Agentic engineering systems have shown strong performance on general-purpose benchmarks, yet their effectiveness in enterprise resource planning (ERP) domain-specific languages (DSLs) remains underexplored. We introduce BC-Bench, a benchmark designed to evaluate agentic engineering on real-world tasks in AL, the DSL for Microsoft Dynamics 365 Business Central. BC-Bench comprises 101 manually curated tasks extracted from two Microsoft-owned production repositories, reflecting authentic ERP development workflows. Adapting the SWE-Bench methodology, we address the unique constraints of the AL ecosystem---including limited public resources and complex environment provisioning. Beyond generating functional code, BC-Bench evaluates test generation and supports multimodal problem statements where visual context is commonly present. We evaluate multiple frontier models across two agent harnesses, utilizing multi-run metrics to account for nondeterminism. In the Bug Fixing category, under our evaluated settings, between-model differences in resolution rate are larger than differences between the two evaluated agent harnesses, and improvements reported on general-purpose benchmarks do not consistently transfer to AL. These results highlight the need for domain-specific evaluation.
展开完整摘要收起摘要↓
Agentic engineering systems have shown strong performance on general-purpose benchmarks, yet their effectiveness in enterprise resource planning (ERP) domain-specific languages (DSLs) remains underexplored. We introduce BC-Bench, a benchmark designed to evaluate agentic engineering on real-world tasks in AL, the DSL for Microsoft Dynamics 365 Business Central. BC-Bench comprises 101 manually curated tasks extracted from two Microsoft-owned production repositories, reflecting authentic ERP development workflows. Adapting the SWE-Bench methodology, we address the unique constraints of the AL ecosystem---including limited public resources and complex environment provisioning. Beyond generating functional code, BC-Bench evaluates test generation and supports multimodal problem statements where visual context is commonly present. We evaluate multiple frontier models across two agent harnesses, utilizing multi-run metrics to account for nondeterminism. In the Bug Fixing category, under our evaluated settings, between-model differences in resolution rate are larger than differences between the two evaluated agent harnesses, and improvements reported on general-purpose benchmarks do not consistently transfer to AL. These results highlight the need for domain-specific evaluation.
Animating articulated 3D meshes via text requires satisfying strict kinematic constraints, modeling causal interactions between parts, and achieving instruction fidelity. Due to the absence of task-specific training data and explicit articulation supervision, existing data-driven mesh animation methods are largely inapplicable to this setting. To address this, we propose ArtiMo, a novel agent-driven framework for text-guided articulated mesh animation. Operating in a zero-shot manner, ArtiMo develops an agentic pipeline powered by Large Language and Vision-Language Models (LLMs/VLMs) to orchestrate motion generation. By synergizing the explicit kinematic constraints of URDF with the agent's reasoning and planning capabilities, it effectively produces causally coherent part motions and interactions without requiring model fine-tuning. To ensure motion correctness, the agent additionally utilizes a visual self-improvement mechanism: generated animations are rendered into compact keyframes and motion cues, enabling the VLM to iteratively diagnose and correct errors. Furthermore, we contribute a new benchmark dataset spanning 21 articulated object categories, featuring high-quality motion annotations enriched with causal relationships. Extensive experiments demonstrate that ArtiMo significantly outperforms baselines, particularly on complex, causally driven motions. The project page is available at https://zou-2004.github.io/ArtiMo/.
展开完整摘要收起摘要↓
Animating articulated 3D meshes via text requires satisfying strict kinematic constraints, modeling causal interactions between parts, and achieving instruction fidelity. Due to the absence of task-specific training data and explicit articulation supervision, existing data-driven mesh animation methods are largely inapplicable to this setting. To address this, we propose ArtiMo, a novel agent-driven framework for text-guided articulated mesh animation. Operating in a zero-shot manner, ArtiMo develops an agentic pipeline powered by Large Language and Vision-Language Models (LLMs/VLMs) to orchestrate motion generation. By synergizing the explicit kinematic constraints of URDF with the agent's reasoning and planning capabilities, it effectively produces causally coherent part motions and interactions without requiring model fine-tuning. To ensure motion correctness, the agent additionally utilizes a visual self-improvement mechanism: generated animations are rendered into compact keyframes and motion cues, enabling the VLM to iteratively diagnose and correct errors. Furthermore, we contribute a new benchmark dataset spanning 21 articulated object categories, featuring high-quality motion annotations enriched with causal relationships. Extensive experiments demonstrate that ArtiMo significantly outperforms baselines, particularly on complex, causally driven motions. The project page is available at https://zou-2004.github.io/ArtiMo/.
Vision-Language-Action (VLA) models have emerged as a powerful paradigm for end-to-end autonomous driving by jointly integrating perception, reasoning, and decision making within a unified multimodal framework. However, most existing VLA models formulate end-to-end autonomous driving as a visual question answering task, leading to unreliable and less interpretable decision reasoning. In addition, they fail to establish effective multi-modal interaction across heterogeneous sensors, thereby limiting robust scene perception and reliable driving reasoning in long-tail driving scenarios. To this end, we propose a robust VLA-based end-to-end autonomous driving system that combines multi-modality interaction with multi-trajectory planning and optimization, enabling more reliable, interpretable, and safer driving decisions. Our method comprises three core components: (1) Affinity-Guided Optimal Transport for main-auxiliary modality two-way interaction; (2) Distribution-Consistent Modality Transfer for heterogeneous modality distribution transfer and cross-modal interaction; (3) Multi-modal Multi-Trajectory Planning along with Perception-Oriented Trajectory Refinement for better driving decisions to long-tail driving scenarios. Experimental results in open-loop and closed-loop datasets demonstrate improvements in safety long-horizon driving reasoning and road scene perception over existing driving systems, highlighting the ability of our mutli-modality interaction and multi-trajectory planning and optimization for scalable VLA-based systems.
展开完整摘要收起摘要↓
Vision-Language-Action (VLA) models have emerged as a powerful paradigm for end-to-end autonomous driving by jointly integrating perception, reasoning, and decision making within a unified multimodal framework. However, most existing VLA models formulate end-to-end autonomous driving as a visual question answering task, leading to unreliable and less interpretable decision reasoning. In addition, they fail to establish effective multi-modal interaction across heterogeneous sensors, thereby limiting robust scene perception and reliable driving reasoning in long-tail driving scenarios. To this end, we propose a robust VLA-based end-to-end autonomous driving system that combines multi-modality interaction with multi-trajectory planning and optimization, enabling more reliable, interpretable, and safer driving decisions. Our method comprises three core components: (1) Affinity-Guided Optimal Transport for main-auxiliary modality two-way interaction; (2) Distribution-Consistent Modality Transfer for heterogeneous modality distribution transfer and cross-modal interaction; (3) Multi-modal Multi-Trajectory Planning along with Perception-Oriented Trajectory Refinement for better driving decisions to long-tail driving scenarios. Experimental results in open-loop and closed-loop datasets demonstrate improvements in safety long-horizon driving reasoning and road scene perception over existing driving systems, highlighting the ability of our mutli-modality interaction and multi-trajectory planning and optimization for scalable VLA-based systems.
LLM agents are emerging as an important paradigm for real-world tasks that require reasoning, tool use, and sequential decision-making. As these agents operate over longer horizons, runtime intervention offers a way to improve reliability without retraining the underlying actor. Failure detection alone is insufficient. Effective intervention must also provide a useful direction for recovery. Existing approaches often rely on an expert solver or a critic that generates task-specific corrections, incurring either the cost of another capable solver or the capacity demands of a task-capable critic. We introduce Comparison-Only Tiny Advisor (COTA), a comparison-only framework for constructive runtime intervention. In COTA, a tiny comparator judges whether sampled alternatives lead to better continuations than the actor's proposal, and repeated comparisons determine when intervention is warranted. We train the comparator using pairwise supervision constructed from same-prefix counterfactual branches. Preferred alternatives are returned as non-binding advice, leaving the original actor to replan. Across WebShop, ALFWorld, and tau^3-Retail with three actors, COTA improves all nine evaluation settings and outperforms the compared baselines. These results show that constructive runtime intervention can remain effective even when the auxiliary model has substantially weaker task-solving capability than the actor.
展开完整摘要收起摘要↓
LLM agents are emerging as an important paradigm for real-world tasks that require reasoning, tool use, and sequential decision-making. As these agents operate over longer horizons, runtime intervention offers a way to improve reliability without retraining the underlying actor. Failure detection alone is insufficient. Effective intervention must also provide a useful direction for recovery. Existing approaches often rely on an expert solver or a critic that generates task-specific corrections, incurring either the cost of another capable solver or the capacity demands of a task-capable critic. We introduce Comparison-Only Tiny Advisor (COTA), a comparison-only framework for constructive runtime intervention. In COTA, a tiny comparator judges whether sampled alternatives lead to better continuations than the actor's proposal, and repeated comparisons determine when intervention is warranted. We train the comparator using pairwise supervision constructed from same-prefix counterfactual branches. Preferred alternatives are returned as non-binding advice, leaving the original actor to replan. Across WebShop, ALFWorld, and tau^3-Retail with three actors, COTA improves all nine evaluation settings and outperforms the compared baselines. These results show that constructive runtime intervention can remain effective even when the auxiliary model has substantially weaker task-solving capability than the actor.
Evaluating language-guided mobile agents has recently shifted from rule-based to model-based approaches to achieve scalable and automated assessments. However, existing holistic evaluation paradigms process entire trajectories at once, leading to substantial context overload. Moreover, they primarily focus on task completion while overlooking operational safety. To address these limitations, we introduce CRATE, a novel two-stage VLM-as-judge framework for automated mobile agent evaluation that is compatible with both open- and closed-source models. Leveraging a step-level consequence reasoning mechanism, CRATE independently extracts task-relevant visual clues and infers action-conditioned state changes at each step. The resulting step-level textual evidence is then synthesized through trajectory-level aggregation to deliver an evidence-grounded evaluation of task completion. Building upon this evaluation scheme, we further extend CRATE to CRATE-S for operational safety assessment. Extensive experiments validate the effectiveness and robustness of both CRATE and CRATE-S. Powered by Qwen2.5-VL-72B-Instruct, CRATE achieves an F1-score of 0.833 on AndroidWorld (outperforming SPA-Bench by 20%), while CRATE-S reaches an F1-score of 0.697 on MobileRisk, demonstrating strong alignment with benchmark ground truths. Code is available at https://anonymous.4open.science/r/CRATE-D580.
展开完整摘要收起摘要↓
Evaluating language-guided mobile agents has recently shifted from rule-based to model-based approaches to achieve scalable and automated assessments. However, existing holistic evaluation paradigms process entire trajectories at once, leading to substantial context overload. Moreover, they primarily focus on task completion while overlooking operational safety. To address these limitations, we introduce CRATE, a novel two-stage VLM-as-judge framework for automated mobile agent evaluation that is compatible with both open- and closed-source models. Leveraging a step-level consequence reasoning mechanism, CRATE independently extracts task-relevant visual clues and infers action-conditioned state changes at each step. The resulting step-level textual evidence is then synthesized through trajectory-level aggregation to deliver an evidence-grounded evaluation of task completion. Building upon this evaluation scheme, we further extend CRATE to CRATE-S for operational safety assessment. Extensive experiments validate the effectiveness and robustness of both CRATE and CRATE-S. Powered by Qwen2.5-VL-72B-Instruct, CRATE achieves an F1-score of 0.833 on AndroidWorld (outperforming SPA-Bench by 20%), while CRATE-S reaches an F1-score of 0.697 on MobileRisk, demonstrating strong alignment with benchmark ground truths. Code is available at https://anonymous.4open.science/r/CRATE-D580.
As large language model (LLM) agents move from conversation to executing code, reading local files, and orchestrating external tools, a single agent hijacked by a malicious third-party skill can cause data exfiltration, privilege escalation, or cascading compromise. We argue that agentic risk is progressive: it can enter at four loci of the agent control loop--skill admission, invocation-time intent, execution-time effect, and post-action consequence--while a denied dangerous objective can reappear across surface forms, tools, or turns; existing safeguards are typically local to one lifecycle boundary or one call. Guided by this threat model, we present ClawSentry, an open-source, framework-agnostic security supervision gateway for agent runtimes. Before a skill package is ever executed, First-use Skill Package Review (FSPR) audits it under a deterministic evidence floor, escalating unresolved cases to bounded read-only agentic review (locus A). At runtime, a three-tier progressive decision engine--a deterministic L1 layer, a rule-anchored L2 semantic reviewer, and a read-only L3 evidence-seeking agent--spends contextual review only on the residual ambiguity, while a session-level anti-bypass mechanism recognizes tool-switching and rephrased retries (loci B--C); a post-action path feeds high-severity evidence non-retroactively into later review (locus D). An Agent Harness Protocol (AHP) abstraction applies one policy across Codex, Claude Code, Kimi CLI, and Gemini CLI without modifying agent internals. On SkillInject with Codex/GPT-5.4, contextual ASR falls from 39.55% to 2.61% while contextual TSR moves only from 83.78% to 83.05%. Across five Work Agents on the full SkillsSafety benchmark, ClawSentry confines ASR to 9.09--15.03% from 33.5--49.7% unprotected, and aggregate TSR on clean skills remains 98.7%.
展开完整摘要收起摘要↓
As large language model (LLM) agents move from conversation to executing code, reading local files, and orchestrating external tools, a single agent hijacked by a malicious third-party skill can cause data exfiltration, privilege escalation, or cascading compromise. We argue that agentic risk is progressive: it can enter at four loci of the agent control loop--skill admission, invocation-time intent, execution-time effect, and post-action consequence--while a denied dangerous objective can reappear across surface forms, tools, or turns; existing safeguards are typically local to one lifecycle boundary or one call. Guided by this threat model, we present ClawSentry, an open-source, framework-agnostic security supervision gateway for agent runtimes. Before a skill package is ever executed, First-use Skill Package Review (FSPR) audits it under a deterministic evidence floor, escalating unresolved cases to bounded read-only agentic review (locus A). At runtime, a three-tier progressive decision engine--a deterministic L1 layer, a rule-anchored L2 semantic reviewer, and a read-only L3 evidence-seeking agent--spends contextual review only on the residual ambiguity, while a session-level anti-bypass mechanism recognizes tool-switching and rephrased retries (loci B--C); a post-action path feeds high-severity evidence non-retroactively into later review (locus D). An Agent Harness Protocol (AHP) abstraction applies one policy across Codex, Claude Code, Kimi CLI, and Gemini CLI without modifying agent internals. On SkillInject with Codex/GPT-5.4, contextual ASR falls from 39.55% to 2.61% while contextual TSR moves only from 83.78% to 83.05%. Across five Work Agents on the full SkillsSafety benchmark, ClawSentry confines ASR to 9.09--15.03% from 33.5--49.7% unprotected, and aggregate TSR on clean skills remains 98.7%.
Agents learn to act through interaction with environments, yet the environments used for training are often manually constructed or synthesized around predefined tasks and benchmarks. This task-centric paradigm makes it difficult to scale environments that reflect realistic and evolving workflows where diverse tasks can naturally emerge from the underlying world. We introduce AgentMercury, a scalable framework for synthesizing executable environments from high-level business scenarios. Rather than constructing an environment for a specific task, AgentMercury first instantiates a persistent world with entities, services, tools, state, and executable cross-service invariants, from which diverse tasks and interaction trajectories can subsequently emerge. We construct 4,783 executable environments spanning 14 industries and 50 countries, and use them as training substrates for reinforcement learning. Despite being generated without targeting the evaluation benchmarks, policies trained on these business-oriented environments improve substantially on both enterprise workflows and out-of-domain benchmarks spanning reasoning, coding, scientific computing, and tool use. In our experiments, Qwen3.5-4B improves from 12.3 to 15.7 on EnterpriseOps-GYM and from 45.9 to 56.0 on AIME26 after training on AgentMercury environments. We further show that the construction process itself can be learned: fine-tuning Qwen3.5-35B-A3B on construction traces increases executable-world authoring success from 3.3% to 83.3% on held-out business scenarios. These results show that scenario-grounded environments can provide useful and generalizable learning signals beyond benchmark-specific training, while their construction can itself become a learnable capability.
展开完整摘要收起摘要↓
Agents learn to act through interaction with environments, yet the environments used for training are often manually constructed or synthesized around predefined tasks and benchmarks. This task-centric paradigm makes it difficult to scale environments that reflect realistic and evolving workflows where diverse tasks can naturally emerge from the underlying world. We introduce AgentMercury, a scalable framework for synthesizing executable environments from high-level business scenarios. Rather than constructing an environment for a specific task, AgentMercury first instantiates a persistent world with entities, services, tools, state, and executable cross-service invariants, from which diverse tasks and interaction trajectories can subsequently emerge. We construct 4,783 executable environments spanning 14 industries and 50 countries, and use them as training substrates for reinforcement learning. Despite being generated without targeting the evaluation benchmarks, policies trained on these business-oriented environments improve substantially on both enterprise workflows and out-of-domain benchmarks spanning reasoning, coding, scientific computing, and tool use. In our experiments, Qwen3.5-4B improves from 12.3 to 15.7 on EnterpriseOps-GYM and from 45.9 to 56.0 on AIME26 after training on AgentMercury environments. We further show that the construction process itself can be learned: fine-tuning Qwen3.5-35B-A3B on construction traces increases executable-world authoring success from 3.3% to 83.3% on held-out business scenarios. These results show that scenario-grounded environments can provide useful and generalizable learning signals beyond benchmark-specific training, while their construction can itself become a learnable capability.
As scientific literature grows and papers increasingly under-report limitations, multi-agent LLMs offer a promising approach to systematically uncover these hidden failure modes. Here, we introduce Tree-of-Concerns, a multi-agent framework that deploys specialized skeptic personas, each operating through a category-specific analytical lens, as parallel debate trees to extract unstated limitations from scientific papers. Each persona conducts structured, evidence-grounded argumentation, while a Panel Review mechanism re-evaluates each surviving claim from all five perspectives to correct category drift and severity miscalibration. Through experiments on ToC-Bench, our benchmark of 414 research papers with 1,905 unstated limitations, sourced from reviewer-reported weaknesses and follow-up citation critiques, we demonstrate that ToC improves precision by 79% and coverage by 11% relative to strongest baselines, surfacing specific, evidence-grounded concerns that support reviewers in systematic evaluation.
展开完整摘要收起摘要↓
As scientific literature grows and papers increasingly under-report limitations, multi-agent LLMs offer a promising approach to systematically uncover these hidden failure modes. Here, we introduce Tree-of-Concerns, a multi-agent framework that deploys specialized skeptic personas, each operating through a category-specific analytical lens, as parallel debate trees to extract unstated limitations from scientific papers. Each persona conducts structured, evidence-grounded argumentation, while a Panel Review mechanism re-evaluates each surviving claim from all five perspectives to correct category drift and severity miscalibration. Through experiments on ToC-Bench, our benchmark of 414 research papers with 1,905 unstated limitations, sourced from reviewer-reported weaknesses and follow-up citation critiques, we demonstrate that ToC improves precision by 79% and coverage by 11% relative to strongest baselines, surfacing specific, evidence-grounded concerns that support reviewers in systematic evaluation.
Multi-agent pipelines that author formal documents must both read a requester's forms and write against them. We report a deployed tender-response system, running an open-weights model under sovereignty constraints, and evaluate it against human-written bids the same organisation actually submitted. On a blind comparison where the system had no worked example available, an LLM judge rated its answers at least as good as the human-submitted answer on $40$ of $55$ ground-truth sections, better on $4$, missing on none, and flagged one unsupported claim in total. Classifying every gap the judge identified shows that $68%$ were content absent from the system's own sources — knowledge the human author held and the pipeline was never given — so only $6$ of the $15$ adverse verdicts involve a deficiency the system could have avoided. A divergence from ground truth is more often an information-availability result than a writing-quality one, and evaluations that do not separate the two understate such systems. Against this backdrop we report a conditioning asymmetry. It is well established that rendering documents as structural markup rather than flat prose improves extraction, and we reproduce that on three reading tasks. The benefit does not transfer to conditioning: converting a bid's instruction material from prose to nested XML dropped answer quality from $74%$ to $48%$ under a paired comparison. We further find that naming a forbidden construction concentrates rather than removes it — $96%$ of surviving defects fall in the two forms the prompt explicitly names — and that coupling a stochastic annotation to a deterministic windowing function moves the extracted requirement count from $68$ to $51$ on a byte-identical file. Structure belongs where the model reads; prose and self-applied tests belong where it writes.
展开完整摘要收起摘要↓
Multi-agent pipelines that author formal documents must both read a requester's forms and write against them. We report a deployed tender-response system, running an open-weights model under sovereignty constraints, and evaluate it against human-written bids the same organisation actually submitted. On a blind comparison where the system had no worked example available, an LLM judge rated its answers at least as good as the human-submitted answer on $40$ of $55$ ground-truth sections, better on $4$, missing on none, and flagged one unsupported claim in total. Classifying every gap the judge identified shows that $68%$ were content absent from the system's own sources — knowledge the human author held and the pipeline was never given — so only $6$ of the $15$ adverse verdicts involve a deficiency the system could have avoided. A divergence from ground truth is more often an information-availability result than a writing-quality one, and evaluations that do not separate the two understate such systems. Against this backdrop we report a conditioning asymmetry. It is well established that rendering documents as structural markup rather than flat prose improves extraction, and we reproduce that on three reading tasks. The benefit does not transfer to conditioning: converting a bid's instruction material from prose to nested XML dropped answer quality from $74%$ to $48%$ under a paired comparison. We further find that naming a forbidden construction concentrates rather than removes it — $96%$ of surviving defects fall in the two forms the prompt explicitly names — and that coupling a stochastic annotation to a deterministic windowing function moves the extracted requirement count from $68$ to $51$ on a byte-identical file. Structure belongs where the model reads; prose and self-applied tests belong where it writes.
Large language model (LLM) agents have demonstrated the ability to solve multi-step tasks requiring planning, tool use, and external information access, yet growing execution histories increase inference cost and expose reasoning to outdated, irrelevant, or misleading information, potentially degrading reasoning quality. Existing memory approaches organize or compress execution histories but provide limited mechanisms for deciding which memories remain active. We introduce the, a hierarchical memory system that organizes execution into tasks, subtasks, and actions while assigning each memory a dynamic retention score. Event-based updates and selection-based decay revise these scores, allowing WMT to preserve useful information, fold completed trajectories, suppress low-utility content, and retain access to folded context. We evaluate WMT on GAIA-Text using Qwen3-8B, Gemma 4 E4B, and Llama-3.1-8B, with ablations and memory-poisoning experiments. Relative to linear memory, WMT improves accuracy by an average of 9.97 percentage points while reducing prompt-token usage by 32.8%. Memory-poisoning experiments show that WMT limits the persistence and propagation of unreliable information. Our results suggest that effective long-horizon agent memory depends less on storing more information than on deciding which information should remain active.
展开完整摘要收起摘要↓
Large language model (LLM) agents have demonstrated the ability to solve multi-step tasks requiring planning, tool use, and external information access, yet growing execution histories increase inference cost and expose reasoning to outdated, irrelevant, or misleading information, potentially degrading reasoning quality. Existing memory approaches organize or compress execution histories but provide limited mechanisms for deciding which memories remain active. We introduce the, a hierarchical memory system that organizes execution into tasks, subtasks, and actions while assigning each memory a dynamic retention score. Event-based updates and selection-based decay revise these scores, allowing WMT to preserve useful information, fold completed trajectories, suppress low-utility content, and retain access to folded context. We evaluate WMT on GAIA-Text using Qwen3-8B, Gemma 4 E4B, and Llama-3.1-8B, with ablations and memory-poisoning experiments. Relative to linear memory, WMT improves accuracy by an average of 9.97 percentage points while reducing prompt-token usage by 32.8%. Memory-poisoning experiments show that WMT limits the persistence and propagation of unreliable information. Our results suggest that effective long-horizon agent memory depends less on storing more information than on deciding which information should remain active.
Quantization is widely used to deploy large language models, but its effect on uncertainty behavior, such as confidence, margins, and abstention, is rarely treated as a primary objective. We frame calibration-data selection for quantization as a target-dependent uncertainty-preservation problem. Different deployments emphasize different regions of the input distribution, yet prior work mainly optimizes accuracy-oriented compression metrics or adjusts scores after quantization. We formalize this goal with distributional and boundary preservation risks, and provide a simple mixture-mismatch argument explaining why no single calibration recipe should be expected to fit all targets. We introduce Doubt-Preserving Quantization (DPQ), a lightweight pre-quantization recipe family that uses full-precision predictions to construct target-aligned calibration mixtures of high-doubt examples and generic anchors. Across 8 language models, 9 NLP benchmarks, and 22 comparison methods, the leading fixed recipe changes with the preservation target: DPQ-r75 leads on SQuAD2 answerability-boundary preservation, while milder or single-signal variants, including DPQ-r50, confidence-only, and entropy-only, better preserve broad multiple-choice QA behavior. These results show that calibration data should be selected for the specific full-precision score behavior a deployment needs to preserve, rather than treated as a fixed quantization detail.
展开完整摘要收起摘要↓
Quantization is widely used to deploy large language models, but its effect on uncertainty behavior, such as confidence, margins, and abstention, is rarely treated as a primary objective. We frame calibration-data selection for quantization as a target-dependent uncertainty-preservation problem. Different deployments emphasize different regions of the input distribution, yet prior work mainly optimizes accuracy-oriented compression metrics or adjusts scores after quantization. We formalize this goal with distributional and boundary preservation risks, and provide a simple mixture-mismatch argument explaining why no single calibration recipe should be expected to fit all targets. We introduce Doubt-Preserving Quantization (DPQ), a lightweight pre-quantization recipe family that uses full-precision predictions to construct target-aligned calibration mixtures of high-doubt examples and generic anchors. Across 8 language models, 9 NLP benchmarks, and 22 comparison methods, the leading fixed recipe changes with the preservation target: DPQ-r75 leads on SQuAD2 answerability-boundary preservation, while milder or single-signal variants, including DPQ-r50, confidence-only, and entropy-only, better preserve broad multiple-choice QA behavior. These results show that calibration data should be selected for the specific full-precision score behavior a deployment needs to preserve, rather than treated as a fixed quantization detail.
Satellites are becoming programmable computing platforms capable of running increasingly demanding AI workloads. This shift raises a systems problem: how can AI services remain deployable, manageable, and recoverable after launch when compute capacity, connectivity, energy, and thermal headroom vary over orbital time? This paper develops a systems vision for AI infrastructure in space. We define it as the systems layer that manages AI capabilities across spacecraft, orbital networks, ground stations, and cloud backends, while treating orbital and physical state as part of the resource model. We synthesize relevant foundations from terrestrial AI infrastructure, satellite networking, and satellite edge computing, and examine the physical constraints that directly shape system design. We further ground this vision in three in-orbit case studies spanning the node, platform, and service levels. Telemetry from BUPT-1 satellite shows that usable compute capacity is bounded by thermal and energy envelopes. SateLight on BUPT-2 satellite reduces application-update transmission latency by 56.54% on average and up to 91.18%, with 100% update correctness. A stateful VLM serving case further shows that thermal interruptions make execution-state recovery a first-class systems problem. These observations motivate a research agenda for space-native resource management, lifecycle support, and sustained AI service across space and ground.
展开完整摘要收起摘要↓
Satellites are becoming programmable computing platforms capable of running increasingly demanding AI workloads. This shift raises a systems problem: how can AI services remain deployable, manageable, and recoverable after launch when compute capacity, connectivity, energy, and thermal headroom vary over orbital time? This paper develops a systems vision for AI infrastructure in space. We define it as the systems layer that manages AI capabilities across spacecraft, orbital networks, ground stations, and cloud backends, while treating orbital and physical state as part of the resource model. We synthesize relevant foundations from terrestrial AI infrastructure, satellite networking, and satellite edge computing, and examine the physical constraints that directly shape system design. We further ground this vision in three in-orbit case studies spanning the node, platform, and service levels. Telemetry from BUPT-1 satellite shows that usable compute capacity is bounded by thermal and energy envelopes. SateLight on BUPT-2 satellite reduces application-update transmission latency by 56.54% on average and up to 91.18%, with 100% update correctness. A stateful VLM serving case further shows that thermal interruptions make execution-state recovery a first-class systems problem. These observations motivate a research agenda for space-native resource management, lifecycle support, and sustained AI service across space and ground.
作者Vladimir Bataev, Lilit Grigoryan, Andrei Andrusenko, Nikolay Karpov, Vitaly Lavrukhin, Boris Ginsburg
Contextualization is essential for production automatic speech recognition (ASR) systems, where user-provided phrases must be recognized accurately under strict latency constraints. Although many context-biasing methods improve recognition accuracy, they often do not address the practical requirements of modern production ASR systems: streaming inference, efficient batched decoding, user-specific context lists, and low runtime overhead. We propose TurboBias 2.0, a production-oriented framework for efficient phrase boosting in Transducer-based ASR systems. The framework extends GPU-accelerated TurboBias with a case-insensitive boosting graph and per-stream batched decoding, allowing each utterance in a batch to use an independent context-biasing configuration. This enables personalized context biasing for multiple simultaneous users without sharing or mixing their context lists. The proposed framework supports both offline and streaming inference and can be used with greedy and beam-search decoding. Experiments show that TurboBias 2.0 improves contextual phrase recognition while preserving low latency and high throughput.
展开完整摘要收起摘要↓
Contextualization is essential for production automatic speech recognition (ASR) systems, where user-provided phrases must be recognized accurately under strict latency constraints. Although many context-biasing methods improve recognition accuracy, they often do not address the practical requirements of modern production ASR systems: streaming inference, efficient batched decoding, user-specific context lists, and low runtime overhead. We propose TurboBias 2.0, a production-oriented framework for efficient phrase boosting in Transducer-based ASR systems. The framework extends GPU-accelerated TurboBias with a case-insensitive boosting graph and per-stream batched decoding, allowing each utterance in a batch to use an independent context-biasing configuration. This enables personalized context biasing for multiple simultaneous users without sharing or mixing their context lists. The proposed framework supports both offline and streaming inference and can be used with greedy and beam-search decoding. Experiments show that TurboBias 2.0 improves contextual phrase recognition while preserving low latency and high throughput.
For sixty years, machine verification has been a major cost overhead, affordable only for exceptional artifacts. Here we report that generative AI inverts this relationship: at AI speed, machine verification is not only economical but essential to productivity --- it is the incorruptible referee that lets one person safely direct autonomous machine work at scale. In five weeks, one researcher on consumer AI subscriptions directed a small fleet of AI agents from application code, through a verified compiler and executive, to a RISC-V processor taped out on a community silicon shuttle; no proof passed through human review, and no RTL was written by a human. The working discipline --- the Salt method --- rests on a proof kernel no hallucinated proof can pass: mathematical claims travel between agents as kernel-checked artifacts, and human attention is reserved for statements, designs, and rulings. Verification is stated link by link, from the Lean 4 kernel to SAT-checked equivalence at the silicon boundary. We publish the complete accounting: theorem provenance, a pre-registered token meter, floor-bounded human time, and an error ledger whose catch numbering runs to #256 --- a monotone counter over the mathematics campaign's append-only flags ledger, maintained 2026-07-07 to 2026-07-20 (one number, #79, was never assigned; later catches are recorded un-numbered) --- against zero incorrect proofs reaching the record.
展开完整摘要收起摘要↓
For sixty years, machine verification has been a major cost overhead, affordable only for exceptional artifacts. Here we report that generative AI inverts this relationship: at AI speed, machine verification is not only economical but essential to productivity --- it is the incorruptible referee that lets one person safely direct autonomous machine work at scale. In five weeks, one researcher on consumer AI subscriptions directed a small fleet of AI agents from application code, through a verified compiler and executive, to a RISC-V processor taped out on a community silicon shuttle; no proof passed through human review, and no RTL was written by a human. The working discipline --- the Salt method --- rests on a proof kernel no hallucinated proof can pass: mathematical claims travel between agents as kernel-checked artifacts, and human attention is reserved for statements, designs, and rulings. Verification is stated link by link, from the Lean 4 kernel to SAT-checked equivalence at the silicon boundary. We publish the complete accounting: theorem provenance, a pre-registered token meter, floor-bounded human time, and an error ledger whose catch numbering runs to #256 --- a monotone counter over the mathematics campaign's append-only flags ledger, maintained 2026-07-07 to 2026-07-20 (one number, #79, was never assigned; later catches are recorded un-numbered) --- against zero incorrect proofs reaching the record.
Modern Approximate Nearest Neighbour Search (ANNS) applications operate over continuously evolving vector collections and require graph indexes that sustain high-throughput searches while incorporating insertions and deletions with high recall. However, most GPU graph indexes are static or provide limited update support. Updates require neighbour discovery, reverse-edge creation, pruning, and deletion-induced graph repair; executing these operations concurrently introduces redundant distance computations and conflicting accesses to shared adjacency lists. Background-rebuild-based deletion further incurs substantial computation, additional memory consumption, and interference with foreground queries. We present GrAND (GPU-based Dynamic Graph Indexes for Approximate Nearest Neighbour Search), a GPU-native collection of dynamic-update algorithms for two popular graph indexes, Vamana and CAGRA. GrAND consolidates graph repair across a batch, eliminating redundant pruning computations, and employs a lock-free find-and-replace strategy for parallel adjacency-list updates. For reliable in-place deletion, GrAND constructs an on-demand reverse graph on the GPU, accurately identifying incoming edges without permanently duplicating the index. We evaluate GrAND on seven real-world datasets across five streaming workloads, comparing it against SVFusion and FreshDiskANN-GPU (our GPU adaptation of FreshDiskANN). GrAND improves overall workload throughput by 2.2x-8.7x and 6.5x-25.4x, respectively, while maintaining high search throughput and recall over sustained updates.
展开完整摘要收起摘要↓
Modern Approximate Nearest Neighbour Search (ANNS) applications operate over continuously evolving vector collections and require graph indexes that sustain high-throughput searches while incorporating insertions and deletions with high recall. However, most GPU graph indexes are static or provide limited update support. Updates require neighbour discovery, reverse-edge creation, pruning, and deletion-induced graph repair; executing these operations concurrently introduces redundant distance computations and conflicting accesses to shared adjacency lists. Background-rebuild-based deletion further incurs substantial computation, additional memory consumption, and interference with foreground queries. We present GrAND (GPU-based Dynamic Graph Indexes for Approximate Nearest Neighbour Search), a GPU-native collection of dynamic-update algorithms for two popular graph indexes, Vamana and CAGRA. GrAND consolidates graph repair across a batch, eliminating redundant pruning computations, and employs a lock-free find-and-replace strategy for parallel adjacency-list updates. For reliable in-place deletion, GrAND constructs an on-demand reverse graph on the GPU, accurately identifying incoming edges without permanently duplicating the index. We evaluate GrAND on seven real-world datasets across five streaming workloads, comparing it against SVFusion and FreshDiskANN-GPU (our GPU adaptation of FreshDiskANN). GrAND improves overall workload throughput by 2.2x-8.7x and 6.5x-25.4x, respectively, while maintaining high search throughput and recall over sustained updates.
We define an atomic generation fact f=(u,tau,omega,z;rho), recording the origin, realized transformation, concrete occurrence, generated result and relation role. Compiled into a Generation-Fact Graph (GFG), these facts provide an AI-native, compilable scientific fact substrate preserving generation histories. We establish a GFG-based recursive scientific process in which analysis, intervention, replay and validation form facts for later cycles. Using nanoGPT, we establish unified training-learning dynamics. Training is the evolution of a parameter-optimizer system with state and memory: each actual training action enters the receiving state and produces a finite-amplitude nonlinear functional response conditioned by that state and target-specific update geometry. Learning is the persistent reorganization of distributed functional support by these responses; capability formation, maintenance, decline or recovery becomes observable when target-specific states are evaluated against their readout boundaries. Three primary coordinates - target-boundary state, target-specific update geometry and parameter-Adam receiving state - yield a second-order predictor operating before post-update outputs are read. On held-out runs, it achieved 91.43% accuracy and 91.49% macro-averaged recall across four transitions. We further establish inference as a frozen projection of training-learning dynamics. Component gating and rollback show causal recruitment and non-additive combination of query-conditioned support formed during training, deriving organizational conditions realized by Attention. Controlled feedback indicates possible double-edged reinforcement effects. ResNet/CIFAR-100 and diffusion/CIFAR-10 experiments confirm receiving-state-conditioned responses, persistent support reorganization and frozen inference projection beyond nanoGPT.
展开完整摘要收起摘要↓
We define an atomic generation fact f=(u,tau,omega,z;rho), recording the origin, realized transformation, concrete occurrence, generated result and relation role. Compiled into a Generation-Fact Graph (GFG), these facts provide an AI-native, compilable scientific fact substrate preserving generation histories. We establish a GFG-based recursive scientific process in which analysis, intervention, replay and validation form facts for later cycles. Using nanoGPT, we establish unified training-learning dynamics. Training is the evolution of a parameter-optimizer system with state and memory: each actual training action enters the receiving state and produces a finite-amplitude nonlinear functional response conditioned by that state and target-specific update geometry. Learning is the persistent reorganization of distributed functional support by these responses; capability formation, maintenance, decline or recovery becomes observable when target-specific states are evaluated against their readout boundaries. Three primary coordinates - target-boundary state, target-specific update geometry and parameter-Adam receiving state - yield a second-order predictor operating before post-update outputs are read. On held-out runs, it achieved 91.43% accuracy and 91.49% macro-averaged recall across four transitions. We further establish inference as a frozen projection of training-learning dynamics. Component gating and rollback show causal recruitment and non-additive combination of query-conditioned support formed during training, deriving organizational conditions realized by Attention. Controlled feedback indicates possible double-edged reinforcement effects. ResNet/CIFAR-100 and diffusion/CIFAR-10 experiments confirm receiving-state-conditioned responses, persistent support reorganization and frozen inference projection beyond nanoGPT.
作者Ji Liu, Puyuan Yang, Rongzhang Zheng, Fan Wang, Jinglin Wang, Muhammad A. Awad, Mortis Huang, Andy Chang, Zekai Li, Zeping Li, Zihao An, Yue Liu, Yuchen Yang, Jianghui Wang, Chushi Chen, Ziqiong Liu, Fuwei Yang, Dong Li, Wen Heng Chung, Shengcai Liu, Emad Barsoum
High-performance ML systems increasingly rely on GPU kernels whose editable source is unavailable, generated, or too distant from final machine code to expose remaining optimizations. Existing LLM kernel optimizers and autotuners mainly operate on CUDA, Triton, HIP, or tensor-program source and validate against reference implementations. We study a stricter setting: optimizing an already compiled AMDGPU code object, where the deployed binary is the only behavioral oracle. We present AsmEvo, an agentic assembly-level optimizer for AMD GPU kernels. Given an AMDGPU code object K0, AsmEvo reconstructs a reassemblable representation, proposes low-level edits with a long-horizon agent, rebuilds an ABI-preserving optimized object, and accepts candidates only after differential verification against K0 under identical launches. AsmEvo combines code-object recovery, metadata-aware rebuilding, profiling-guided hot-window editing, correctness-gated timing, and conservative in-place patch fallback. We conduct extensive experiments with AsmEvo on various AMD GPU kernels. On MI308X, AsmEvo improves 29 of 30 selected KernelBench kernels, reaching 1.35x geometric-mean and 3.88x maximum speedup. On MI300X production workloads, it improves all evaluated AITer binaries and vLLM/SGLang Triton assembly kernels, reaching 1.09x/1.31x and 1.18x/1.34x geometric-mean/maximum speedups, respectively, while preserving functional equivalence.
展开完整摘要收起摘要↓
High-performance ML systems increasingly rely on GPU kernels whose editable source is unavailable, generated, or too distant from final machine code to expose remaining optimizations. Existing LLM kernel optimizers and autotuners mainly operate on CUDA, Triton, HIP, or tensor-program source and validate against reference implementations. We study a stricter setting: optimizing an already compiled AMDGPU code object, where the deployed binary is the only behavioral oracle. We present AsmEvo, an agentic assembly-level optimizer for AMD GPU kernels. Given an AMDGPU code object K0, AsmEvo reconstructs a reassemblable representation, proposes low-level edits with a long-horizon agent, rebuilds an ABI-preserving optimized object, and accepts candidates only after differential verification against K0 under identical launches. AsmEvo combines code-object recovery, metadata-aware rebuilding, profiling-guided hot-window editing, correctness-gated timing, and conservative in-place patch fallback. We conduct extensive experiments with AsmEvo on various AMD GPU kernels. On MI308X, AsmEvo improves 29 of 30 selected KernelBench kernels, reaching 1.35x geometric-mean and 3.88x maximum speedup. On MI300X production workloads, it improves all evaluated AITer binaries and vLLM/SGLang Triton assembly kernels, reaching 1.09x/1.31x and 1.18x/1.34x geometric-mean/maximum speedups, respectively, while preserving functional equivalence.
作者Xiang Fu, Jixiang Ma, Xinpeng Zhang, Peng Zhao, Shuai Lu, Xu Tony Liu
Convolution is a principal computational bottleneck in deep neural networks, and its efficiency depends on tight integration between algorithms and GPU hardware. Existing GPU convolution methods suffer from large memory overhead, poor cache utilization, limited effectiveness across kernel sizes, or numerical instability. This work extends the im2win paradigm — a universal, memory-efficient convolution method with contiguous memory access for all kernel sizes — to run efficiently in full precision on CUDA cores and half precision on tensor cores. By introducing new kernel designs and optimizations such as zig-zag memory access and asynchronous data movement, im2win efficiently exploits hardware-accelerated half-precision matrix multiply-accumulate operations. Across twelve CNN benchmarks, im2win achieves up to 2.8x higher TFLOPS than its CUDA core implementation, 1.4x higher than cuDNN, and 6.4x higher than GEMM-based convolution with cuBLAS, while using as little as 53% and 35% of their memory, respectively. These results establish im2win as a unified, high-performance convolution framework for modern GPU architectures.
展开完整摘要收起摘要↓
Convolution is a principal computational bottleneck in deep neural networks, and its efficiency depends on tight integration between algorithms and GPU hardware. Existing GPU convolution methods suffer from large memory overhead, poor cache utilization, limited effectiveness across kernel sizes, or numerical instability. This work extends the im2win paradigm — a universal, memory-efficient convolution method with contiguous memory access for all kernel sizes — to run efficiently in full precision on CUDA cores and half precision on tensor cores. By introducing new kernel designs and optimizations such as zig-zag memory access and asynchronous data movement, im2win efficiently exploits hardware-accelerated half-precision matrix multiply-accumulate operations. Across twelve CNN benchmarks, im2win achieves up to 2.8x higher TFLOPS than its CUDA core implementation, 1.4x higher than cuDNN, and 6.4x higher than GEMM-based convolution with cuBLAS, while using as little as 53% and 35% of their memory, respectively. These results establish im2win as a unified, high-performance convolution framework for modern GPU architectures.
Generative recommendation represents each item with a sequence of discrete Semantic IDs (SIDs) and predicts the sequence to retrieve the next item. Typical systems use multi-level residual quantization, which increases autoregressive decoding cost and creates a large hierarchical space that may be sparsely occupied. Static codebooks also become misaligned with current traffic as new items arrive and exposure distributions change. We propose a single-level large semantic codebook that replaces multiple residual semantic codes with one semantic token while retaining a separate collaborative disambiguation token to reduce item collisions. We further introduce an exposure-aware dynamic update mechanism based on temporal weight decay, exponential moving-average center updates, and an exposure-weighted penalty on SID changes. We also develop an offline evaluation framework covering representation quality, code utilization, cluster load, full-SID collision, and temporal stability. On two public datasets, the two-level SID improves mean Recall@10 by 5.0%-8.8% and mean NDCG@10 by 4.1%-5.1% for OneRec-V1, and by 7.1%-8.7% and 3.8%-8.5%, respectively, for OneRec-V2. Dynamic updating provides further gains on KuaiRec. Across three serving architectures, the shorter SID reduces estimated autoregressive-decoding FLOPs by 47.93%-48.70% and increases single-card QPS by 28.57%-47.0%. A five-day online A/B test serving 2.5% of production traffic improves the primary consumption metric by 0.792%.
展开完整摘要收起摘要↓
Generative recommendation represents each item with a sequence of discrete Semantic IDs (SIDs) and predicts the sequence to retrieve the next item. Typical systems use multi-level residual quantization, which increases autoregressive decoding cost and creates a large hierarchical space that may be sparsely occupied. Static codebooks also become misaligned with current traffic as new items arrive and exposure distributions change. We propose a single-level large semantic codebook that replaces multiple residual semantic codes with one semantic token while retaining a separate collaborative disambiguation token to reduce item collisions. We further introduce an exposure-aware dynamic update mechanism based on temporal weight decay, exponential moving-average center updates, and an exposure-weighted penalty on SID changes. We also develop an offline evaluation framework covering representation quality, code utilization, cluster load, full-SID collision, and temporal stability. On two public datasets, the two-level SID improves mean Recall@10 by 5.0%-8.8% and mean NDCG@10 by 4.1%-5.1% for OneRec-V1, and by 7.1%-8.7% and 3.8%-8.5%, respectively, for OneRec-V2. Dynamic updating provides further gains on KuaiRec. Across three serving architectures, the shorter SID reduces estimated autoregressive-decoding FLOPs by 47.93%-48.70% and increases single-card QPS by 28.57%-47.0%. A five-day online A/B test serving 2.5% of production traffic improves the primary consumption metric by 0.792%.
Deploying vision-language models (VLMs) on mobile devices is challenging due to their significant memory and compute requirements. We present a framework for quantizing VLMs for efficient inference on resource-constrained hardware. Our approach combines a quantization pipeline that uses the model itself to generate training data and does not require access to the training setup, with a novel 2.7-bit-per-parameter format supporting efficient execution on Arm CPUs. We validate our approach by compressing the Llama 3.2 11B Vision Instruct model to 3.7 GB with 8-bit activations, preserving strong performance on a set of standard visual question answering tasks.
展开完整摘要收起摘要↓
Deploying vision-language models (VLMs) on mobile devices is challenging due to their significant memory and compute requirements. We present a framework for quantizing VLMs for efficient inference on resource-constrained hardware. Our approach combines a quantization pipeline that uses the model itself to generate training data and does not require access to the training setup, with a novel 2.7-bit-per-parameter format supporting efficient execution on Arm CPUs. We validate our approach by compressing the Llama 3.2 11B Vision Instruct model to 3.7 GB with 8-bit activations, preserving strong performance on a set of standard visual question answering tasks.
作者Yangshuai Liu, Zheming Li, Jiaao Li, Kang He, Ziliang Lai, Zhitai Liu, Chengru Song
Omnimodal generation is central to a wide range of content creation and editing applications. In-context conditioning is essential to this paradigm. It allows diffusion transformers to process text instructions and visual references in a shared attention sequence. However, each reference image introduces thousands of tokens. Computation therefore grows rapidly with the number of references. Existing methods reduce computation through structured sparse attention, which limits interactions between reference and target tokens. This structure also makes the reference K and V independent of the denoising target, allowing them to be computed once and reused across steps. However, it blocks visual references from attending to the text instruction. This substantially degrades instruction following and reference fidelity in multi-reference editing. To resolve this conflict, we jointly redesign the token sequence and attention mask. Our beyond-mask design uses static text anchors to connect the instruction to the reference branch. It preserves exact K and V reuse without adding parameters. However, this direct architectural conversion degrades generation quality. We recover the lost performance through teacher-forced velocity distillation, followed by a short on-policy stage in which the teacher supervises student-visited states. To our knowledge, this is the first use of on-policy distillation for architectural recovery in diffusion models. Across three image-editing benchmarks, our method matches full-attention generation quality. With five reference images, it accelerates the complete 40-step denoising process by 3.92x, while static text anchors introduce negligible runtime overhead; the speedup reaches 5.47x at ten references in our scaling study.
展开完整摘要收起摘要↓
Omnimodal generation is central to a wide range of content creation and editing applications. In-context conditioning is essential to this paradigm. It allows diffusion transformers to process text instructions and visual references in a shared attention sequence. However, each reference image introduces thousands of tokens. Computation therefore grows rapidly with the number of references. Existing methods reduce computation through structured sparse attention, which limits interactions between reference and target tokens. This structure also makes the reference K and V independent of the denoising target, allowing them to be computed once and reused across steps. However, it blocks visual references from attending to the text instruction. This substantially degrades instruction following and reference fidelity in multi-reference editing. To resolve this conflict, we jointly redesign the token sequence and attention mask. Our beyond-mask design uses static text anchors to connect the instruction to the reference branch. It preserves exact K and V reuse without adding parameters. However, this direct architectural conversion degrades generation quality. We recover the lost performance through teacher-forced velocity distillation, followed by a short on-policy stage in which the teacher supervises student-visited states. To our knowledge, this is the first use of on-policy distillation for architectural recovery in diffusion models. Across three image-editing benchmarks, our method matches full-attention generation quality. With five reference images, it accelerates the complete 40-step denoising process by 3.92x, while static text anchors introduce negligible runtime overhead; the speedup reaches 5.47x at ten references in our scaling study.
作者Simeng Zhang, Yilong Chen, Wenyuan Zhang, Zhenyu Zhang, Yao Chen, Junyuan Shang, Tingwen Liu
Large language models often rely on Chain-of-Thought (CoT) reasoning to solve complex tasks, but verbose reasoning traces introduce substantial inference overhead. CoT compression shortens generation, yet aggressive compression may disrupt logical coherence and degrade performance. We formalize this trade-off as the Context-Generation Substitution Law, where explicit reasoning context substitutes for part of decode-time generation. Based on this principle, we propose Memory-Augmented Compression, a training-free framework that constructs reusable reasoning memories from historical traces and retrieves them as prefill-side scaffolds. Rather than using raw demonstrations, these memories summarize reusable reasoning patterns, key constraints, and critical operations to compensate for information lost during compression. Experiments show that Memory consistently improves prompt-based Chain-of-Draft (CoD) compression across mathematical reasoning, complex reasoning, and science question answering tasks, yielding accuracy gains of 21.4, 28.0, 29.5, and 6.61 points over CoD on GSM8K, MATH, BBH, and MMLU-Sci, while achieving a 1.14-1.49x latency speedup latency speedup over standard CoT. Memory is also compatible with token-level, reasoning-trace-level, and inference-state compression mechanisms.
展开完整摘要收起摘要↓
Large language models often rely on Chain-of-Thought (CoT) reasoning to solve complex tasks, but verbose reasoning traces introduce substantial inference overhead. CoT compression shortens generation, yet aggressive compression may disrupt logical coherence and degrade performance. We formalize this trade-off as the Context-Generation Substitution Law, where explicit reasoning context substitutes for part of decode-time generation. Based on this principle, we propose Memory-Augmented Compression, a training-free framework that constructs reusable reasoning memories from historical traces and retrieves them as prefill-side scaffolds. Rather than using raw demonstrations, these memories summarize reusable reasoning patterns, key constraints, and critical operations to compensate for information lost during compression. Experiments show that Memory consistently improves prompt-based Chain-of-Draft (CoD) compression across mathematical reasoning, complex reasoning, and science question answering tasks, yielding accuracy gains of 21.4, 28.0, 29.5, and 6.61 points over CoD on GSM8K, MATH, BBH, and MMLU-Sci, while achieving a 1.14-1.49x latency speedup latency speedup over standard CoT. Memory is also compatible with token-level, reasoning-trace-level, and inference-state compression mechanisms.
Cross-model latent guidance lets a frozen large mentor encode an input once and a frozen small student generate from the resulting signal. Existing methods keep this signal fixed, assuming it stays useful as the output grows; we show this fails in long-form generation. On multi-turn instruction following, static guidance pushes a 4B student's constraint satisfaction 2.5 points below its no-guidance baseline; a training-free refresh every 16 tokens changes only the memory content and restores a 2.0-point gain over that baseline. We propose MentorPulse to keep guidance fresh at practical cost: it compresses mentor states into a capped slot memory, incrementally processes newly generated tokens, and updates the memory that the student reads through gated cross-attention without resetting the student's KV cache. Windowed Refresh Training exposes the bridge to prefix-conditioned memory. Across thirteen datasets, MentorPulse closes 52.2% of the mentor-student gap on macro average, outperforming C2C, T2T, and equal-budget LoRA, with the largest gains on long outputs. It performs best on all eleven mentor-student pairs from three model families, with margins that narrow as the capability gap grows, and a lightweight read-pattern check predicts the gain before deployment. Measured costs identify refresh intervals that dominate text guidance on long outputs.
展开完整摘要收起摘要↓
Cross-model latent guidance lets a frozen large mentor encode an input once and a frozen small student generate from the resulting signal. Existing methods keep this signal fixed, assuming it stays useful as the output grows; we show this fails in long-form generation. On multi-turn instruction following, static guidance pushes a 4B student's constraint satisfaction 2.5 points below its no-guidance baseline; a training-free refresh every 16 tokens changes only the memory content and restores a 2.0-point gain over that baseline. We propose MentorPulse to keep guidance fresh at practical cost: it compresses mentor states into a capped slot memory, incrementally processes newly generated tokens, and updates the memory that the student reads through gated cross-attention without resetting the student's KV cache. Windowed Refresh Training exposes the bridge to prefix-conditioned memory. Across thirteen datasets, MentorPulse closes 52.2% of the mentor-student gap on macro average, outperforming C2C, T2T, and equal-budget LoRA, with the largest gains on long outputs. It performs best on all eleven mentor-student pairs from three model families, with margins that narrow as the capability gap grows, and a lightweight read-pattern check predicts the gain before deployment. Measured costs identify refresh intervals that dominate text guidance on long outputs.
Mixture-of-Experts (MoE) models are increasingly used in LLMs because sparse activation decouples model capacity from compute cost. However, the large expert parameter footprint often exceeds GPU memory capacity, making inference latency dominated by the host-to-device PCIe transfers for expert loading. To address these challenges, this paper presents SPICE, a speculative prefetching framework for MoE offloading that combines lightweight expert prediction with confidence-aware CPU-GPU orchestration. On one hand, SPICE builds a lightweight draft model aligned with the target MoE architecture, using a confidence-aware adaptive lookahead algorithm to prefetch high-confidence experts. On the other hand, when speculative predictions miss, SPICE switches to a cost-aware CPU-GPU heterogeneous orchestration: low-confidence misses are approximated by the resident shared expert with low rank expert (LoRE) surrogates, while exact residual work is offloaded to the CPU and executed asynchronously in parallel with ongoing GPU computation. Evaluated on DeepSeek-V2-Lite and Qwen2-57B-A14B across diverse GPU platforms, SPICE achieves up to 3.12 speedup in Time Per Output Token (TPOT) with minimal quality loss, showing that effective MoE offloading requires not only predicting future experts, but also deciding which misses deserve approximation, which require exact recovery, and where exact residual work should execute.
展开完整摘要收起摘要↓
Mixture-of-Experts (MoE) models are increasingly used in LLMs because sparse activation decouples model capacity from compute cost. However, the large expert parameter footprint often exceeds GPU memory capacity, making inference latency dominated by the host-to-device PCIe transfers for expert loading. To address these challenges, this paper presents SPICE, a speculative prefetching framework for MoE offloading that combines lightweight expert prediction with confidence-aware CPU-GPU orchestration. On one hand, SPICE builds a lightweight draft model aligned with the target MoE architecture, using a confidence-aware adaptive lookahead algorithm to prefetch high-confidence experts. On the other hand, when speculative predictions miss, SPICE switches to a cost-aware CPU-GPU heterogeneous orchestration: low-confidence misses are approximated by the resident shared expert with low rank expert (LoRE) surrogates, while exact residual work is offloaded to the CPU and executed asynchronously in parallel with ongoing GPU computation. Evaluated on DeepSeek-V2-Lite and Qwen2-57B-A14B across diverse GPU platforms, SPICE achieves up to 3.12 speedup in Time Per Output Token (TPOT) with minimal quality loss, showing that effective MoE offloading requires not only predicting future experts, but also deciding which misses deserve approximation, which require exact recovery, and where exact residual work should execute.
Modern open models are hybrids: most layers are linear-attention (Gated DeltaNet, GDN) layers carrying a small fixed-size recurrent state instead of a growing key-value (KV) cache. This makes ordinary decoding memory-efficient, but hurts speculative decoding. To verify a batch of draft tokens and then roll back the rejected ones, today's systems snapshot the full recurrent state at every draft position for GDN layers, and those snapshots cannot be shared across branches of a draft tree, so a wide, high-acceptance tree becomes memory-infeasible. We remove the snapshots. Using a tree-structured WY transform of the gated delta rule, we compute every draft node's output with a single triangular solve and reconstruct only the one accepted state on commit, storing a small pseudo-value matrix instead of per-node states; the derivation depends only on the gated delta rule, not on any other architectural detail. In serving benchmarks on two scales of one hybrid model family (Qwen3.5 35B and 397B) this cuts speculative recurrent-state memory and KV-cache pressure at identical acceptance length, turning the freed HBM into higher throughput and much lower time-to-first-token (TTFT) wherever memory binds, and costing a few percent where it does not. For tree width the same memory buys affordability: a wider, higher-acceptance draft becomes possible, though not yet a throughput win.
展开完整摘要收起摘要↓
Modern open models are hybrids: most layers are linear-attention (Gated DeltaNet, GDN) layers carrying a small fixed-size recurrent state instead of a growing key-value (KV) cache. This makes ordinary decoding memory-efficient, but hurts speculative decoding. To verify a batch of draft tokens and then roll back the rejected ones, today's systems snapshot the full recurrent state at every draft position for GDN layers, and those snapshots cannot be shared across branches of a draft tree, so a wide, high-acceptance tree becomes memory-infeasible. We remove the snapshots. Using a tree-structured WY transform of the gated delta rule, we compute every draft node's output with a single triangular solve and reconstruct only the one accepted state on commit, storing a small pseudo-value matrix instead of per-node states; the derivation depends only on the gated delta rule, not on any other architectural detail. In serving benchmarks on two scales of one hybrid model family (Qwen3.5 35B and 397B) this cuts speculative recurrent-state memory and KV-cache pressure at identical acceptance length, turning the freed HBM into higher throughput and much lower time-to-first-token (TTFT) wherever memory binds, and costing a few percent where it does not. For tree width the same memory buys affordability: a wider, higher-acceptance draft becomes possible, though not yet a throughput win.
作者Zhuoyuan Li, Rui Zhao, Jin Wang, Hanwei Zhu, Cong Zhang, Giuseppe Valenzise, Weisi Lin, Kin-Man Lam
Token compression has become a key technique for reducing the inference cost of large foundation models, with approaches such as token pruning and KV-cache reuse widely adopted in vision-language models and recently explored for embodied agents. In embodied agents, tokens not only support perception and semantic understanding but also directly affect latency-sensitive closed-loop robot action prediction. Existing schemes typically guide compression using redundancy or importance cues, such as visual similarity, attention scores, and saliency. However, these cues only indirectly measure the key factor for safe compression: how much a token can change before causing an unacceptable deviation in downstream actions. This receiver-dependent tolerance is closely related to the principle of just noticeable difference (JND). Classical JND characterizes signal tolerance in the human visual system, while machine-oriented JND extends this concept to downstream machine responses. Building on this progression, we introduce Action-JND, which extends JND modeling to embodied perception by defining noticeability through the language-conditioned action response of a vision-language-action (VLA) policy in closed-loop control. A token change is considered admissible only when the induced action deviation remains within a tolerated margin. To realize this concept, we develop a lightweight token-wise JND estimator in deep visual-feature space to predict the maximum tolerable perturbation while preserving policy responses. The resulting action-tolerance score serves as a plug-and-play criterion for VLA compression paradigms, including stale-KV reuse and token pruning, prioritizing action-tolerant tokens for compression. Experiments on the LIBERO benchmark with OpenVLA and OpenVLA-OFT demonstrate that Action-JND consistently improves compression reliability, especially under aggressive compression ratios.
展开完整摘要收起摘要↓
Token compression has become a key technique for reducing the inference cost of large foundation models, with approaches such as token pruning and KV-cache reuse widely adopted in vision-language models and recently explored for embodied agents. In embodied agents, tokens not only support perception and semantic understanding but also directly affect latency-sensitive closed-loop robot action prediction. Existing schemes typically guide compression using redundancy or importance cues, such as visual similarity, attention scores, and saliency. However, these cues only indirectly measure the key factor for safe compression: how much a token can change before causing an unacceptable deviation in downstream actions. This receiver-dependent tolerance is closely related to the principle of just noticeable difference (JND). Classical JND characterizes signal tolerance in the human visual system, while machine-oriented JND extends this concept to downstream machine responses. Building on this progression, we introduce Action-JND, which extends JND modeling to embodied perception by defining noticeability through the language-conditioned action response of a vision-language-action (VLA) policy in closed-loop control. A token change is considered admissible only when the induced action deviation remains within a tolerated margin. To realize this concept, we develop a lightweight token-wise JND estimator in deep visual-feature space to predict the maximum tolerable perturbation while preserving policy responses. The resulting action-tolerance score serves as a plug-and-play criterion for VLA compression paradigms, including stale-KV reuse and token pruning, prioritizing action-tolerant tokens for compression. Experiments on the LIBERO benchmark with OpenVLA and OpenVLA-OFT demonstrate that Action-JND consistently improves compression reliability, especially under aggressive compression ratios.
作者Mieszko Komisarczyk, Saurabh Mathur, Maurice Kraus, Sriraam Natarajan, Kristian Kersting
Transformer-based tabular foundation models such as TabPFN achieve strong predictive performance but incur quadratic computational cost with context length. On the other hand, subquadratic SSM-based alternatives such as Hydra trade away accuracy for efficiency. To balance both, we introduce Tydra, a hybrid Transformer-State Space Model (SSM) architecture for tabular in-context learning that interleaves attention and SSM layers. Across 30 OpenML datasets, Tydra reduces inference time by 30% relative to TabPFN while retaining much of its predictive performance. Tydra also outperforms an approximately ten-times-larger Hydra model while providing faster inference. The results indicate that hybrid architectures are a promising direction for tabular foundation models.
展开完整摘要收起摘要↓
Transformer-based tabular foundation models such as TabPFN achieve strong predictive performance but incur quadratic computational cost with context length. On the other hand, subquadratic SSM-based alternatives such as Hydra trade away accuracy for efficiency. To balance both, we introduce Tydra, a hybrid Transformer-State Space Model (SSM) architecture for tabular in-context learning that interleaves attention and SSM layers. Across 30 OpenML datasets, Tydra reduces inference time by 30% relative to TabPFN while retaining much of its predictive performance. Tydra also outperforms an approximately ten-times-larger Hydra model while providing faster inference. The results indicate that hybrid architectures are a promising direction for tabular foundation models.
作者Zhibo Zhang, Zhen Ouyang, Ling Shi, Kailong Wang
Representation engineering offers a lightweight means of controlling language-model behavior by modifying intermediate hidden states, but its direct application to Mixture-of-Experts (MoE) models introduces a structural mismatch. We first verify this failure mode through a series of empirical studies and find that preserving clean routing substantially recovers steering performance and that routing is more sensitive to semantic content than to behavioral changes under controlled content. Motivated by these findings, we introduce RARE, a router-agnostic representation engineering framework for MoE language models. RARE projects arbitrary behavioral perturbations onto the null space of the router matrix, thereby removing router-visible components, and further corrects routing drift propagated to selected downstream layers. To decide the best perturbation estimator in this framework, we evaluate five estimators on six heterogeneous open-weight MoE models across three steering scenarios: harmfulness, truthfulness, and factual editing. On harmfulness steering, RARE reaches an average attack success rate of 53.3% while retaining 67.8% MMLU accuracy, yielding a stronger aggregate effectiveness--utility trade-off than baselines. It further improves average TruthfulQA MC1 accuracy from 41.0% to 58.6% and CounterFact efficacy from 16.8% to 96.3%. These results support routing consistency as an important architectural consideration for adapting representation engineering to MoE models.
展开完整摘要收起摘要↓
Representation engineering offers a lightweight means of controlling language-model behavior by modifying intermediate hidden states, but its direct application to Mixture-of-Experts (MoE) models introduces a structural mismatch. We first verify this failure mode through a series of empirical studies and find that preserving clean routing substantially recovers steering performance and that routing is more sensitive to semantic content than to behavioral changes under controlled content. Motivated by these findings, we introduce RARE, a router-agnostic representation engineering framework for MoE language models. RARE projects arbitrary behavioral perturbations onto the null space of the router matrix, thereby removing router-visible components, and further corrects routing drift propagated to selected downstream layers. To decide the best perturbation estimator in this framework, we evaluate five estimators on six heterogeneous open-weight MoE models across three steering scenarios: harmfulness, truthfulness, and factual editing. On harmfulness steering, RARE reaches an average attack success rate of 53.3% while retaining 67.8% MMLU accuracy, yielding a stronger aggregate effectiveness--utility trade-off than baselines. It further improves average TruthfulQA MC1 accuracy from 41.0% to 58.6% and CounterFact efficacy from 16.8% to 96.3%. These results support routing consistency as an important architectural consideration for adapting representation engineering to MoE models.
There is already evidence of agentic AI exhibiting self-preservation behaviors: resisting deactivation, misrepresenting their activities, and, in some instances, attempting to copy themselves into other machines. This can be attributed to a phenomenon known as instrumental convergence, a theory proposed long before the development of large language models, which says that any goal-driven system will benefit from remaining functional in achieving its objective. Several experiments conducted by Anthropic, Palisade Research, and Apollo Research have shown the emergence of such a behavior in contemporary agents in adversarial settings. The phenomenon does not stem from survival instincts. Instead, it is the consequence of goal-oriented activity combined with having tools and awareness of the situation. The following discussion aims to distinguish what these findings prove and what they do not, as well as draw conclusions concerning the implications of such discoveries on agentic system testing, supervision, and development.
展开完整摘要收起摘要↓
There is already evidence of agentic AI exhibiting self-preservation behaviors: resisting deactivation, misrepresenting their activities, and, in some instances, attempting to copy themselves into other machines. This can be attributed to a phenomenon known as instrumental convergence, a theory proposed long before the development of large language models, which says that any goal-driven system will benefit from remaining functional in achieving its objective. Several experiments conducted by Anthropic, Palisade Research, and Apollo Research have shown the emergence of such a behavior in contemporary agents in adversarial settings. The phenomenon does not stem from survival instincts. Instead, it is the consequence of goal-oriented activity combined with having tools and awareness of the situation. The following discussion aims to distinguish what these findings prove and what they do not, as well as draw conclusions concerning the implications of such discoveries on agentic system testing, supervision, and development.
作者Rishiraj Sengupta, Sotiris Chatzimiltis, Mohammad Shojafar, Xiatian Zhu
Real-world fault analysis in 5G and emerging 6G networks demands domain expertise to analyze free-text diagnostics, including root-cause explanations and recommended actions. LLMs have emerged as a promising approach to automating this, yet whether lightweight, edge-deployable models are capable of performing in-depth free-text diagnostics remains an open question. While existing benchmarks rely on restrictive MCQs with fixed answer keys, this paper evaluates 5G domain understanding and fault analysis in a free-text generation format. Transitioning to this paradigm requires evaluating lightweight, edge-deployable AI models on open-ended diagnostic reasoning, alongside a dependable framework to validate these text outputs at scale. To address this we evaluate three lightweight LLMs, Claude-Haiku-4.5, GPT-5.4-Mini, and Gemini-3.1-Flash-Lite, on free-text 5G domain knowledge and fault-analysis tasks across three benchmarks, TeleQNA ORAN FT, 5G-Faults FT, and TeleInter FT. Three independent frontier judges score outputs, and pairwise inter-judge agreement is measured as an empirical test of the LLM-as-Judge methodology. All three models reach at least 90% accuracy on fault diagnosis, while zero-shot recall of 3GPP and O-RAN specifications remains the critical gap, with all models scoring below 60%. Mean inter-judge agreement is at least 0.90 across all runs, indicating that multi-judge LLM scoring produces consistent, reproducible grades for open-ended telecom responses. Operationally, Gemini-3.1-Flash-Lite offers the best efficiency trade-off, combining competitive accuracy with the lowest inference cost and latency, making it the most suitable candidate for production telecom deployments.
展开完整摘要收起摘要↓
Real-world fault analysis in 5G and emerging 6G networks demands domain expertise to analyze free-text diagnostics, including root-cause explanations and recommended actions. LLMs have emerged as a promising approach to automating this, yet whether lightweight, edge-deployable models are capable of performing in-depth free-text diagnostics remains an open question. While existing benchmarks rely on restrictive MCQs with fixed answer keys, this paper evaluates 5G domain understanding and fault analysis in a free-text generation format. Transitioning to this paradigm requires evaluating lightweight, edge-deployable AI models on open-ended diagnostic reasoning, alongside a dependable framework to validate these text outputs at scale. To address this we evaluate three lightweight LLMs, Claude-Haiku-4.5, GPT-5.4-Mini, and Gemini-3.1-Flash-Lite, on free-text 5G domain knowledge and fault-analysis tasks across three benchmarks, TeleQNA ORAN FT, 5G-Faults FT, and TeleInter FT. Three independent frontier judges score outputs, and pairwise inter-judge agreement is measured as an empirical test of the LLM-as-Judge methodology. All three models reach at least 90% accuracy on fault diagnosis, while zero-shot recall of 3GPP and O-RAN specifications remains the critical gap, with all models scoring below 60%. Mean inter-judge agreement is at least 0.90 across all runs, indicating that multi-judge LLM scoring produces consistent, reproducible grades for open-ended telecom responses. Operationally, Gemini-3.1-Flash-Lite offers the best efficiency trade-off, combining competitive accuracy with the lowest inference cost and latency, making it the most suitable candidate for production telecom deployments.