), whereas the ratio aggregates likelihood ratios as a product (multiplicative variance that grows exponentially with support size, with no accompanying bias reduction). The resulting design principle is unambiguous: aggregate neighbors in the advantage, sized to the coupling neighborhood, and keep the ratio per-agent.
Computer Use Agents (CUAs) can autonomously execute complex, multi-step tasks within GUIs, enhancing efficiency through parallel multitasking. However, our formative studies with CUA experts and GenAI users indicated that current feedback is primarily text-based, requiring sustained attention to monitor progress and offering limited visibility to trace past GUI interactions. Based on the findings, we developed a prototype system, Sidekick, for communicating CUAs' status with multimodal feedback across different stages of interaction: (i) When CUAs run in the background, Sidekick signals its execution state through ambient cues. (ii) Upon resuming interaction with CUAs, Sidekick provides multimodal summaries of completed actions to support rapid context resumption. (iii) When CUAs operate in the foreground, Sidekick enhances transparency by verbalizing and visualizing the agent's reasoning. A study with 30 participants demonstrated that Sidekick significantly improved multitasking performance with CUAs compared to baseline systems that presented textual feedback either in a typical chat or in an ambient display. Sidekick supported progress awareness, and error and action traceability more effectively. Finally, we demonstrate the promise of Sidekick through several example applications, and discuss implications for long-horizon human-agent collaboration.
展开完整摘要收起摘要↓
Computer Use Agents (CUAs) can autonomously execute complex, multi-step tasks within GUIs, enhancing efficiency through parallel multitasking. However, our formative studies with CUA experts and GenAI users indicated that current feedback is primarily text-based, requiring sustained attention to monitor progress and offering limited visibility to trace past GUI interactions. Based on the findings, we developed a prototype system, Sidekick, for communicating CUAs' status with multimodal feedback across different stages of interaction: (i) When CUAs run in the background, Sidekick signals its execution state through ambient cues. (ii) Upon resuming interaction with CUAs, Sidekick provides multimodal summaries of completed actions to support rapid context resumption. (iii) When CUAs operate in the foreground, Sidekick enhances transparency by verbalizing and visualizing the agent's reasoning. A study with 30 participants demonstrated that Sidekick significantly improved multitasking performance with CUAs compared to baseline systems that presented textual feedback either in a typical chat or in an ambient display. Sidekick supported progress awareness, and error and action traceability more effectively. Finally, we demonstrate the promise of Sidekick through several example applications, and discuss implications for long-horizon human-agent collaboration.
Verify-repair loops are a standard means for large language model (LLM) agents to correct faulty plans in code generation, mathematical reasoning, and tool use. When both the verifier and the repairer are noisy, repair can damage already-correct plans, and reported acceptance keeps rising while true validity falls, so existing methods lack a principled basis for deciding when repair should stop. We propose VRR-Stop, a robust stopping framework for noisy verify-repair-repeat (VRR) loops. A four-parameter noise model separates verifier false acceptance and false rejection from the repair and damage behavior of the repairer. Belief filtering turns repeated verification votes into an estimate of committed validity, and the loop commits or repairs according to the sign of the true marginal gain, which requires only sign identifiability rather than accurate recovery of all parameters. When verifier discrimination approaches zero, calibration itself fails and estimation error can flip the stopping sign, so we pair VRR-Stop with VRR-Guard, an estimation-free fallback that replaces the incumbent candidate only under a sufficient verification margin. On a GSM8K stress setting, VRR-Stop improves final true validity by 60.6 percentage points over fixed five-round repair at an average cost of 0.72 repair rounds. Across settings, stopping reliability is governed jointly by verifier discrimination and the decision margin rather than by the absolute size of estimation error.
展开完整摘要收起摘要↓
Verify-repair loops are a standard means for large language model (LLM) agents to correct faulty plans in code generation, mathematical reasoning, and tool use. When both the verifier and the repairer are noisy, repair can damage already-correct plans, and reported acceptance keeps rising while true validity falls, so existing methods lack a principled basis for deciding when repair should stop. We propose VRR-Stop, a robust stopping framework for noisy verify-repair-repeat (VRR) loops. A four-parameter noise model separates verifier false acceptance and false rejection from the repair and damage behavior of the repairer. Belief filtering turns repeated verification votes into an estimate of committed validity, and the loop commits or repairs according to the sign of the true marginal gain, which requires only sign identifiability rather than accurate recovery of all parameters. When verifier discrimination approaches zero, calibration itself fails and estimation error can flip the stopping sign, so we pair VRR-Stop with VRR-Guard, an estimation-free fallback that replaces the incumbent candidate only under a sufficient verification margin. On a GSM8K stress setting, VRR-Stop improves final true validity by 60.6 percentage points over fixed five-round repair at an average cost of 0.72 repair rounds. Across settings, stopping reliability is governed jointly by verifier discrimination and the decision margin rather than by the absolute size of estimation error.
作者Kemal Devrim Kafadar, Eren Özaltun, Mahmud Efnan Şanlı, Feyza Orak, Emirhan Gazi, Kubilay Kağan Kömürcü, Nazım Kemal Üre
Robust multi-agent coordination relies heavily on inter-agent communication, which is frequently disrupted by physical and environmental constraints in real-world deployments. To maintain operation during these intermittent communication failures, agents can employ internal prediction models to estimate missing shared state information. However, predictors trained with standard reconstruction objectives treat all transitions equally. In a Reinforcement Learning context, this forces the model to waste capacity learning stochastic exploration noise and the outdated dynamics of suboptimal policies. In this paper, we propose a value-aware extension of Multi-Agent Observation Sharing under Communication Dropout (MARO) to patch communication gaps; we refer to this method as Value-Aware MARO. By dynamically weighting the predictor's loss function using advantage estimates derived from the underlying actor-critic architecture, our objective explicitly couples the predictor's learning process to the policy's evolution. This formulation focuses the model's capacity on the intentional, high-return dynamics actively reinforced by the agents. We evaluate our framework on several tasks within the Multi-Agent Particle Environment under varying communication reliability levels. Experimental results demonstrate that our approach maintains performance under declining communication reliability, particularly below 40%. While our method performs comparably in tasks where the baseline already maintains high coordination, our value-aware weighting effectively prevents the performance collapse observed in the standard predictor during high-attrition scenarios. In these environments, our method achieves an average improvement in mean returns of more than 20% and reduces performance variance by a mean of 64.7% compared to the standard unweighted baseline.
展开完整摘要收起摘要↓
Robust multi-agent coordination relies heavily on inter-agent communication, which is frequently disrupted by physical and environmental constraints in real-world deployments. To maintain operation during these intermittent communication failures, agents can employ internal prediction models to estimate missing shared state information. However, predictors trained with standard reconstruction objectives treat all transitions equally. In a Reinforcement Learning context, this forces the model to waste capacity learning stochastic exploration noise and the outdated dynamics of suboptimal policies. In this paper, we propose a value-aware extension of Multi-Agent Observation Sharing under Communication Dropout (MARO) to patch communication gaps; we refer to this method as Value-Aware MARO. By dynamically weighting the predictor's loss function using advantage estimates derived from the underlying actor-critic architecture, our objective explicitly couples the predictor's learning process to the policy's evolution. This formulation focuses the model's capacity on the intentional, high-return dynamics actively reinforced by the agents. We evaluate our framework on several tasks within the Multi-Agent Particle Environment under varying communication reliability levels. Experimental results demonstrate that our approach maintains performance under declining communication reliability, particularly below 40%. While our method performs comparably in tasks where the baseline already maintains high coordination, our value-aware weighting effectively prevents the performance collapse observed in the standard predictor during high-attrition scenarios. In these environments, our method achieves an average improvement in mean returns of more than 20% and reduces performance variance by a mean of 64.7% compared to the standard unweighted baseline.
Most reinforcement learning controllers for these networks assume stationary conditions, and the few that handle change react to the external environment while leaving the network's internal state unexamined. We show that sustained non-stationarity damages this internal state directly: as objectives shift, neurons progressively fall dormant and the shared policy loses the capacity to learn. The obvious remedy, resetting dormant neurons, is unsafe under shared-parameter multi-agent training: many neurons that appear inactive are still receiving strong training gradients, and whether a neuron appears dormant depends on which agent's observations it processes. PRIME (Plasticity Recovery In Multi-agent Environments) therefore verifies both directions before intervening. Extending the bidirectional Silent Neuron framework to cooperative multi-agent reinforcement learning, it aggregates activation and gradient statistics over the full team batch, reads the backward signal from the gradient the training loss has already deposited , not from a hand-crafted proxy, and reinitializes only neurons that are simultaneously activation-dormant and gradient-silent. Useful representations are preserved while learning capacity is restored. On a phase-switching UAV emergency communication simulator, PRIME improves interquartile mean return by 24.9% over MAPPO and holds dormant neuron fractions at 10--20% versus 40--45%; ablations attribute the gains to the gradient signal and team-level aggregation rather than to the specific reset operator. A dynamic regret bound shows that the perturbation cost scales with the small silent-subspace dimension rather than the full parameter count.
展开完整摘要收起摘要↓
Most reinforcement learning controllers for these networks assume stationary conditions, and the few that handle change react to the external environment while leaving the network's internal state unexamined. We show that sustained non-stationarity damages this internal state directly: as objectives shift, neurons progressively fall dormant and the shared policy loses the capacity to learn. The obvious remedy, resetting dormant neurons, is unsafe under shared-parameter multi-agent training: many neurons that appear inactive are still receiving strong training gradients, and whether a neuron appears dormant depends on which agent's observations it processes. PRIME (Plasticity Recovery In Multi-agent Environments) therefore verifies both directions before intervening. Extending the bidirectional Silent Neuron framework to cooperative multi-agent reinforcement learning, it aggregates activation and gradient statistics over the full team batch, reads the backward signal from the gradient the training loss has already deposited , not from a hand-crafted proxy, and reinitializes only neurons that are simultaneously activation-dormant and gradient-silent. Useful representations are preserved while learning capacity is restored. On a phase-switching UAV emergency communication simulator, PRIME improves interquartile mean return by 24.9% over MAPPO and holds dormant neuron fractions at 10--20% versus 40--45%; ablations attribute the gains to the gradient signal and team-level aggregation rather than to the specific reset operator. A dynamic regret bound shows that the perturbation cost scales with the small silent-subspace dimension rather than the full parameter count.
作者Di Lu, Bo Zhang, Xiyuan Li, Yongzhi Liao, Xuewen Dong, Yulong Shen, Zhiquan Liu, Jianfeng Ma
Natural-language control offers a promising interface for unmanned aerial vehicles (UAVs), but directly applying self-hosted computer-use agents (SHCUAs) to UAV control introduces a structural mismatch. SHCUAs are designed for interactive host-side tool use, where delayed agent iterations are often acceptable. UAV control, however, is coupled with continuously changing physical states, strict timing constraints, safety risks, and security accountability. A stale, unauthorized, or tampered agent decision may therefore lead to unsafe or untraceable vehicle behavior. This paper proposes a real-time and security-oriented restructuring of SHCUA-based UAV control. Instead of allowing an SHCUA to directly issue flight commands, we transform its outputs into contract-bound UAV skill invocations with explicit timing, state, authority, fallback, and evidence semantics. Based on this abstraction, we design an architecture that separates semantic reasoning from onboard execution and security/safety enforcement. Slow cloud or edge reasoning is used for mission understanding, while onboard components validate and dispatch only timely, authorized, and state-consistent skills. Security-critical enforcement points can be protected by TEE-style or microcontroller isolation mechanisms without moving the full language agent or high-frequency flight-control loop into trusted components. Prototype evaluation shows that RT-SHCUA maintains bounded task-level responsiveness while supporting degraded handling, trusted admission, and auditable evidence preservation for SHCUA-mediated UAV actions.
展开完整摘要收起摘要↓
Natural-language control offers a promising interface for unmanned aerial vehicles (UAVs), but directly applying self-hosted computer-use agents (SHCUAs) to UAV control introduces a structural mismatch. SHCUAs are designed for interactive host-side tool use, where delayed agent iterations are often acceptable. UAV control, however, is coupled with continuously changing physical states, strict timing constraints, safety risks, and security accountability. A stale, unauthorized, or tampered agent decision may therefore lead to unsafe or untraceable vehicle behavior. This paper proposes a real-time and security-oriented restructuring of SHCUA-based UAV control. Instead of allowing an SHCUA to directly issue flight commands, we transform its outputs into contract-bound UAV skill invocations with explicit timing, state, authority, fallback, and evidence semantics. Based on this abstraction, we design an architecture that separates semantic reasoning from onboard execution and security/safety enforcement. Slow cloud or edge reasoning is used for mission understanding, while onboard components validate and dispatch only timely, authorized, and state-consistent skills. Security-critical enforcement points can be protected by TEE-style or microcontroller isolation mechanisms without moving the full language agent or high-frequency flight-control loop into trusted components. Prototype evaluation shows that RT-SHCUA maintains bounded task-level responsiveness while supporting degraded handling, trusted admission, and auditable evidence preservation for SHCUA-mediated UAV actions.
We introduce WC2026-Agents, a benchmark and dataset for evaluating large language models (LLMs) as autonomous forecasting agents on real, future events. For every one of the 104 matches of the 2026 FIFA World Cup, four frontier models — Claude Opus 4.8, ChatGPT (GPT-5.5, high reasoning), Gemini 3.1 Pro, and Grok (Expert Mode) — ran an identical search-act-reflect loop: gather evidence with a web tool, commit to a 1X2 (team-A win / draw / team-B win) distribution and a virtual 100-USD bet, and, after the match, reflect given only the final score. Because every match kicked off after the models' training cutoffs, the benchmark is contamination-free by construction. Crucially, we pair the four agents with a fifth competitor drawn from the same information environment — the pre-match betting market — collected as per-match 1X2 odds, giving an economically grounded baseline and letting us score not just what an agent predicts but what it does with money. The release contains 416 forecasts and 414 reflections with verbatim reasoning, ground truth (including penalty shootouts), odds, and a reproducible evaluation suite. A reference evaluation surfaces findings that raw accuracy hides: the four agents issue an identical top pick in 92% of matches and none beats the market's Brier score; indeed, a naive flat stake on the market favorite out-earns all four agents. Yet the agents diverge sharply as decision-makers: betting return-on-investment ranges from -18% to +10%, fading the market is unprofitable for all four, the share of forecasts that cite the market ranges from 12% to 100%, and self-reported error rates on wrong picks range from 36% to 86%. The benchmark thus measures calibration, decision quality, and self-knowledge — axes on which frontier models differ even when their predictions do not. Data and code: https://github.com/graphuofm/FIFA2026LLM
展开完整摘要收起摘要↓
We introduce WC2026-Agents, a benchmark and dataset for evaluating large language models (LLMs) as autonomous forecasting agents on real, future events. For every one of the 104 matches of the 2026 FIFA World Cup, four frontier models — Claude Opus 4.8, ChatGPT (GPT-5.5, high reasoning), Gemini 3.1 Pro, and Grok (Expert Mode) — ran an identical search-act-reflect loop: gather evidence with a web tool, commit to a 1X2 (team-A win / draw / team-B win) distribution and a virtual 100-USD bet, and, after the match, reflect given only the final score. Because every match kicked off after the models' training cutoffs, the benchmark is contamination-free by construction. Crucially, we pair the four agents with a fifth competitor drawn from the same information environment — the pre-match betting market — collected as per-match 1X2 odds, giving an economically grounded baseline and letting us score not just what an agent predicts but what it does with money. The release contains 416 forecasts and 414 reflections with verbatim reasoning, ground truth (including penalty shootouts), odds, and a reproducible evaluation suite. A reference evaluation surfaces findings that raw accuracy hides: the four agents issue an identical top pick in 92% of matches and none beats the market's Brier score; indeed, a naive flat stake on the market favorite out-earns all four agents. Yet the agents diverge sharply as decision-makers: betting return-on-investment ranges from -18% to +10%, fading the market is unprofitable for all four, the share of forecasts that cite the market ranges from 12% to 100%, and self-reported error rates on wrong picks range from 36% to 86%. The benchmark thus measures calibration, decision quality, and self-knowledge — axes on which frontier models differ even when their predictions do not. Data and code: https://github.com/graphuofm/FIFA2026LLM
作者Martino M. L. Pulici, Cuong Xuan Chu, Evgeny Kharlamov, Zifeng Ding, Volker Tresp, Yunpu Ma
Large language models achieve strong reasoning performance, but often at prohibitive training cost - a challenge that is especially acute for compact models ($\leq 4 \, \mathrm{B}$ parameters) trained under limited budgets. We introduce MADA-RL, a post-training framework that specializes compact models into generator and critic roles and trains them with a debate-aware learning signal, fine-tuning only a small subset of parameters via LoRA adapters. Our central contribution is a counterfactual critic advantage: a dynamic, role-conditioned baseline that redefines the critic's advantage as its reward minus the generator ensemble's per-instance accuracy. This explicitly optimizes critics to improve over generator consensus rather than to merely reproduce a correct answer, yielding more targeted credit assignment than static mean-reward normalization. At deployment, the specialized agents are composed in a lightweight multi-round protocol. Across five mathematical reasoning benchmarks, MADA-RL raises the accuracy of the DeepSeek-R1-Distill-Qwen-1.5B model from $39.9 \, %$ to $41.9 \, %$ ($+2.0$ points, $p < 0.001$) using $16$ times fewer trainable parameters than fully fine-tuned baselines, placing it on the accuracy-trainable-parameter Pareto front. It approaches, but does not surpass, the strongest baselines (DeepScaleR, STILL-3), which are trained on substantially larger datasets; we analyse this gap and the associated inference-time cost directly. A controlled study isolates the source of MADA-RL's gains: the counterfactual advantage produces the highest critic improvement rate of any model evaluated, indicating that trained critics learn to correct generator errors rather than to imitate them.
展开完整摘要收起摘要↓
Large language models achieve strong reasoning performance, but often at prohibitive training cost - a challenge that is especially acute for compact models ($\leq 4 \, \mathrm{B}$ parameters) trained under limited budgets. We introduce MADA-RL, a post-training framework that specializes compact models into generator and critic roles and trains them with a debate-aware learning signal, fine-tuning only a small subset of parameters via LoRA adapters. Our central contribution is a counterfactual critic advantage: a dynamic, role-conditioned baseline that redefines the critic's advantage as its reward minus the generator ensemble's per-instance accuracy. This explicitly optimizes critics to improve over generator consensus rather than to merely reproduce a correct answer, yielding more targeted credit assignment than static mean-reward normalization. At deployment, the specialized agents are composed in a lightweight multi-round protocol. Across five mathematical reasoning benchmarks, MADA-RL raises the accuracy of the DeepSeek-R1-Distill-Qwen-1.5B model from $39.9 \, %$ to $41.9 \, %$ ($+2.0$ points, $p < 0.001$) using $16$ times fewer trainable parameters than fully fine-tuned baselines, placing it on the accuracy-trainable-parameter Pareto front. It approaches, but does not surpass, the strongest baselines (DeepScaleR, STILL-3), which are trained on substantially larger datasets; we analyse this gap and the associated inference-time cost directly. A controlled study isolates the source of MADA-RL's gains: the counterfactual advantage produces the highest critic improvement rate of any model evaluated, indicating that trained critics learn to correct generator errors rather than to imitate them.
Recovering high dynamic range (HDR) radiance from a single standard dynamic range (SDR) image is highly ill-posed. Extreme luminance variation and severe quantization in dark regions make accurate reconstruction challenging, often leading to visual artifacts and color distortions. To address this problem, we propose Bio-SFT, a bio-inspired spiking frequency transformer for single-image HDR reconstruction. Bio-SFT incorporates three biologically motivated components. First, a learnable Naka--Rushton retinal adaptation frontend stabilizes the input under complex lighting conditions. Second, an explicit Parvo--Magno split introduces asymmetric Parvo-to-Magno guidance, allowing high-frequency structural cues to modulate low-frequency reconstruction. Third, an event-driven SNN hard gating module applies all-or-none spiking to suppress dark-region noise while preserving structural details. The module is trained with a sparsity prior to encourage efficient feature utilization. Built for end-to-end training within a transformer backbone, these lightweight components provide strong parameter efficiency. Experiments on HDRTV1K show that Bio-SFT achieves competitive perceptual quality and consistently improves HDR-VDP-3 and $ΔE_{ITP}$ while reducing artifact propagation in symmetric guidance pipelines.
展开完整摘要收起摘要↓
Recovering high dynamic range (HDR) radiance from a single standard dynamic range (SDR) image is highly ill-posed. Extreme luminance variation and severe quantization in dark regions make accurate reconstruction challenging, often leading to visual artifacts and color distortions. To address this problem, we propose Bio-SFT, a bio-inspired spiking frequency transformer for single-image HDR reconstruction. Bio-SFT incorporates three biologically motivated components. First, a learnable Naka--Rushton retinal adaptation frontend stabilizes the input under complex lighting conditions. Second, an explicit Parvo--Magno split introduces asymmetric Parvo-to-Magno guidance, allowing high-frequency structural cues to modulate low-frequency reconstruction. Third, an event-driven SNN hard gating module applies all-or-none spiking to suppress dark-region noise while preserving structural details. The module is trained with a sparsity prior to encourage efficient feature utilization. Built for end-to-end training within a transformer backbone, these lightweight components provide strong parameter efficiency. Experiments on HDRTV1K show that Bio-SFT achieves competitive perceptual quality and consistently improves HDR-VDP-3 and $ΔE_{ITP}$ while reducing artifact propagation in symmetric guidance pipelines.
作者Chang Eun Song, Sumukh Pinge, Tianqi Zhang, Sung Eun Kim, Tajana S. Rosing, Mingu Kang
Retrieval-Augmented Generation (RAG) enhances the factual grounding of large language model (LLM) inference by retrieving relevant information from external knowledge bases. However, its dense vector retrieval introduces significant latency and energy overhead, becoming the primary performance bottleneck. Although recent in-storage accelerators aim to reduce data movement, they still rely on host or embedded processors outside the memory, where nearly 70% of the total retrieval time is spent. As a result, they cannot fully overcome the bandwidth limitations, leading to yet another memory bottleneck. To tackle these limitations, we present D-NOVA, a hardware-software co-designed in-storage retrieval accelerator. D-NOVA executes an inverted file (IVF)-based hierarchical retrieval pipeline by deeply embedding the search functionality directly into the NAND memory array. This is achieved by incorporating a new distance metric, Dual-Bound Tight Similarity Sensing (DTS), which is specifically tailored for searching within the NAND string. In addition, we introduce a lightweight contrastive adapter that maps embedding vectors into a DTS-friendly domain, recovering near-software recall while improving performance and energy efficiency. D-NOVA is up to 41.7x faster and 71x more energy-efficient than a CPU baseline, and achieves 12.13x higher throughput while being up to 1.26x more energy-efficient than state-of-the-art in-storage RAG accelerators, demonstrating the potential of fully in-storage vector search for scalable RAG acceleration.
展开完整摘要收起摘要↓
Retrieval-Augmented Generation (RAG) enhances the factual grounding of large language model (LLM) inference by retrieving relevant information from external knowledge bases. However, its dense vector retrieval introduces significant latency and energy overhead, becoming the primary performance bottleneck. Although recent in-storage accelerators aim to reduce data movement, they still rely on host or embedded processors outside the memory, where nearly 70% of the total retrieval time is spent. As a result, they cannot fully overcome the bandwidth limitations, leading to yet another memory bottleneck. To tackle these limitations, we present D-NOVA, a hardware-software co-designed in-storage retrieval accelerator. D-NOVA executes an inverted file (IVF)-based hierarchical retrieval pipeline by deeply embedding the search functionality directly into the NAND memory array. This is achieved by incorporating a new distance metric, Dual-Bound Tight Similarity Sensing (DTS), which is specifically tailored for searching within the NAND string. In addition, we introduce a lightweight contrastive adapter that maps embedding vectors into a DTS-friendly domain, recovering near-software recall while improving performance and energy efficiency. D-NOVA is up to 41.7x faster and 71x more energy-efficient than a CPU baseline, and achieves 12.13x higher throughput while being up to 1.26x more energy-efficient than state-of-the-art in-storage RAG accelerators, demonstrating the potential of fully in-storage vector search for scalable RAG acceleration.
作者Hanchen Yang, Kaiwen Yang, Junpeng Zhuang, Yang He, Keting Cen, Bochao Liu, Zhongbo Sun, An Liu, Zhongteng Han, Chenyi Lei
User experience is a first-class objective in industrial e-commerce recommender systems (RS). Post-ranking strategies, which govern diversity, similarity, and exposure over a ranked list, are widely deployed in industrial RS for their simplicity and low serving cost. However, as the online recommendation environment evolves continuously, these statically configured strategies gradually become stale, thereby degrading the user experience. Refining them typically relies on manual inspection, diagnosis, and updates, making it slow, costly, and difficult to scale or reuse. Although recent LLM-based agents (e.g., RecUserSim, SimUSER, and Self-EvolveRec) offer promising directions, none of them close the full loop of automated, self-evolving strategy refinement. To bridge this gap, we introduce SR-Agent, which, to the best of our knowledge, is the first agentic framework deployed to refine post-ranking strategies in industrial RS. SR-Agent unifies three components: (i) a UserSim agent that applies inspection skills to surface user-perceived bad cases; (ii) an Analysis agent that consolidates recurring bad cases into structured, reusable diagnoses; and (iii) a constrained Strategy Refinement Harness that maps diagnoses to typed and bounded actions, gated by a four-stage reward pipeline with reversible rollback. Deployed on the Kuaishou e-commerce platform, SR-Agent continuously runs this refinement loop and, in a one-month online A/B test, increases order volume by 0.71%, browsing depth by 0.34%, and clicked-category diversity by 0.48%, while markedly shortening the refinement cycle and lowering operational cost.
展开完整摘要收起摘要↓
User experience is a first-class objective in industrial e-commerce recommender systems (RS). Post-ranking strategies, which govern diversity, similarity, and exposure over a ranked list, are widely deployed in industrial RS for their simplicity and low serving cost. However, as the online recommendation environment evolves continuously, these statically configured strategies gradually become stale, thereby degrading the user experience. Refining them typically relies on manual inspection, diagnosis, and updates, making it slow, costly, and difficult to scale or reuse. Although recent LLM-based agents (e.g., RecUserSim, SimUSER, and Self-EvolveRec) offer promising directions, none of them close the full loop of automated, self-evolving strategy refinement. To bridge this gap, we introduce SR-Agent, which, to the best of our knowledge, is the first agentic framework deployed to refine post-ranking strategies in industrial RS. SR-Agent unifies three components: (i) a UserSim agent that applies inspection skills to surface user-perceived bad cases; (ii) an Analysis agent that consolidates recurring bad cases into structured, reusable diagnoses; and (iii) a constrained Strategy Refinement Harness that maps diagnoses to typed and bounded actions, gated by a four-stage reward pipeline with reversible rollback. Deployed on the Kuaishou e-commerce platform, SR-Agent continuously runs this refinement loop and, in a one-month online A/B test, increases order volume by 0.71%, browsing depth by 0.34%, and clicked-category diversity by 0.48%, while markedly shortening the refinement cycle and lowering operational cost.
作者Tingjia Zhang, Bo Chen, Shengzhong Liu, Fan Wu, Guihai Chen
Recent breakthroughs in 3D Gaussian Splatting (3DGS) have advanced neural rendering with high fidelity and speed. However, its performance degrades significantly in large-scale scenes due to the computational burden of tile-based rasterization. Existing optimization efforts either require costly scene re-training or focus on narrow aspects of the pipeline, overlooking critical inefficiencies in real-world deployments. Through a comprehensive analysis, we identify three primary sources of redundancy and low GPU utilization: redundant inter-frame pre-processing, viewpoint-based occlusion redundancy, and severe tile-level load imbalance. To address these issues, we propose CaT-GS, a novel and efficient 3DGS rendering pipeline. CaT-GS introduces a speculative multi-frame preprocessing method to eliminate redundant computations across consecutive frames, and an inter-frame caching mechanism to eliminate viewpoint redundant rendering stages. Furthermore, it refactors rasterization tasks with a dedicated kernel to mitigate tile load imbalance, significantly boosting GPU utilization. Extensive experiments demonstrate that CaT-GS achieves a speedup of up to 10 times over the original 3DGS and up to 70% over previous state-of-the-art methods, establishing a new benchmark for high-fidelity, real-time rendering of large-scale scenes.
展开完整摘要收起摘要↓
Recent breakthroughs in 3D Gaussian Splatting (3DGS) have advanced neural rendering with high fidelity and speed. However, its performance degrades significantly in large-scale scenes due to the computational burden of tile-based rasterization. Existing optimization efforts either require costly scene re-training or focus on narrow aspects of the pipeline, overlooking critical inefficiencies in real-world deployments. Through a comprehensive analysis, we identify three primary sources of redundancy and low GPU utilization: redundant inter-frame pre-processing, viewpoint-based occlusion redundancy, and severe tile-level load imbalance. To address these issues, we propose CaT-GS, a novel and efficient 3DGS rendering pipeline. CaT-GS introduces a speculative multi-frame preprocessing method to eliminate redundant computations across consecutive frames, and an inter-frame caching mechanism to eliminate viewpoint redundant rendering stages. Furthermore, it refactors rasterization tasks with a dedicated kernel to mitigate tile load imbalance, significantly boosting GPU utilization. Extensive experiments demonstrate that CaT-GS achieves a speedup of up to 10 times over the original 3DGS and up to 70% over previous state-of-the-art methods, establishing a new benchmark for high-fidelity, real-time rendering of large-scale scenes.
Vision Transformers process spatially redundant tokens efficiently only when coarse token summaries preserve the evidence required by exponential attention aggregation. We identify a boundary-minority underestimation failure in which a spatially small, high-response region contributes dominant Gibbs mass while remaining nearly invisible to a block mean. We formalize the failure through the discrepancy between normalized log-mean-exp free energy and mean summarization, prove that minority Gibbs mass can remain non-vanishing as its spatial support and mean contribution vanish, and characterize the limitations of finite-order moment corrections. Building on the resulting analysis, we introduce Boundary-Minority Free-Energy Adaptive Screening (BMFA), which constructs a hierarchical piecewise-constant approximation and recursively refines blocks according to a computable lower-bound increment of local free energy. Controlled synthetic tests, COCO and LVIS diagnostic probes, closed-loop DeiT-Tiny evaluations, and ImageNet-1K experiments establish a consistent evidence chain. BMFA reduces the mean synthetic underestimate from 2.582 to 0.261 at a 5.794% leaf ratio, lowers the COCO image-edge mean gap from 2.254 to 0.526, and preserves 71.520% ImageNet Top-1 accuracy at a 55.861% leaf ratio. The current prototype evaluates selection quality after full QK computation; the reported leaf ratio therefore characterizes representation granularity rather than verified sparse-kernel speedup.
展开完整摘要收起摘要↓
Vision Transformers process spatially redundant tokens efficiently only when coarse token summaries preserve the evidence required by exponential attention aggregation. We identify a boundary-minority underestimation failure in which a spatially small, high-response region contributes dominant Gibbs mass while remaining nearly invisible to a block mean. We formalize the failure through the discrepancy between normalized log-mean-exp free energy and mean summarization, prove that minority Gibbs mass can remain non-vanishing as its spatial support and mean contribution vanish, and characterize the limitations of finite-order moment corrections. Building on the resulting analysis, we introduce Boundary-Minority Free-Energy Adaptive Screening (BMFA), which constructs a hierarchical piecewise-constant approximation and recursively refines blocks according to a computable lower-bound increment of local free energy. Controlled synthetic tests, COCO and LVIS diagnostic probes, closed-loop DeiT-Tiny evaluations, and ImageNet-1K experiments establish a consistent evidence chain. BMFA reduces the mean synthetic underestimate from 2.582 to 0.261 at a 5.794% leaf ratio, lowers the COCO image-edge mean gap from 2.254 to 0.526, and preserves 71.520% ImageNet Top-1 accuracy at a 55.861% leaf ratio. The current prototype evaluates selection quality after full QK computation; the reported leaf ratio therefore characterizes representation granularity rather than verified sparse-kernel speedup.
作者Yue Shui, Chenyu Ma, Hangfei Xu, Shengzhao Wen, Yanpeng Wang
Large language models (LLMs) can assist GPU kernel generation, but their practical effectiveness depends on whether generated code can be reliably constrained, validated, profiled, and selected. This paper presents a harness-centered system for LLM-driven GPU kernel optimization in the MLSys 2026 FlashInfer AI Kernel Generation Contest on NVIDIA Blackwell B200 GPUs. The system separates an evaluation harness from a profile-backed optimization controller: the harness enforces compilation, correctness, official-aligned timing, and artifact archival, while the controller turns profiler and workload evidence into bounded candidate-generation decisions. Human-authored skills capture operator constraints, references, profiling procedures, and promotion rules, while Codex and Claude Code agents generate candidate kernels inside those constraints. Across five operator definitions, the retained official-aligned artifacts achieved mean-latency speedups over supplied FlashInfer baselines of 1.62x, 18.05x, 29.68x, 1.12x, and 13.70x. The Agent-Assisted kernels outperform the Full-Agent artifacts across the evaluated definitions, indicating that expert-provided optimization directions, high-quality references, and workload context remain critical for reliable AI-driven kernel optimization.
展开完整摘要收起摘要↓
Large language models (LLMs) can assist GPU kernel generation, but their practical effectiveness depends on whether generated code can be reliably constrained, validated, profiled, and selected. This paper presents a harness-centered system for LLM-driven GPU kernel optimization in the MLSys 2026 FlashInfer AI Kernel Generation Contest on NVIDIA Blackwell B200 GPUs. The system separates an evaluation harness from a profile-backed optimization controller: the harness enforces compilation, correctness, official-aligned timing, and artifact archival, while the controller turns profiler and workload evidence into bounded candidate-generation decisions. Human-authored skills capture operator constraints, references, profiling procedures, and promotion rules, while Codex and Claude Code agents generate candidate kernels inside those constraints. Across five operator definitions, the retained official-aligned artifacts achieved mean-latency speedups over supplied FlashInfer baselines of 1.62x, 18.05x, 29.68x, 1.12x, and 13.70x. The Agent-Assisted kernels outperform the Full-Agent artifacts across the evaluated definitions, indicating that expert-provided optimization directions, high-quality references, and workload context remain critical for reliable AI-driven kernel optimization.
作者Haiyue Ma, Lauren Malek, Joseph Forzani, David Wentzlaff
As more capable AI models are increasingly integrated into critical computer systems, the lack of control over AI intent motivates safety mechanisms. Existing software safeguards impose only behavioral constraints that can potentially be bypassed by sufficiently intelligent models. While hardware-level safety enforcement has been recognized as an essential last line of defense, few mechanisms have been proposed beyond policy regulations on unauthorized accesses or coarse full-chip shutdown. What is missing is a fine-grained, dynamic intervention mechanism at the architecture level. In this paper, we introduce a set of microarchitecture knobs which dynamically control the available hardware resources to limit AI performance at runtime. We evaluate candidate knobs spanning the GPU memory subsystem, across capacity, bandwidth, latency and frequency dimensions, and narrow down to four strong candidates: L2 size, L2 latency, L2 bandwidth, and shared memory port access rate. To minimize new logic and extra design cost, we build all four mechanisms from well-established microarchitectural primitives: cache way masking, credit-based rate limiting, latency insertion, and bank arbitration. We show that these knobs achieve high performance sensitivity (up to 80% performance cut at 1/8 resource availability), negligible implementation cost (<~10K flip flops), fast stabilization after dynamic throttling (5-80K cycles), and minimal collateral impact on the rest of the chip. Further, multi-knob analysis reveals combinations of knobs that amplify the performance degradation beyond the effect of each knob individually, which enables a broader range of performance targets.
展开完整摘要收起摘要↓
As more capable AI models are increasingly integrated into critical computer systems, the lack of control over AI intent motivates safety mechanisms. Existing software safeguards impose only behavioral constraints that can potentially be bypassed by sufficiently intelligent models. While hardware-level safety enforcement has been recognized as an essential last line of defense, few mechanisms have been proposed beyond policy regulations on unauthorized accesses or coarse full-chip shutdown. What is missing is a fine-grained, dynamic intervention mechanism at the architecture level. In this paper, we introduce a set of microarchitecture knobs which dynamically control the available hardware resources to limit AI performance at runtime. We evaluate candidate knobs spanning the GPU memory subsystem, across capacity, bandwidth, latency and frequency dimensions, and narrow down to four strong candidates: L2 size, L2 latency, L2 bandwidth, and shared memory port access rate. To minimize new logic and extra design cost, we build all four mechanisms from well-established microarchitectural primitives: cache way masking, credit-based rate limiting, latency insertion, and bank arbitration. We show that these knobs achieve high performance sensitivity (up to 80% performance cut at 1/8 resource availability), negligible implementation cost (<~10K flip flops), fast stabilization after dynamic throttling (5-80K cycles), and minimal collateral impact on the rest of the chip. Further, multi-knob analysis reveals combinations of knobs that amplify the performance degradation beyond the effect of each knob individually, which enables a broader range of performance targets.
作者Shizeng Jiang, Hao Zhang, Xuerui Ma, Ying Hu, Tao Zhang
3D Gaussian Splatting (3DGS) achieves high-quality real-time rendering by representing a scene with a large collection of anisotropic Gaussian primitives. However, complex scenes often require millions of Gaussians, resulting in substantial storage and rendering costs. Existing compression methods mainly reduce redundancy through primitive-wise pruning, attribute quantization, clustering, or neural coding, while redundancy caused by strongly overlapping and non-orthogonal Gaussian basis functions remains largely unexplored. We present QIRF, a quantum-inspired non-orthogonal function-space compression method for 3D Gaussian Splatting. QIRF models neighboring Gaussian primitives as a local non-orthogonal basis and formulates primitive reduction as a subspace-aware selection problem. Specifically, an analytic Gaussian overlap matrix and a radiance-response density matrix are constructed to characterize functional redundancy and rendering relevance. Generalized eigendecomposition is then used to identify the dominant local subspace and select representative Gaussian primitives. An RRDM-based response model and detail-aware safeguarding further preserve visually important high-frequency structures under aggressive pruning. Experiments on 13 scenes from Mip-NeRF 360, Tanks and Temples, and Deep Blending show that QIRF reduces the Gaussian count and raw PLY storage by 71.7 percent on average, corresponding to approximately 3.54 times compression, while maintaining reconstruction quality comparable to 3DGS and achieving a marginal average PSNR improvement of 0.10 dB. QIRF also improves the average rendering speed over 3DGS by 34.3 percent. These results suggest that non-orthogonal function-space redundancy is an important yet underexplored source of representational redundancy in explicit Gaussian radiance fields.
展开完整摘要收起摘要↓
3D Gaussian Splatting (3DGS) achieves high-quality real-time rendering by representing a scene with a large collection of anisotropic Gaussian primitives. However, complex scenes often require millions of Gaussians, resulting in substantial storage and rendering costs. Existing compression methods mainly reduce redundancy through primitive-wise pruning, attribute quantization, clustering, or neural coding, while redundancy caused by strongly overlapping and non-orthogonal Gaussian basis functions remains largely unexplored. We present QIRF, a quantum-inspired non-orthogonal function-space compression method for 3D Gaussian Splatting. QIRF models neighboring Gaussian primitives as a local non-orthogonal basis and formulates primitive reduction as a subspace-aware selection problem. Specifically, an analytic Gaussian overlap matrix and a radiance-response density matrix are constructed to characterize functional redundancy and rendering relevance. Generalized eigendecomposition is then used to identify the dominant local subspace and select representative Gaussian primitives. An RRDM-based response model and detail-aware safeguarding further preserve visually important high-frequency structures under aggressive pruning. Experiments on 13 scenes from Mip-NeRF 360, Tanks and Temples, and Deep Blending show that QIRF reduces the Gaussian count and raw PLY storage by 71.7 percent on average, corresponding to approximately 3.54 times compression, while maintaining reconstruction quality comparable to 3DGS and achieving a marginal average PSNR improvement of 0.10 dB. QIRF also improves the average rendering speed over 3DGS by 34.3 percent. These results suggest that non-orthogonal function-space redundancy is an important yet underexplored source of representational redundancy in explicit Gaussian radiance fields.
作者Simla Burcu Harma, Danila Mishin, Zhengyuan Su, Ayan Chakraborty, Elizaveta Kostenok, Dongho Ha, Babak Falsafi, Martin Jaggi, Yunho Oh, Amir Yazdanbakhsh
4-bit quantization enables efficient LLM inference, but suffers from significant accuracy degradation due to outliers. Prior work addresses this problem via data rotation or mixed-precision integer quantization, but often relies on software-managed scaling and frequent dequantization, incurring substantial overhead. Microscaling formats, such as MXINT, eliminate these inefficiencies by encoding scales in hardware, yet remain incompatible with rotation-based methods. Our analysis reveals that outliers vary in severity, from rare extremes to frequent mild deviations, and that quantization sensitivity is unevenly distributed across layers and columns. These insights motivate a fine-grained, sensitivity-guided approach. We introduce MXSens, a training-free method that assigns mixed mantissa bitwidths (4/6/8) based on column- and layer-wise sensitivity, naturally leveraging the block-wise structure of MXINT. MXSens outperforms state-of-the-art quantization methods across a range of models and tasks. Under the W4A4KV4 setting, MXSens achieves perplexities of 3.77 and 7.63 on LLaMA-2-70B and LLaMA-3-8B, respectively, substantially improving over existing baselines on WikiText-2. Our work establishes a new balance between accuracy and resource efficiency for LLM quantization.
展开完整摘要收起摘要↓
4-bit quantization enables efficient LLM inference, but suffers from significant accuracy degradation due to outliers. Prior work addresses this problem via data rotation or mixed-precision integer quantization, but often relies on software-managed scaling and frequent dequantization, incurring substantial overhead. Microscaling formats, such as MXINT, eliminate these inefficiencies by encoding scales in hardware, yet remain incompatible with rotation-based methods. Our analysis reveals that outliers vary in severity, from rare extremes to frequent mild deviations, and that quantization sensitivity is unevenly distributed across layers and columns. These insights motivate a fine-grained, sensitivity-guided approach. We introduce MXSens, a training-free method that assigns mixed mantissa bitwidths (4/6/8) based on column- and layer-wise sensitivity, naturally leveraging the block-wise structure of MXINT. MXSens outperforms state-of-the-art quantization methods across a range of models and tasks. Under the W4A4KV4 setting, MXSens achieves perplexities of 3.77 and 7.63 on LLaMA-2-70B and LLaMA-3-8B, respectively, substantially improving over existing baselines on WikiText-2. Our work establishes a new balance between accuracy and resource efficiency for LLM quantization.
作者Nikola Pižurica, Matteo Risso, Nikola Milović, Alessio Burrello, Igor Jovančević, Conor Heins, Miguel de Prado
Bayesian inference provides a principled foundation for reasoning under uncertainty, but its computational cost hinders deployment on resource-constrained edge devices. In this paper, we present a hardware-oriented methodology for accelerating discrete Bayesian inference on commercial off-the-shelf embedded GPUs. We identify that the latency of a broad class of variational message-passing algorithms is dominated by tensor contractions. Our approach restructures the memory layout of these operations using two complementary merging strategies that produce compact, regularly-shaped primitives better suited for efficient GPU execution. We then introduce optional sparse array representations and a tensor-clustering scheme to reduce the memory footprint. We instantiate the methodology and produce optimized variants of three message-passing algorithms for Hidden Markov Models (HMMs), namely variational filtering, variational message passing, and marginal message passing. Furthermore, we complement this with a machine-learning-based autotuner that automatically selects the best-performing algorithmic variant for a given generative model specification. Benchmarked on an NVIDIA Jetson Orin AGX across 770 randomly sampled realistic Partially Observable Markov Decision Process (POMDP) configurations, our implementations achieve speedups of up to 5x, with typical gains of 2-2.5x, while producing numerically identical outputs to the baseline implementations.
展开完整摘要收起摘要↓
Bayesian inference provides a principled foundation for reasoning under uncertainty, but its computational cost hinders deployment on resource-constrained edge devices. In this paper, we present a hardware-oriented methodology for accelerating discrete Bayesian inference on commercial off-the-shelf embedded GPUs. We identify that the latency of a broad class of variational message-passing algorithms is dominated by tensor contractions. Our approach restructures the memory layout of these operations using two complementary merging strategies that produce compact, regularly-shaped primitives better suited for efficient GPU execution. We then introduce optional sparse array representations and a tensor-clustering scheme to reduce the memory footprint. We instantiate the methodology and produce optimized variants of three message-passing algorithms for Hidden Markov Models (HMMs), namely variational filtering, variational message passing, and marginal message passing. Furthermore, we complement this with a machine-learning-based autotuner that automatically selects the best-performing algorithmic variant for a given generative model specification. Benchmarked on an NVIDIA Jetson Orin AGX across 770 randomly sampled realistic Partially Observable Markov Decision Process (POMDP) configurations, our implementations achieve speedups of up to 5x, with typical gains of 2-2.5x, while producing numerically identical outputs to the baseline implementations.
作者Alon Abudraham, Xingyu Chen, Itamar Levi, Ari Trachtenberg
Modern cloud platforms increasingly combine strong software isolation mechanisms with shared hardware resources to improve performance and resource efficiency. Conventional containers do this by sharing the host kernel directly, whereas sandboxed runtimes (e.g., gVisor) and VM-based runtimes (e.g., Kata, QEMU/KVM) provide progressively stronger isolation. In all cases, when tenants access host-backed filesystem state, the host page cache can remain shared and observable. Although OS-managed, this page-cache channel forms an OS-mediated microarchitectural timing side channel whose signal is shaped by the processor microarchitecture, memory and storage hierarchies, and virtualization mechanisms. We thus investigate whether unprivileged timing measurements can reveal page-cache residency across these isolation boundaries. Our evaluation covers Docker; gVisor with systrap and KVM; Kata Containers using QEMU and Cloud Hypervisor with shared host filesystems; Kata using QEMU, Cloud Hypervisor, and Firecracker with block-device-backed storage; and QEMU/KVM virtual machines under multiple host cache policies. Our results show that the timing signal persists whenever the I/O path exposes shared, host-cacheable file-backed objects, including under OverlayFS layers, virtio-fs exports, and loop-backed block devices. However, direct I/O and dedicated block devices substantially attenuate or eliminate the signal. Virtualization therefore reshapes leakage through added latency and algorithmic noise but does not remove the underlying dependence on shared hardware and cache state. We showcase this through a case study in which we recover coarse-grained activity from a WordPress deployment backed by MySQL. These results place page-cache attacks within the broader class of OS-mediated microarchitectural timing channels and motivate coordinated hardware, virtualization, and OS support for timing isolation.
展开完整摘要收起摘要↓
Modern cloud platforms increasingly combine strong software isolation mechanisms with shared hardware resources to improve performance and resource efficiency. Conventional containers do this by sharing the host kernel directly, whereas sandboxed runtimes (e.g., gVisor) and VM-based runtimes (e.g., Kata, QEMU/KVM) provide progressively stronger isolation. In all cases, when tenants access host-backed filesystem state, the host page cache can remain shared and observable. Although OS-managed, this page-cache channel forms an OS-mediated microarchitectural timing side channel whose signal is shaped by the processor microarchitecture, memory and storage hierarchies, and virtualization mechanisms. We thus investigate whether unprivileged timing measurements can reveal page-cache residency across these isolation boundaries. Our evaluation covers Docker; gVisor with systrap and KVM; Kata Containers using QEMU and Cloud Hypervisor with shared host filesystems; Kata using QEMU, Cloud Hypervisor, and Firecracker with block-device-backed storage; and QEMU/KVM virtual machines under multiple host cache policies. Our results show that the timing signal persists whenever the I/O path exposes shared, host-cacheable file-backed objects, including under OverlayFS layers, virtio-fs exports, and loop-backed block devices. However, direct I/O and dedicated block devices substantially attenuate or eliminate the signal. Virtualization therefore reshapes leakage through added latency and algorithmic noise but does not remove the underlying dependence on shared hardware and cache state. We showcase this through a case study in which we recover coarse-grained activity from a WordPress deployment backed by MySQL. These results place page-cache attacks within the broader class of OS-mediated microarchitectural timing channels and motivate coordinated hardware, virtualization, and OS support for timing isolation.
Multi-head Latent Attention (MLA) ships two implementations in Megatron-Core: an explicit form used for training and an absorbed form — which slashes collective communication by gathering only the compressed latent — that is fully implemented but hard-asserted out of training (the forward opens with "assert not (self.training and self.cache_mla_latents)"), allowed only in inference decode. The library documents no reason. We show the restriction is well-founded and quantify why: ported to training, the absorbed form is a memory trap — its intermediates live in n_h x d_kv dimensions per token, larger than the per-head K/V they replace — inflating activation memory by 20-34%, up to 9.2 GB at DeepSeek-V3 scale (n_h=128, seq=16384, SP=8, eager kernel; the gap widens to 19.2 GB under a fused kernel), enough to change device-fit. This measurement, validated on two axes (linear in seq and n_h) and cross-verified on NVIDIA A100, explains the otherwise-undocumented restriction and leaves practitioners with no low-communication MLA training path. We then provide one. LAGA (Latent All-Gather Attention) keeps the absorbed form's latent-gather communication but rejects the absorb reformulation, instead reconstructing per-head K/V locally from the gathered latent. On 8x Ascend 910B at real DeepSeek-V3 dimensions, LAGA cuts collective communication 1.98x, matches explicit memory within 0.5%, is bit-identical to explicit at SP=1 and equivalent to within 1e-3 at SP=2-8, and under a fused attention kernel improves attention-block throughput 1.04-1.06x single-node and 1.07-1.24x cross-node — leading at all sequence lengths in the cross-node regime MLA is deployed for.
展开完整摘要收起摘要↓
Multi-head Latent Attention (MLA) ships two implementations in Megatron-Core: an explicit form used for training and an absorbed form — which slashes collective communication by gathering only the compressed latent — that is fully implemented but hard-asserted out of training (the forward opens with "assert not (self.training and self.cache_mla_latents)"), allowed only in inference decode. The library documents no reason. We show the restriction is well-founded and quantify why: ported to training, the absorbed form is a memory trap — its intermediates live in n_h x d_kv dimensions per token, larger than the per-head K/V they replace — inflating activation memory by 20-34%, up to 9.2 GB at DeepSeek-V3 scale (n_h=128, seq=16384, SP=8, eager kernel; the gap widens to 19.2 GB under a fused kernel), enough to change device-fit. This measurement, validated on two axes (linear in seq and n_h) and cross-verified on NVIDIA A100, explains the otherwise-undocumented restriction and leaves practitioners with no low-communication MLA training path. We then provide one. LAGA (Latent All-Gather Attention) keeps the absorbed form's latent-gather communication but rejects the absorb reformulation, instead reconstructing per-head K/V locally from the gathered latent. On 8x Ascend 910B at real DeepSeek-V3 dimensions, LAGA cuts collective communication 1.98x, matches explicit memory within 0.5%, is bit-identical to explicit at SP=1 and equivalent to within 1e-3 at SP=2-8, and under a fused attention kernel improves attention-block throughput 1.04-1.06x single-node and 1.07-1.24x cross-node — leading at all sequence lengths in the cross-node regime MLA is deployed for.
Real-time multimodal applications, including voice agents and interactive video generation, compose heterogeneous models into pipelines whose efficient deployment requires application-specific decisions about placement, streaming, and intra-model parallelism. Existing serving systems and auto-parallelism compilers commit to limited transformations and fixed workload assumptions, so achieving high performance on a new application requires hand-crafting an efficient implementation. We present FlashRT, an agent harness that guides coding agents to lift simple developer-written reference implementations into optimized multi-GPU deployments that flexibly weigh target metrics like latency and throughput. Using a new chain-of-program paradigm, FlashRT directs a generic coding agent through a multi-pass transformation process where an agent transforms the reference into an intermediate representation (IR) to capture data dependencies and persistent-state scopes, validates this IR via a sequential interpreter, and performs static analyses to identify candidate transformations. Then, the agent iteratively implements, verifies, and benchmarks each candidate under a measurement-gated optimization loop to produce effective deployments that span different hardware budgets. Across various applications, including video world models and multimodal LLMs, FlashRT converts reference implementations into highly efficient deployments, delivering up to ~70x latency reduction and 2.8x throughput improvement on NVIDIA B200 GPUs. On AMD MI355X GPUs, FlashRT matches the peak latency reduction while increasing peak throughput improvement to 3.6x, demonstrating that agent-driven optimization can be more scalable on platforms with less mature expert optimization. In fact, for Qwen3-Omni text-to-audio inference, FlashRT reduces response latency by 65% compared to the expert vLLM-Omni implementation on AMD MI355X.
展开完整摘要收起摘要↓
Real-time multimodal applications, including voice agents and interactive video generation, compose heterogeneous models into pipelines whose efficient deployment requires application-specific decisions about placement, streaming, and intra-model parallelism. Existing serving systems and auto-parallelism compilers commit to limited transformations and fixed workload assumptions, so achieving high performance on a new application requires hand-crafting an efficient implementation. We present FlashRT, an agent harness that guides coding agents to lift simple developer-written reference implementations into optimized multi-GPU deployments that flexibly weigh target metrics like latency and throughput. Using a new chain-of-program paradigm, FlashRT directs a generic coding agent through a multi-pass transformation process where an agent transforms the reference into an intermediate representation (IR) to capture data dependencies and persistent-state scopes, validates this IR via a sequential interpreter, and performs static analyses to identify candidate transformations. Then, the agent iteratively implements, verifies, and benchmarks each candidate under a measurement-gated optimization loop to produce effective deployments that span different hardware budgets. Across various applications, including video world models and multimodal LLMs, FlashRT converts reference implementations into highly efficient deployments, delivering up to ~70x latency reduction and 2.8x throughput improvement on NVIDIA B200 GPUs. On AMD MI355X GPUs, FlashRT matches the peak latency reduction while increasing peak throughput improvement to 3.6x, demonstrating that agent-driven optimization can be more scalable on platforms with less mature expert optimization. In fact, for Qwen3-Omni text-to-audio inference, FlashRT reduces response latency by 65% compared to the expert vLLM-Omni implementation on AMD MI355X.
作者Xiangyu Chen, Jixiang Luo, Yuankai Fan, Haibin Huang, Chi Zhang, Xuelong Li
Under the AI Flow framework, communication is shifting from transmitting fidelity-oriented information flows toward delivering task-oriented and perception-oriented token flows across heterogeneous network resources. Video communication is a fundamental component of modern information networks. However, under ultra-low-bandwidth and weak-network conditions, conventional video coding and transmission methods, which are primarily optimized for pixel-level fidelity, often struggle to balance visual usability, transmission efficiency, and robustness to unstable links. With the rapid advancement of generativemodels, video communication is also moving from precise signal reconstruction toward receiver-side perceptual utility and system-level usability. In this paper, we propose Generative Transmission (GenTrans) for video communication under ultra-low-bandwidth and weak-network conditions. Built upon Generative Video Compression (GVC), GenTrans formulates video transmission as a joint optimization problem involving bandwidth, computation, and memory, rather than treating it merely as a signal coding task. By leveraging generative priors, cross-clip memory reuse, runtime state reuse, and weak-network-aware transport, GenTrans significantly reduces transmission overhead while enabling visually coherent and practically useful reconstruction. Experimental results show that GenTrans supports effective video transmission under ultra-low-bitrate and weak-network conditions, achieving improved transmission efficiency, decoding efficiency, and robustness while preserving perceptual quality.
展开完整摘要收起摘要↓
Under the AI Flow framework, communication is shifting from transmitting fidelity-oriented information flows toward delivering task-oriented and perception-oriented token flows across heterogeneous network resources. Video communication is a fundamental component of modern information networks. However, under ultra-low-bandwidth and weak-network conditions, conventional video coding and transmission methods, which are primarily optimized for pixel-level fidelity, often struggle to balance visual usability, transmission efficiency, and robustness to unstable links. With the rapid advancement of generativemodels, video communication is also moving from precise signal reconstruction toward receiver-side perceptual utility and system-level usability. In this paper, we propose Generative Transmission (GenTrans) for video communication under ultra-low-bandwidth and weak-network conditions. Built upon Generative Video Compression (GVC), GenTrans formulates video transmission as a joint optimization problem involving bandwidth, computation, and memory, rather than treating it merely as a signal coding task. By leveraging generative priors, cross-clip memory reuse, runtime state reuse, and weak-network-aware transport, GenTrans significantly reduces transmission overhead while enabling visually coherent and practically useful reconstruction. Experimental results show that GenTrans supports effective video transmission under ultra-low-bitrate and weak-network conditions, achieving improved transmission efficiency, decoding efficiency, and robustness while preserving perceptual quality.
作者Mateusz Piechocki, Alessandro Capotondi, Marek Kraft
On-device model adaptation is essential to enable lifelong personalization on resource-constrained hardware, but compute, power, and memory limitations of such devices make end-to-end backpropagation impractical for modern deep neural networks. This work proposes a heterogeneous adaptation pipeline that repurposes a commercial edge AI inference accelerator, Hailo-8L, for frozen-backbone feature extraction during on-device training. The computational graph is partitioned so that the pre-trained backbone is quantized to INT8 and run on the accelerator, while only a lightweight FP32 classification head is fine-tuned on the host CPU, enabling frequent, energy-efficient in-field updates with most weights remaining fixed. Across multiple architectures and datasets, this pipeline achieves up to 15.4x faster wall-clock training time compared to a Raspberry Pi 5 CPU baseline, offers competitive throughput in favorable settings, and consistently reduces energy per sample. Post-training quantization restoration is shown to be crucial for preserving the quality of accelerator-generated features and mitigating accuracy loss in quantization-sensitive architectures. Overall, the results demonstrate a practical approach to efficient on-device adaptation using inference-oriented edge accelerators. The implementation is available at https://github.com/MatPiech/accelerator-training.
展开完整摘要收起摘要↓
On-device model adaptation is essential to enable lifelong personalization on resource-constrained hardware, but compute, power, and memory limitations of such devices make end-to-end backpropagation impractical for modern deep neural networks. This work proposes a heterogeneous adaptation pipeline that repurposes a commercial edge AI inference accelerator, Hailo-8L, for frozen-backbone feature extraction during on-device training. The computational graph is partitioned so that the pre-trained backbone is quantized to INT8 and run on the accelerator, while only a lightweight FP32 classification head is fine-tuned on the host CPU, enabling frequent, energy-efficient in-field updates with most weights remaining fixed. Across multiple architectures and datasets, this pipeline achieves up to 15.4x faster wall-clock training time compared to a Raspberry Pi 5 CPU baseline, offers competitive throughput in favorable settings, and consistently reduces energy per sample. Post-training quantization restoration is shown to be crucial for preserving the quality of accelerator-generated features and mitigating accuracy loss in quantization-sensitive architectures. Overall, the results demonstrate a practical approach to efficient on-device adaptation using inference-oriented edge accelerators. The implementation is available at https://github.com/MatPiech/accelerator-training.
As large language models (LLMs) process increasingly longer prompts, computation and KV-cache memory costs have emerged as major bottlenecks in inference systems. Existing input-level prompt compression methods address this, but rank each sentence by a scalar relevance score, treating the document as an unstructured pool of words and sentences. Under tight budgets, this causes theme collapse, where the dominant theme(s) of a document consumes the budget, discarding less-frequent yet task-relevant themes. Preserving thematic coverage instead requires allocating the budget across recurring themes rather than scoring sentences in isolation. To this end, we propose SALT, a model-agnostic extractive framework that organizes per-sentence keywords into a trie ordered by sentence frequency (SF), a lightweight, reusable proxy for document thematic structure. This trie-based organization smooths memory allocation and prevents dominant themes from monopolizing the budget. Multi-anchor retrieval activates trie nodes labeled by query keywords at any depth, and the trie persists across dialogue turns, supporting multi-turn use without re-encoding the document. By preserving document themes, SALT reduces the prefill computation and memory cost of long-context prompts while remaining composable with KV-cache methods that target decoding-time latency and memory.
展开完整摘要收起摘要↓
As large language models (LLMs) process increasingly longer prompts, computation and KV-cache memory costs have emerged as major bottlenecks in inference systems. Existing input-level prompt compression methods address this, but rank each sentence by a scalar relevance score, treating the document as an unstructured pool of words and sentences. Under tight budgets, this causes theme collapse, where the dominant theme(s) of a document consumes the budget, discarding less-frequent yet task-relevant themes. Preserving thematic coverage instead requires allocating the budget across recurring themes rather than scoring sentences in isolation. To this end, we propose SALT, a model-agnostic extractive framework that organizes per-sentence keywords into a trie ordered by sentence frequency (SF), a lightweight, reusable proxy for document thematic structure. This trie-based organization smooths memory allocation and prevents dominant themes from monopolizing the budget. Multi-anchor retrieval activates trie nodes labeled by query keywords at any depth, and the trie persists across dialogue turns, supporting multi-turn use without re-encoding the document. By preserving document themes, SALT reduces the prefill computation and memory cost of long-context prompts while remaining composable with KV-cache methods that target decoding-time latency and memory.
LLMs scale Mixture-of-Experts (MoE) parameters for superior intelligence, but massive weights and dynamic computation impede efficient serving. Existing instance-level prefill-decode disaggregation isolates the phases on separate full-model replicas. As MoE weights grow, each instance may span tens to hundreds of GPUs, making resource allocation increasingly coarse. Configured prefill-to-decode ratios thus often mismatch demand, overprovisioning one phase while overloading the other. Prefill-decode colocation avoids this duplication, but existing Green Context solutions partition each GPU by phase and fix phase resources during a kernel. They cannot track resource changes across operations or layerwise variation in routed expert load, causing head-of-line blocking or idle reserved resources. Partitioning every GPU also leaves each phase with fewer local resources, forces wider parallelism and more communication, and lets prefill and decode traffic interfere on the shared network. We present ExpertPlex, which shares massive MoE experts across phases while disaggregating lightweight attention modules. Expert sharing eliminates over 95% of duplicate model weights and multiplexes dynamically sparse computation, while attention disaggregation reduces attention communication cost. ExpertPlex further uses (1) adaptive persistent kernels to schedule dynamic expert computation at tile granularity for efficient, isolated execution; (2) attention-initiated MoE communication to avoid network interference and enable cross-phase communication-computation overlap; and (3) a tile-to-cluster model to optimize these mechanisms for maximum goodput. Experiments serving MiniMax-M2.7 and GLM-5.1-FP8 show that ExpertPlex improves goodput by up to 2.01$\times$ over instance-level prefill-decode disaggregation and 1.66$\times$ over prefill-decode colocation.
展开完整摘要收起摘要↓
LLMs scale Mixture-of-Experts (MoE) parameters for superior intelligence, but massive weights and dynamic computation impede efficient serving. Existing instance-level prefill-decode disaggregation isolates the phases on separate full-model replicas. As MoE weights grow, each instance may span tens to hundreds of GPUs, making resource allocation increasingly coarse. Configured prefill-to-decode ratios thus often mismatch demand, overprovisioning one phase while overloading the other. Prefill-decode colocation avoids this duplication, but existing Green Context solutions partition each GPU by phase and fix phase resources during a kernel. They cannot track resource changes across operations or layerwise variation in routed expert load, causing head-of-line blocking or idle reserved resources. Partitioning every GPU also leaves each phase with fewer local resources, forces wider parallelism and more communication, and lets prefill and decode traffic interfere on the shared network. We present ExpertPlex, which shares massive MoE experts across phases while disaggregating lightweight attention modules. Expert sharing eliminates over 95% of duplicate model weights and multiplexes dynamically sparse computation, while attention disaggregation reduces attention communication cost. ExpertPlex further uses (1) adaptive persistent kernels to schedule dynamic expert computation at tile granularity for efficient, isolated execution; (2) attention-initiated MoE communication to avoid network interference and enable cross-phase communication-computation overlap; and (3) a tile-to-cluster model to optimize these mechanisms for maximum goodput. Experiments serving MiniMax-M2.7 and GLM-5.1-FP8 show that ExpertPlex improves goodput by up to 2.01$\times$ over instance-level prefill-decode disaggregation and 1.66$\times$ over prefill-decode colocation.
Block-wise diffusion large language models (dLLMs) decode sequentially at the block level, enabling effective KV-cache reuse across blocks but making inter-block decoding strictly serial. Prior work has attempted to unlock inter-block parallelism through post-training methods, but achieves only modest speedups and often degrades accuracy. We observe that self-correcting dLLMs offer a training-free alternative: token-to-token (T2T) editing can repair tokens drafted with a slightly stale upstream context, so a downstream block requires only an informative draft rather than a finalized predecessor. This turns block finality from a hard dependency into a scheduling resource. We propose \textbf{\flowblock{}}, a training-free parallel decoding framework built on two mechanisms. (i) Gated Wavefront Decoding admits blocks into a bounded wavefront only when a readiness gate is satisfied, jointly refines active blocks via T2T editing, and commits blocks in order under a windowed block-causal mask that preserves exact frozen-prefix KV caches reuse. (ii) Heterogeneous Wavefront Packing assigns each request an independent wavefront while packing asynchronous windows into dense, shape-stable batched forwards. Across different benchmarks, \flowblock{} improves tokens per second (TPS) over LLaDA-2.1 and LLaDA-2.0, two serial block-wise dLLMs, by up to 2.95$\times$ and 4.01$\times$, while reducing latency by up to 53.6% and 77.1%, respectively. It also improves average accuracy by 1.3 points. Compared with D2F, a training-based inter-block-parallel baseline, \flowblock{} achieves higher accuracy and up to 16$\times$ higher batched serving throughput.
展开完整摘要收起摘要↓
Block-wise diffusion large language models (dLLMs) decode sequentially at the block level, enabling effective KV-cache reuse across blocks but making inter-block decoding strictly serial. Prior work has attempted to unlock inter-block parallelism through post-training methods, but achieves only modest speedups and often degrades accuracy. We observe that self-correcting dLLMs offer a training-free alternative: token-to-token (T2T) editing can repair tokens drafted with a slightly stale upstream context, so a downstream block requires only an informative draft rather than a finalized predecessor. This turns block finality from a hard dependency into a scheduling resource. We propose \textbf{\flowblock{}}, a training-free parallel decoding framework built on two mechanisms. (i) Gated Wavefront Decoding admits blocks into a bounded wavefront only when a readiness gate is satisfied, jointly refines active blocks via T2T editing, and commits blocks in order under a windowed block-causal mask that preserves exact frozen-prefix KV caches reuse. (ii) Heterogeneous Wavefront Packing assigns each request an independent wavefront while packing asynchronous windows into dense, shape-stable batched forwards. Across different benchmarks, \flowblock{} improves tokens per second (TPS) over LLaDA-2.1 and LLaDA-2.0, two serial block-wise dLLMs, by up to 2.95$\times$ and 4.01$\times$, while reducing latency by up to 53.6% and 77.1%, respectively. It also improves average accuracy by 1.3 points. Compared with D2F, a training-based inter-block-parallel baseline, \flowblock{} achieves higher accuracy and up to 16$\times$ higher batched serving throughput.
作者Hakbeom Jang, Inho Song, Sam H. Noh, Jongryool Kim
Long-context, multi-turn, and agentic LLM workloads increasingly reuse previously processed context, making KV-cache reuse essential for reducing redundant computation. However, this reuse shifts the bottleneck to the memory tier that stores and serves reusable KV states at cluster scale. GPU HBM and host DRAM are too costly to scale to TB-scale shared context capacity, motivating remote tiers built from lower-cost, higher-capacity media. This paper presents HyMCache, a KV-cache framework that integrates CXL-hybrid memory (CXL-HM) for multi-turn LLM serving. CXL-HM combines a small amount of in-device DRAM with large SSD-backed capacity behind a CXL interface. By exploiting the read-dominant, predictable, and append-only nature of multi-turn KV-cache access, HyMCache rethinks DRAM management within CXL-HM to efficiently support TB-scale SSD-backed KV reuse. It uses request-level prefix prefetching and opportunistic write buffering to stage latency-critical reads in device DRAM, enabling DRAM-scale KV-cache efficiency at SSD-level cost. We evaluate HyMCache on a real CXL-HM prototype under both single-aggregator and PD-disaggregated serving configurations. Under the same DRAM budget, HyMCache outperforms local LMCache by 3.0x in single-node serving and 1.45x in PD-disaggregated serving. Compared with 1 TB distributed-DRAM Mooncake, HyMCache incurs about 30% lower performance but uses 16x less DRAM.
展开完整摘要收起摘要↓
Long-context, multi-turn, and agentic LLM workloads increasingly reuse previously processed context, making KV-cache reuse essential for reducing redundant computation. However, this reuse shifts the bottleneck to the memory tier that stores and serves reusable KV states at cluster scale. GPU HBM and host DRAM are too costly to scale to TB-scale shared context capacity, motivating remote tiers built from lower-cost, higher-capacity media. This paper presents HyMCache, a KV-cache framework that integrates CXL-hybrid memory (CXL-HM) for multi-turn LLM serving. CXL-HM combines a small amount of in-device DRAM with large SSD-backed capacity behind a CXL interface. By exploiting the read-dominant, predictable, and append-only nature of multi-turn KV-cache access, HyMCache rethinks DRAM management within CXL-HM to efficiently support TB-scale SSD-backed KV reuse. It uses request-level prefix prefetching and opportunistic write buffering to stage latency-critical reads in device DRAM, enabling DRAM-scale KV-cache efficiency at SSD-level cost. We evaluate HyMCache on a real CXL-HM prototype under both single-aggregator and PD-disaggregated serving configurations. Under the same DRAM budget, HyMCache outperforms local LMCache by 3.0x in single-node serving and 1.45x in PD-disaggregated serving. Compared with 1 TB distributed-DRAM Mooncake, HyMCache incurs about 30% lower performance but uses 16x less DRAM.
作者Chuheng Du, Junyi Chen, Hanlin Tang, Kan Liu, Tao Lan, Lin Qu, Chaoyue Niu, Shengzhong Liu, Guihai Chen, Fan Wu
Long-context inference is central to modern large language model (LLM) applications such as retrieval-augmented generation and multi-document reasoning. To mitigate the growing inference cost, recent work has explored key-value (KV) cache reuse to reduce redundant prefill computation. However, existing reuse methods primarily focus on computation savings and overlook a critical bottleneck in long-context LLM serving: the cost of storing and accessing large KV caches. While KV compression appears to be a natural complement, naively combining compression with non-prefix KV reuse often leads to severe accuracy degradation. In this work, we propose C$^2$KV, a unified framework for non-prefix KV reuse that jointly optimizes KV extraction and inference-time concatenation. C$^2$KV learns a composable and compressed KV cache manifold that is explicitly designed to be position-agnostic. Our approach introduces a lightweight sidecar Extractor with learnable compression tokens and a structured attention flow, enabling modular KV representations that can be flexibly reused and concatenated without modifying the frozen base model. We further employ a compression-concatenation co-training strategy to align extraction-time representations with their downstream reuse behavior. Extensive experiments across multiple long-context benchmarks and model families demonstrate that C$^2$KV significantly reduces KV cache storage and transfer costs, achieving up to 17$\times$ inference speedup under long contexts, while preserving generation quality.
展开完整摘要收起摘要↓
Long-context inference is central to modern large language model (LLM) applications such as retrieval-augmented generation and multi-document reasoning. To mitigate the growing inference cost, recent work has explored key-value (KV) cache reuse to reduce redundant prefill computation. However, existing reuse methods primarily focus on computation savings and overlook a critical bottleneck in long-context LLM serving: the cost of storing and accessing large KV caches. While KV compression appears to be a natural complement, naively combining compression with non-prefix KV reuse often leads to severe accuracy degradation. In this work, we propose C$^2$KV, a unified framework for non-prefix KV reuse that jointly optimizes KV extraction and inference-time concatenation. C$^2$KV learns a composable and compressed KV cache manifold that is explicitly designed to be position-agnostic. Our approach introduces a lightweight sidecar Extractor with learnable compression tokens and a structured attention flow, enabling modular KV representations that can be flexibly reused and concatenated without modifying the frozen base model. We further employ a compression-concatenation co-training strategy to align extraction-time representations with their downstream reuse behavior. Extensive experiments across multiple long-context benchmarks and model families demonstrate that C$^2$KV significantly reduces KV cache storage and transfer costs, achieving up to 17$\times$ inference speedup under long contexts, while preserving generation quality.