每天知道
你的领域
出了什么

Large Language Models (LLMs) are evolving from performing end-to-end mathematical reasoning to integrating agentic intelligence. However, most existing math benchmarks evaluate only final answers. This outcome-oriented evaluation provides limited diagnostic value for identifying process-level failures or rigorous logic, failing to guide the transformation of LLMs into robust agents. To bridge this gap, we present a process-level benchmark designed to evaluate the inherent agentic mathematical reasoning abilities of LLMs. Our framework aligns problem-solving agentic behaviors with a structured taxonomy of reusable mathematical atomic capabilities. We design a comprehensive suite of planning, action, and feedback tasks across both textual and multimodal contexts, supported by an automated pipeline that synthesizes high-quality trajectories and produces fine-grained annotations via controlled LLM rewriting. Experiments reveal that models with similar end-to-end accuracy can exhibit markedly different agentic capability profiles. This demonstrates that process-level evaluation is crucial for interpreting the true potential of LLMs and guiding the development of next-generation mathematical agents.

Large language model agents are increasingly deployed as autonomous loops. Starting from one human goal, such a system repeatedly discovers work, plans, executes tool calls, verifies outcomes and persists state across many unattended iterations. The agent safeguards in wide use, however, are defined over a single trajectory, and their safety state is re-initialized when the next trajectory begins. We show that this is a failure of composition rather than an implementation detail. Our central result is a separation: against an attack whose evidence is fragmented across several iterations, every trajectory-scoped monitor has a true-positive rate equal to its false-positive rate, however expressive it is, because the evidence it would need never appears in the window it sees, whereas a monitor retaining cross-iteration state separates the two perfectly. We further show that the obvious repair of carrying a geometrically decaying risk score is insufficient, because the cooling-off period a patient adversary must wait is a constant that does not grow with the horizon $N$. We then present LoopHarness, which restores a persistent, non-decaying safety state at the loop level. Under mediated commits and an arbiter detection floor $δ_M$, it bounds the expected number of unauthorized irreversible actions by $B+m-1+m/δ_M$, a constant in $N$, of which the $B+m-1$ term is decided by a model-free rule and therefore survives a fully colluding verifier. We give a complete evaluation protocol on native Agent-SafetyBench tasks with paired clean and attacked episodes, an outer-state attack suite whose decisive evidence exists only across iterations, per-module ablations, and an adaptive white-box red team.

Autonomous agents are increasingly adopted to complete complex, multi-tool workflows in real-world settings. However, existing benchmarks typically separate tasks by application or capability and evaluate agents in environments that are cleaner and more stable than those encountered in practice. We introduce DuMateBench, a real-session benchmark reconstructed from anonymized and privacy-screened user sessions collected from a large-scale production agent platform. Each task preserves the relevant pre-solution interaction history, persistent configurations, and workspace state, and is then validated through human verification. The resulting benchmark comprises 200 tasks spanning 8 broad scenarios and 17 fine-grained capability categories, with most tasks requiring multiple capability coordination. We execute these tasks in isolated Docker containers injected with three forms of real-world environmental complexity: Insufficient, Unstable, and Noisy, and assess performance using a hybrid deterministic and LLM-as-Judge evaluation protocol. Experiments across five representative autonomous-agent frameworks paired with four state-of-the-art LLMs reveal substantial gaps in strict task completion. Complementary robustness, efficiency, and diagnostic analyses further show that performance under environmental perturbations is jointly shaped by the capabilities of the LLM and the surrounding agent framework. The code and data are publicly available at https://dumatebench.com/.

Large language models can summarize financial information, but an operational stock-research system must first assemble heterogeneous evidence, expose unavailable data and model capabilities, and control how generated opinions affect a final report. We present DSA, an evidence-aware orchestration framework for multi-market stock research with large language model (LLM) agents. DSA organizes the workflow into evidence acquisition, structured context construction, model-routed analysis, optional role and Strategy Skill reasoning, and report generation with selected context and diagnostics. A default report profile and an optional agentic profile share evidence and model-routing services but use profile-specific output validation and risk safeguards. In the agentic profile, core role outputs are processed by role-specific parsers, whereas Strategy Skill opinions undergo an additional signal-eligibility partition before synthesis; disagreement is supplied explicitly to the decision agent, followed by a conservative risk override. The reference implementation includes six regional market paths, fifteen bundled Strategy Skills, hosted and local model routes, and multiple execution and delivery surfaces. At a frozen software snapshot, a selected manifest of 1,457 portable offline backend contract tests passed; 596 cases were retrospectively mapped to six contract families central to the reported LLM-agent architecture. This evidence establishes implementation conformance for the tested software contracts, not superior report quality, forecasting accuracy, or investment returns.

LLM judges are widely used to evaluate agentic tool-calling systems, yet their reliability on structured, dependency-driven workflows remains largely unexamined. We present AgentJudgeBench, the first benchmark to systematically study LLM-as-a-judge reliability for agentic tool-calling over workflow DAGs, as distinct from the broader LLM-as-a-judge task of open-ended text or preference evaluation. The benchmark comprises 3,808 instances spanning six DAG topologies and three difficulty tiers, evaluated with five generators (3B-70B open-weight models and GPT-5.4) and six judges (20B to frontier scale) under paired with- and without-ground-truth conditions. Judge alignment degrades monotonically with task difficulty, 1.5x faster without ground truth, and on hard queries without ground truth all six judges converge to a narrow 77-82% band regardless of scale, revealing a structural ceiling driven primarily by task difficulty, though its height is partly prompt-dependent for weaker generators, that model capacity alone cannot overcome. Ground-truth exposure is not uniformly beneficial: it reduces alignment for GPT-5.4 (1.5 pp) and Gemini-2.5-Pro (3.9 pp), consistent with over-anchoring. Among mitigation strategies, chain-of-thought reasoning and judge temperature both have negligible effect, while structured evaluation rubrics improve alignment by up to 6.5 pp but do not generalize uniformly across judge-generator pairs. With ground truth, QwQ-32B best matches the programmatic reference, while a human validation study identifies GPT-OSS-120B as the most human-aligned judge; without it, frontier judges lead only marginally within the shared ceiling. These results expose fundamental limitations of current LLM judges and yield practical guidelines for reliable evaluation in agentic systems.

Dense warehouses often contain single-lane aisles, dead ends, and tree-like guidepaths that leave little room for idle agents to wait without blocking others. Existing Multi-Agent Pickup and Delivery (MAPD) guarantees for completing all finitely released tasks typically rely on extra waiting endpoints that planned paths can avoid, or on biconnected topology; these assumptions may fail in such layouts. We study fixed-Haven reservation for online MAPD, where pickup-delivery tasks are released over time. Each agent owns a fixed Safe Haven (Haven for short), usually its start cell, that only the owner may occupy and that other agents treat as blocked. For finite task releases, we prove that this fixed-Haven contract completes all released tasks under Haven-Reachability and explicit planning/progress assumptions. We implement the contract in SHARP, a Safe-Haven Retreat Planner that keeps every busy or retreating agent on a collision-free reserved route ending at its Haven. We compare SHARP with representative TP and PIBT-family MAPD baselines: Token Passing (TP), Priority Inheritance with Backtracking (PIBT), and PIBT with Temporary Priority and Temporary Avoidance (PIBTTP-TA) for biconnected main areas with attached trees. In the robustness sweep, SHARP is the only method with 100% success on all tested configurations, at substantially higher centralized planning cost on tree-like layouts. A TP-style fixed-home-return counterfactual with full-route validation also recovers robustness on tested tree-like layouts, suggesting that fixed return is a central robustness mechanism there. A no-overwrite variant shows that disabling mid-retreat reassignment worsens service time (release-to-delivery latency) by 1.89 times and makespan by 1.53 times in the tested high-load tree condition.

Space-efficient warehouse layouts often contain single-agent-width aisles and dead-end workstations where robots have few places to wait without blocking others. In Multi-Agent Pickup and Delivery (MAPD) on such constrained layouts, robots must accept online pickup-delivery tasks while preserving protected waiting locations called Havens. The Safe HAven Retreat Planner (SHARP) introduced a mechanism that extends each committed task path with a validated retreat to the agent's dedicated initial Haven, but fixed-Haven commitments can send agents toward distant Havens after deliveries. We present A-sharp (Adaptive SHARP), which changes an agent's retreat target at task assignment time. A naive switch can cause two agents to rely on the same waiting location or let another committed path pass through a location that is still occupied or reserved. A-sharp prevents these failures with an availability test for candidate Havens and a pending-release rule that keeps the previous Haven protected until the agent departs. Under explicit Haven-structure and Safe Interval Path Planning (SIPP) assumptions, we prove invariant preservation and finite-release completeness: every task in any finite release sequence is delivered in finite time. Across 72,000 runs on 14,400 paired map-agent-count-rate-seed cases over four maps, both SHARP and A-sharp complete their respective 14,400 runs. For makespan (final delivery time), a prespecified paired comparison with Holm correction over all 138 configurations with more Havens than agents finds A-sharp significantly better in 107 configurations and never significantly worse than SHARP; on the tested tree map, the median reduction is 16.7%.

Internet memes are a pervasive form of multimodal online communication; however, such communication often involves users from diverse linguistic and cultural backgrounds. Therefore, adapting memes across cultures and languages is a central challenge for enabling mutual understanding in online communication. Unlike ordinary translation or standalone text rewriting, cross-cultural meme transcreation must jointly preserve communicative intent, adapt culture-dependent meaning for the target audience, and maintain coherence between text and image. In this work, we first provide an explicit task analysis of cross-cultural meme transcreation and identify three core challenges: culture-specific knowledge understanding, intent and tone preservation, and multimodal consistency. Based on this analysis, we propose a multi-agent framework with specialized agents that are coordinated to address these challenges through cultural adaptation, target text rewriting, revision, and conditional visual adjustment. The framework strengthens target text adaptation with coordinated feedback to handle difficult cases that require deeper cultural or visual intervention. We evaluate the framework on bidirectional Chinese-English meme transcreation using both human evaluation and LLM-as-a-Judge. Our method consistently outperforms all baselines across both evaluation settings. In human evaluation, it achieves the best performance on all four dimensions and delivers a 33.1% average improvement over the strongest baseline, while in LLM-as-a-Judge, it attains the highest Top-1 ranking rate (60% versus 26% for the second-best baseline). Further analysis indicates that each component contributes to the performance. Our error analysis suggests that the remaining bottlenecks lie in humor reconstruction and image-text alignment rather than simple cultural knowledge gaps, pointing to future work on humor transfer.

Scientific LLM agents have shown promise in literature reasoning, tool use, and experiment planning, but it remains unclear whether they can autonomously improve large, tightly coupled scientific machine-learning systems through executable code changes and computationally expensive validation. We study this question in protein folding, where progress requires coordinated architectural modifications, multi-objective evaluation, and domain-aware interpretation. We present AgentFold, a multi-agent framework that formulates folding-model development as a closed-loop search over executable code variants. Starting from ESMFold, AgentFold proposes hypotheses, implements and debugs code-level modifications, evaluates model variants, analyzes experimental outcomes, and stores both successful and failed interventions in structured memory. An MCTS-style policy allocates computational resources across high-scoring search branches. On an engineering-scale protein-folding codebase comprising more than 2,000 lines of code, AgentFold explores approximately 80 model variants using approximately 5,000 GPU-hours and 170 million LLM tokens. Under a matched computational budget, AgentFold improves the best lDDT by 7.5% over independent Codex proposals and outperforms a random-search control. Beyond model improvement, the resulting intervention traces reveal recurring empirical design patterns: stable gains tend to arise from early, soft, learnable priors and gated refinement, whereas direct geometric perturbations and geometry-conditioned feedback often destabilize training. The code and experimental resources are publicly available at https://github.com/lmqfly/AgentFold.

Generated operational programs are often validated with either a few hand-written examples or exhaustive regression suites. The former can miss sparse boundary and interaction faults, while the latter can be unnecessarily expensive. We introduce FaultLens, a method for learning compact behavioral test suites while preserving an auditable connection to executed evidence. It executes a rich probe domain once, stores the fault-probe kill relation as a sparse outcome cache, and learns probe orderings only from earlier program generations. A fault-driven greedy component exploits known kill structure, while a mutation-independent diversity component covers probe families, cases, templates, and temporal bins. Their alternating hybrid remains useful when a new program contains a fault mechanism absent from ordering construction. We evaluate twenty generated operational policies across four environments, ten execution seeds, 1,200 measured run summaries, 2,160 controlled program transformations, and 4,120,200 executed program-probe pairs. Of 1,960 intended faulty transformations, 1,779 alter a contract or output somewhere in the finite audit domain; 200 additional controls preserve behavior. A 32-probe hybrid learned on generations 1-3 covers 576/582 (99.0%) dynamically killable faults in generations 4-5 using 1.2-2.0% of the exhaustive domain. With an entire fault family withheld from training, diversity raises scenario-family macro coverage from 84.6% to 94.9%. In a downstream deployment study, a conservative admission rule reduces severe tail regressions from 15/20 program-environment groups to 0/20. FaultLens provides a prioritized evidence mechanism, not a proof of correctness, and makes its budget, evidence source, generalization split, and misses explicit.

With the growing demand for processing multiple image sequences in real-world applications, various visual token pruning methods have emerged to mitigate the computational and context length constraints faced by Large Vision Language Models (LVLMs). However, most existing pruning approaches rely on static strategies that struggle to adapt across different architectural LVLMs and multi-image scenarios, and are additionally constrained by their dependence on attention computations that are incompatible with efficient techniques like FlashAttention. To address these limitations, we propose a training-free, Adaptive Visual Token Pruning (AVTP) framework, applicable to diverse LVLM architectures. We strategically determine pruning layers based on empirical analysis of visual attention distributions across various LVLMs, and implement adaptive pruning ratios in multi-image contexts where images of higher importance retain proportionally more tokens. We conduct extensive experiments across different LVLMs to demonstrate the effectiveness and robustness of AVTP. Specifically, Qwen3VL-8B achieves 2 times inference speedup while maintaining 96.1% of its original accuracy on multiple multi-image benchmarks, InternVL3.5-8B retains 94.1% accuracy, and LLaVA-OV-7B even exceeds its original baseline performance. Our code is available at \href{https://github.com/zry13/AVTP}{this link}.

Diffusion-based Video Virtual Try-On (VVT) achieves high visual fidelity through bidirectional spatio-temporal modeling, but complete-clip dependence incurs prohibitive latency and computational overhead in practical continuous deployment. Naively enforcing causality disrupts pretrained bidirectional priors and substantially degrades synthesis quality. We introduce LiveVVT, a rolling streaming diffusion framework that preserves bounded bidirectional modeling within causal recurrent generation. Within a fixed-size window, LiveVVT jointly denoises multiple video chunks under bounded look-ahead, preserving local bidirectional interactions while emitting one clean chunk per iteration. Beyond the window, two complementary memories sustain long-term consistency: a bounded temporal memory propagates recent dynamics and occlusion context, whereas a persistent global appearance memory, constructed once from the target garment and a frontal try-on keyframe, anchors garment details and dressed appearance throughout the stream. We further introduce a progressive distillation framework integrating bidirectional VVT learning, teacher-trajectory regression for causal few-step adaptation, and Collaborative Matching Distillation, which couples teacher-distribution matching with rolling flow matching on real videos to align optimization with recurrent inference. Experiments on paired and unpaired long-sequence benchmarks demonstrate superior generation quality over similarly sized models, with $26\times$ lower latency and $11\times$ higher throughput, enabling high-fidelity real-time streaming VVT.

Scaling video generation to long durations reveals a critical bottleneck: current models lack robust long-term memory. This deficiency can be studied along two critical aspects: object permanence, the ability to precisely reproduce the appearance of objects upon re-entry; and memory capacity, the ability to process ultra-long context and use information from distant history. Robust long-term memory requires both: object permanence without sufficient context handling limits the temporal scope, while long context length without permanence fails to maintain identity. To address this, we present Ring Forcing, an autoregressive video diffusion framework designed to robustly construct and precisely utilize long-term memory. Our ring-structured training strategy enforces retrieval from distant history, effectively reconciling the trade-off between strict historical adherence and generative diversity. To expand memory capacity, we introduce a compression and timestep composition strategy. Under fixed sequence length constraints, this method extends the effective historical span to minutes-long durations and achieves a comprehensive receptive field over the entire history. Furthermore, we present a sparse RoPE mechanism to enable flexible, scalable memory adaptation while fully exploiting pre-trained priors. Extensive experiments demonstrate that Ring Forcing achieves superior minutes-long coherence and object permanence, significantly outperforming state-of-the-art methods.

Autoregressive large language models (LLMs) routinely generate factually incorrect outputs with high decoding confidence, limiting their deployment in high-stakes workflows. Existing output-stage uncertainty metrics can fail when models are overconfident on false assertions, while multi-sample verification pipelines introduce substantial memory and latency overhead. This work evaluates whether internal hidden-state transition dynamics during generation can signal factual errors without auxiliary decoding calls. We introduce Prediction of Prediction (PoP), a mechanism that captures layer-transition uncertainty by fusing intermediate hidden representations across depth during a single forward pass. Evaluated on the TruthfulQA benchmark using autoregressive transformer backbones, PoP achieves an area under the receiver operating characteristic curve (AUROC) of 75.5% for factual-correctness classification. The mechanism operates within the base forward pass, adding less than 1.2% runtime latency and requiring zero additional generation passes. The numerical results are reported from the author-verified experimental implementation and are bounded by the evaluation scope described below.

Evolution Strategies (ES) have recently emerged as a memory-efficient post-training paradigm for LLM reasoning. However, the optimization behavior of ES remains understudied, making it hard to define its advantage scope compared to mainstream post-training paradigms (e.g., Group Relative Policy Optimization (GRPO)). By systematically investigating ES dynamics and mechanisms, this paper first identifies a performance advantage of ES over GRPO, theoretically and empirically showing that ES can lead to broader reasoning coverage, thereby better exploiting the reasoning capabilities of pretrained LLMs. Theoretically, we show that verifier-projected Jensen-Shannon diversity across the ES population is helpful to higher Pass@K performances. Empirically, unlike GRPO, which exhibits entropy collapse, ES improves Pass@1 while attaining higher Pass@K than GRPO. We further develop a sequential GRPO-ES training strategy that combines GRPO's strength in Pass@1 with ES's gains in Pass@K. Second, we find that despite substantial whole-model parameter drift, the task-performance gains of ES are only contributed to a sparse subset of larger-magnitude updates. This functional sparsity suggests that large parameter movement need not imply widespread functional change, and held-out evaluations further show that it does not necessarily lead to catastrophic forgetting. Finally, we study how hyperparameter design affects the effectiveness of ES, demonstrating that ES requires a smaller population size in a larger LLM. These findings position ES as a distinct reasoning post-training paradigm rather than a less effective, memory-efficient alternative to GRPO.

This paper measures the performance impact of running large language model inference and training inside a Trusted Execution Environment (TEE) on NVIDIA B200 GPUs, using Intel Trust Domain Extensions (TDX) confidential VMs together with NVIDIA Confidential Computing (CC) on Blackwell GPUs. The performance impact is derived from paired confidential versus non-confidential runs on a single physical host where the only variable is the GPU CC bit and the TDX guest object in the VM launch. The main result is that confidential inference on Blackwell achieves low single-digit throughput overhead when the stack is configured correctly, at about 1-3%. Stock inference stacks incur 30 to 40% penalties due to avoidable configurations rather than the achievable operating point. The cost is not fully represented by a single number because it is governed by two independent axes, a fixed per-host-operation cost that amortizes as batch size grows and a per-NVLink-traffic cost that tracks the share of the step spent in encrypted collectives, and which of the two dominates is set by the workload and the software. We localize each cost to a specific encrypted boundary, give a microbenchmark that predicts the serving penalty to within a submission count, and end with concrete deployment guidance. GPU compute, energy draw, and usable memory capacity are unaffected by CC.

This paper introduces ClusterAttention, a general training-free speedup of bidirectional attention layers. Existing sparse attention methods either rely on structure in the input, such as order in language or spatial proximity in images, or use slow clustering processes amortized over several forward passes. ClusterAttention instead uses a fast recursive clustering method that adapts to the geometry of the keys and queries in each attention head to produce useful clusters. This method allows setting the size of the clusters arbitrarily. We utilize this by setting all clusters to be a fixed size that is a power of two, allowing the block-sparse attention to run at the same latency per query-key interaction as dense attention on GPUs. We also derive an expression for the output error in sparse attention, that explains the counterintuitive experimental finding that tight clusters can lead to larger errors than random clusters. We then derive the error when excluded clusters are compensated through their centroids, and show that this error shrinks with tighter clusters. We integrate this compensation into the method. On large-scale tabular data ClusterAttention speeds up TabPFN-3 arXiv:2605.13986 by two to six times, while retaining at least 99% of the dense accuracy. To our knowledge, it is the first training-free method that can be successfully applied in the setting of unstructured input and a single forward pass. For video generation with Wan 2.1-14B T2V arXiv:2503.20314 , ClusterAttention achieves output closer to dense attention and a larger speedup (1.8x versus 1.4x) compared to SVOO arXiv:2603.18636 , a leading method developed specifically for this domain, both run without offline calibration.

Vision-Language-Action (VLA) models are increasingly promising for robotic manipulation, yet their real-world deployment remains bottlenecked by high inference latency and unstable asynchronous execution. This challenge is particularly pronounced in flow-matching-based VLA models, where action decoding requires multiple iterative steps conditioned on the VLM context. While efficient inference methods improve control frequency and asynchronous methods reduce execution idle time, existing approaches often fail to jointly achieve low-latency inference and accurate, temporally consistent asynchronous execution. We introduce FlashVLA, a streaming action decoding framework that addresses both challenges in a unified formulation. FlashVLA maintains a streaming action buffer with multiple chunks at different noise levels and decodes them using chunk-wise causal attention. This design allows FlashVLA to produce one executable action chunk per inference step. Moreover, its chunk-wise autoregressive formulation implicitly preserves action continuity, enabling smooth asynchronous execution without extra future-state conditioning. Across extensive simulated and real-world experiments, FlashVLA substantially improves inference speed while maintaining strong task performance. It can achieve $\geq$30\,Hz control frequency on a single GPU with smooth asynchronous inference in real-world deployment.

Vision-Language Models (VLMs) demonstrate exceptional visual reasoning capabilities, yet their inference costs escalate rapidly with the proliferation of visual tokens. Existing visual token pruning methods exhibit two fundamental limitations. First, most approaches operate exclusively post-vision encoder, leaving the substantial latency of the visual encoding phase unoptimized. Second, under strict token budgets, these methods often fail to jointly preserve holistic visual contexts and fine-grained details, leading to performance degradation. To address these bottlenecks, we propose PACE (Pixel-Adaptive Condense and Extract), a training-free inference framework that accelerates both the vision encoder and the Large Language Model (LLM) via a unified Condense-and-Extract paradigm. During the Condense stage, an Adaptive Pixel Compressor (APC) evaluates visual information density prior to encoding, adaptively downsampling redundant inputs, curtailing encoder computation while preserving global context and essential visual cues. In the Extract stage, a Dynamic Dual-Attention Extractor (DDAE) selectively retains visual tokens via a fusion of internal visual signals from the encoder and semantic signals from the LLM, safeguarding task-critical details. By integrating PACE into Qwen2.5-VL-7B, the model retains 93.8% of its original performance while utilizing only 10% of the visual tokens, yielding a 3.1x speedup in time to first token (TTFT). Our code is available at https://github.com/jjL357/PACE.

Long autoregressive video generation faces a fundamental memory challenge: with a finite attention window, a model must decide which information from an ever-expanding history to retain. Existing methods organize memory temporally, preserving recent frames while compressing or discarding older ones. We instead propose RECAP-Forcing, organizing memory by appearance novelty. A long video is not merely a sequence of frames, but an evolving cast of subjects, objects, and scenes whose identities must remain consistent over time. We organize memory by retaining the KV cache associated with newly appearing content--such as entering subjects, disoccluded regions, and newly introduced scenes--at the moment it first becomes visible, prioritizing novelty over recency. Memory should scale with the amount of newly introduced content, rather than with video length. This appearance-indexed memory makes long-range consistency an explicit property of the memory structure. Our framework unifies two mechanisms under this single principle. At the beginning of a video, when all visible content is novel, an attention sink preserves the initial scene. As the video evolves, an optical-flow-based novelty bank extends the same principle by selectively retaining newly revealed content. As a training-free inference method with no additional learnable parameters, RECAP-Forcing consistently improves visual quality and semantic fidelity across multiple strong baselines and outperforms existing memory methods.

Small language models (sLLMs) are nowadays hosted on devices with limited memory and computational budget. In an autoregressive setup, inference is memory-bandwidth bound: uniform quantization is often detrimental to such models, since their architecture has limited redundancies and only a few layers are not very sensitive to lower precision. We propose a composite metric that combines two orthogonal criteria: information retention (measured in terms of a normalized SQNR-based coefficient) and throughput gains (modeled using a roofline-based latency analysis). By profiling Gemma 3 1B, we find that Feed-Forward Network blocks and the embedding matrix are the most promising targets for acceleration. For each candidate, we estimate a normalized quality score based on simulated quantization and a normalized speed score based on roofline modeling with no actual execution needed. We combine the two scores in a composite priority coefficient, allowing us to tune the trade-off between speed and quality as needed. Our metric is general and can be used to prioritize individual blocks, their projection sublayers, or transformer layers as a whole. We evaluate our approach on several model architectures, showing that our estimates have at around 4% prediction error for the accelerated speedup. We find that our method generally allocates more resources to the most expressive layers compared to evolutionary search, specialized accelerators, or Shapley-value-based approaches that require expensive approximate inference. Our analytical approach makes sLLM quantization a predictable engineering task.

Mixture-of-Experts (MoE) models route each token to a few of many expert networks, and that routing is data-dependent in a way standard inference optimizations do not expect. This paper measures what three of them actually deliver on OLMoE-1B-7B, DeepSeek-V2-Lite, and Qwen3-30B-A3B. Fused Triton kernels reach 5.6x to 9.0x in isolation but 0.999x end to end against a measured 1.07x ceiling, because the model spends its time waiting on roughly a thousand kernel launches per forward pass rather than on the arithmetic those kernels improve. INT4 quantization changes on average 0.53 of the eight selected experts per token position, yet replaying exactly those changed routes through full-precision weights reproduces only 2.7% of the quality loss, which makes the experts substitutable rather than specialized. Removing all 23 graph breaks from PyTorch's compiler, the step prior work treats as the structural fix, makes the model three times slower. A fourth result ties the three together: leaving the routers in FP16 lowers drift by 20% while raising loss, so routing fidelity and output quality are separable objectives. Every number recomputes from committed per-token route dumps.

Chunked prefill pipeline parallelism (CPP) is a key technique for LLM inference. However, equal-size chunks exhibit imbalanced latency, as later chunks attend longer prefix KV caches and incur higher attention costs, leading to pipeline bubbles. Existing approaches mitigate this imbalance through dynamic chunk resizing (Dynamic CPP, DCPP), but our measurements show that this trades scheduling overhead for load balancing, which becomes unfavorable on long sequences. In this study, we propose Virtual Pipeline Parallelism (VPP), which keeps chunk sizes fixed and optimizes the pipeline layout through virtual stages. A V-shaped virtual-stage traversal overlaps each chunk's expensive middle stages with the lighter head and tail stages of its neighbors, while asynchronous communication and pipelined packing further reduce communication stalls and cross-request drain bubbles. We implement VPP in vLLM-Ascend and evaluate it on three MoE-based LLMs with sequences up to 1M tokens on 16 Ascend 910C NPUs. VPP improves throughput by up to 13.1% over DCPP on long sequences and 6.7% on mixed workloads, while preserving performance on short sequences. On a 512K-token DeepSeek-V3.1 prefill workload, VPP reduces the pipeline bubble ratio from 6.4% to 0.1%, achieving a 98.0% reduction compared with DCPP.

Long-context inference is bottlenecked by the memory footprint of the key-value (KV) cache, especially for small models under tight resource budgets. Existing KV cache eviction methods score tokens using the model's attention distribution or, in attention-free variants, each key's distance from a global reference point. Using a controlled leave-one-out probe, we find that attention magnitude is unrelated to a token's causal contribution to the answer (Spearman $ρ=-0.004$), challenging the premise behind dominant eviction methods. We introduce TwinKV, a training-free, attention-free redundancy signal that detects whether a token's key has a near-duplicate elsewhere in context. Rather than replacing existing policies, TwinKV acts as a composable repair pass: given a policy's fixed retained set, it identifies evicted tokens with no surviving duplicate (orphans) and retained tokens whose information is duplicated elsewhere (redundant donors), then swaps them while preserving the original budget and scoring rule. We compose TwinKV with four recent eviction policies across LongBench, LooGLE, RULER, and a short-context MMLU-Pro no-harm control at compression ratios ${0.3,0.5,0.7}$. On Qwen3-4B, TwinKV improves a majority of configurations for two policies, is near-even for a third, and helps only a minority for a fourth adaptive baseline already near a performance ceiling; gains across the three non-ceiling policies are smallest at the loosest ratio. On RULER with Llama-3.2-1B, however, that fourth policy improves in every evaluated cell because its Alone score leaves substantial room to improve. More broadly, Llama-3.2-1B shows a smaller average LongBench gain but a higher fraction of improved cells on LongBench and LooGLE than Qwen3-4B, plus a clean RULER win. We also identify few-shot classification exemplars as a task structure where TwinKV does not help on either model.

Text-guided drone geo-localization aims to identify a target region in a large-scale image gallery from a natural-language description. Existing methods mainly formulate this task as direct matching between an open-ended text query and candidate images. However, incomplete queries and highly similar candidates often make global cross-modal matching insufficient for reliable fine-grained localization. We propose UniGeo, a unified multimodal large language model (MLLM) for text-guided drone geo-localization. Built on a shared vision-language framework, UniGeo jointly supports geo-semantic understanding, cross-view semantic generation, and candidate-level verification. Specifically, it establishes stable correspondences among local scene elements, spatial relations, and language descriptions through geo-semantic learning, and further models semantic mappings between drone and satellite views through cross-view generation. Based on these capabilities, a plug-and-play verification module performs fine-grained discrimination among highly confusable candidates. We further introduce a multi-stage training strategy that progressively learns geo-semantic understanding, cross-view generation, and candidate verification, improving adaptation to text-guided geo-localization. Experiments demonstrate consistent improvements across multiple retrieval backbones. On GeoText-1652, UniGeo improves R@10 and mAP by 13.59 and 2.83 percentage points, respectively, validating its effectiveness for fine-grained text-guided drone geo-localization.

Graph neural networks (GNNs) are typically conceptualized as message-passing neural networks, yet it remains unclear why neighborhood aggregation reliably outperforms node-wise multilayer perceptrons (MLPs). Despite its empirical success, this paradigm can be computationally expensive and sensitive to imperfect graph structures. In this work, we present a retrieval-augmented view of GNNs: each layer makes predictions by applying an MLP to a node representation together with a permutation-invariant summary of retrieved graph context. Motivated by this perspective, we propose RTA, a simple MLP-based framework that replaces structural message passing with label-aware retrieval and propagation. We provide theoretical insights that (i) connect retrieval-based aggregation to softmax-attention message passing, and (ii) establish the robustness of retrieved-context supervision to mis-retrieved outliers. Experiments on multiple text-attributed graph benchmarks show that RTA matches or even outperforms strong GNN and graph LLM baselines while improving efficiency and robustness across diverse scenarios.

Although batch prompting makes large language model inference more efficient by processing multiple instances simultaneously, it suffers from unpredictable downstream task performance. We propose cascaded batch prompting, a two-stage approach designed to resolve the unpredictability of conventional batch prompting by disentangling complex reasoning from symbol grounding. Experiments on multiple-choice question answering and natural language inference demonstrate that the proposed method outperforms the standard single prompting baseline while achieving a speedup proportional to batch size, establishing a new state of the art on the Pareto frontier.

Cross-lingual alignment in multilingual language models is typically attributed to joint training: shared parameters, mixed-language batches, or explicit alignment objectives. We ask whether monolingual models trained on non-parallel data learn alignable representations without joint training. By testing on strictly monolingual language models, such as the Goldfish model families and independently developed models from different research labs, we find three results. Correlation: these models develop alignable representational geometry across layers, with alignment strengthening as data scale, model scale, or linguistic proximity increases. Construction: a single Procrustes rotation fit on parallel sentences maps hidden states between models. Causation: the same rotation transfers functional content; patching a rotated English residual into a German model on a factual cloze flips the prediction to the donor's capital in most cases. We confirm that cross-lingual alignment can emerge from the structure of language and the information it carries rather than from joint training, and this points to practical future directions including model stitching, merging, and modular multilingual systems built from monolingual components.

Latent chain-of-thought models move intermediate reasoning from emitted text into continuous states, improving compactness but hiding the causal object. We introduce SCIT, the Suffix Cache Interchange Test, a causal protocol that constructs exact source-recipient counterfactuals, patches declared cache segments, and identifies which transformer object carries the counterfactual computation. SCIT combines sufficiency tests with K/V component splits, hidden-state controls, semantic source controls, decoded validation, and matched corruption. On CODI-GPT2 and a Sim-CoT-style GPT-2 reproduction, counterfactual arithmetic transfers primarily through value-cache suffix trajectories rather than hidden states, keys, reusable answer slots, or single-token triggers. Complete sufficiency-and-necessity evidence for the late-value-suffix mechanism holds for the main CODI-GPT2 checkpoint; the Sim-CoT-style checkpoint shows the same sufficiency and decoded-control pattern but insufficient matched-corruption evidence for a necessity call. Beyond these local arithmetic cells, SCIT reveals carrier-regime shifts: arithmetic-like GPT-2/1B cells preserve latent-tail value/KV transfer, whereas competent 8B and repaired non-arithmetic cells route through prompt-prefix or full-cache K/V; boundary cells receive no mechanism call. SCIT therefore contributes a cache-level diagnostic, a checkpoint-specific GPT-2 arithmetic mechanism, and a competence-gated carrier map rather than a universal latent-tail claim.

We ask whether large language models (LLMs) can design effective algorithms for well-specified operations research (OR) problems. We study inventory control, queueing network control, and assortment optimization. We evaluate two levels of LLM use: at level 1, the model receives one problem instance and returns a solution for that instance; at level 2, it receives only the problem class description and broad parameter ranges, and returns an algorithm that maps instance parameters to solutions. Human input is minimal: we give one untuned prompt that describes the problem, and the model has access to a Python sandbox tool with a fixed compute budget. The strongest model we test, gpt-5.6-sol, matches or outperforms the best existing method on almost all evaluated instances. This holds even at level 2, where the returned algorithm is fixed before seeing the evaluation instances. Performance also improves sharply across models released less than eight months apart, suggesting that this capability is moving quickly. Thus, for the well-specified operations problems we study, a single untuned LLM query can already produce algorithms competitive with specialized methods. These results suggest that frontier LLMs can be a serious empirical baseline for algorithm design in well-specified OR problems.