Agent 进展

Computer-use agents have become increasingly capable of executing tasks on live desktops through natural-language instructions, based on trajectories of screenshots, actions, and reasoning. We discover that they can stealthily exhibit inertia, in which they repeat fruitless actions despite recognizing that these actions are ineffective. We hypothesize that inertia is reflected in the agent's internal state, i.e., the activation values of the agent's underlying model, and propose a protocol to measure the relationship between the two. Extensive analysis of high-dimensional activation states shows that inertia corresponds to an absorbing region of activation space, where activation values become stale across actions and even after attempts to steer them. We conjecture that drastically changing the agents' activations by re-initializing them is necessary to escape inertia. Specifically, we propose R$^3$ (Reset, Reroute, Restore), which temporarily resets the agent's context trajectory to escape the absorbing region and then restores the historical context to effectively complete the task. Our approach yields 17-55% lower measured inertia across models relative to unmodified agents. These results suggest that changing the context can interrupt recurrence more effectively than directly steering the resulting activations. Our code is available at https://anonymous.4open.science/r/vlm-agent-defense-D076

Cell Painting is a high-content morphological profiling assay widely used for phenotype-based biological inference, with mechanism of action (MOA) prediction as a central application. Existing approaches largely formulate Cell Painting-based inference as representation matching, assigning predictions from nearby reference perturbations in morphological feature space. However, retrieved neighbors are often noisy and partially misleading evidence due to batch effects, non-specific cytotoxicity, phenotypic convergence, and source-dependent variability. We reformulate Cell Painting-based MOA prediction as a calibrated evidence reasoning problem, where retrieved neighbors are treated as uncertain observations that must be evaluated, compared, and sometimes rejected before supporting a mechanistic conclusion. We propose PhenoAIR, a reliability-aware multi-agent framework that maintains a candidate-centric evidence memory and performs controller-guided refinement over phenotype- and mechanism-side evidence. PhenoAIR uses offline reference-set calibration to weight evidence by source reliability, phenotype stability, and mechanism-level confusion. We evaluate PhenoAIR on a benchmark constructed from JUMP Cell Painting profiles and annotations, covering controlled, realistic, and discovery-oriented open-world MOA prediction settings. PhenoAIR outperforms representation-matching and LLM-based baselines across all settings.

Advances in the coding capabilities of LLM agents allow them to inspect and modify their own instructions, tools, and execution procedures. Existing approaches use this ability to search for improved agents through repeated downstream evaluation, which incurs substantial costs and ties the search to the evaluated tasks. We introduce SelfSearch, a reward-free search procedure in which agents modify themselves using records of previous self-improvement episodes. These records capture the reasoning, tool actions, and outcomes of earlier modification attempts, providing concrete experience for improving both task solving and self-modification. Without downstream reward signals during search, SelfSearch improves population-mean success over the initial agent in all six model--benchmark settings, with individual agents gaining up to 11.2 percentage points on Terminal-Bench 2.1. On SWE-bench Multilingual, an agent improves success by 5.0 percentage points while reducing execution cost by 38.5% on tasks solved by both the initial and evolved agents. SelfSearch achieves competitive task success with evaluation-guided search baselines at lower search cost. With only \$4.03 in search cost, it produces a harness that solves 82.0% of Terminal-Bench 2.1 tasks with DeepSeek V4 Flash under the settings of a public nine-harness comparison, matching the top-scoring harness, Codex. These results suggest that experience gained through self-modification can improve agents' downstream capabilities and efficiency.

Multi-agent LLM systems rely on message passing among specialized agents to accomplish complex tasks. However, an upstream agent may provide useful information or an incorrect answer that causes a downstream agent to override a correct answer supported by its own evidence. Prior work has not clearly separated the benefits of communication from the damage caused by incorrect messages. We study this problem with controlled experiments across five benchmarks and five receivers, keeping the downstream task and evidence fixed while comparing answers under three conditions: no message, the upstream agent's original message, or a message with the opposite conclusion. Our experiments reveal three key findings. First, messages often help when the downstream agent would otherwise answer incorrectly. Second, messages can also hurt: when the downstream agent would answer correctly without a message, an incorrect upstream message changes the answer in up to 32% of cases. Third, in 94% of audited harmful cases, the downstream agent copies the upstream's specific wrong answer--a pattern we term answer substitution. Removing unreliable messages recovers part of the lost accuracy, suggesting that communication should be selective based on upstream reliability and the evidence already available to the downstream agent.

Natural-sounding multi-agent conversational AI is increasingly deployed, fundamentally altering human-machine interaction and human information processing. While prior work largely focuses on algorithmic failure, this study investigates the cognitive ergonomics and socio-cognitive impact of algorithmic competence. We present and evaluate FORMS (Framework for Opinion and Rhetoric in Multi-agent Simulations), a low-latency architecture for spatially mediated human-machine dialogue, driven by distinct LLM-based personas and real-time concurrency resolution. To conduct a system test and evaluation of its psychological impact, we exposed an adolescent cohort (n=120) and an adult pilot group (n=25) to a live, moderated synthetic debate. Our findings reveal that exposure to highly competent multi-agent systems triggers "Cognitive Destabilization," fragmenting users' prior strategic consensus. Concurrently, we observe a "Regulatory Awakening" driven by the "Normality Paradox": fluid human-machine interactions inherently increase the baseline demand for external regulation. Furthermore, our pilot study suggests the presence of a "Truthfulness Paradox": despite understanding the risks of generative AI, participants in the adult cohort rated the synthetic debate as significantly more sincere than equivalent human discourse (Cohen's d=2.04). Supported by robust statistical effect sizes, this paper contributes the FORMS architecture and a replicable evaluation protocol, illustrating how high-fidelity conversational systems can reshape human information processing.

Recent efforts to scale tool-use post-training have largely centered on the synthesis of executable environments, which constitute only one component of a broader agentic interaction system comprising the environment, task, agent harness, and evaluator. Scaling environments in isolation, however, does not guarantee commensurate gains in model performance, because reliable learning signals depend on coherent interactions among all components of the agentic interaction system. To address this problem, we introduce WEFT (Whole-system Evolution For Tool-use Post-training), which couples scalable agentic interaction system construction, execution-driven self-evolution, and stable post-training. WEFT scales agentic interaction system construction across environment breadth, task complexity, and interaction diversity. Execution-driven self-evolution iteratively uses execution traces and state evidence to attribute failures and revise the responsible components, with fresh rollouts evaluating the changes and providing evidence for subsequent evolution rounds. For stable post-training at scale, WEFT addresses both optimization and execution reliability: prefix-preserving sampling retains verified progress and atomic-turn credit assignment localizes learning signals, while MegaMCP maintains isolated, recoverable state across concurrent rollouts over shared tool services. Extensive experiments across various models and benchmarks demonstrate the effectiveness of WEFT for tool-use post-training. WEFT-8B and WEFT-14B outperform all evaluated matched-size environment-scaling baselines on BFCL V4, $τ^2$-Bench, and Claw-Eval. In particular, WEFT-14B improves over Agent-World-14B by 6.41, 2.23, and 12.27 percentage points. WEFT-35B-A3B further extends these gains to more challenging long-horizon workflow benchmarks, including Toolathlon-Verified and AutomationBench.

Many computer tasks recur: the same workflow runs many times, with new inputs and from different starting states. Current computer-use agents re-plan every step of every run, which makes them costly and unreliable on such tasks. We introduce neuro-symbolic computer use, in which a recurring workflow is executed by a learned policy rather than re-derived by an agent on each run. The policy fixes the decisions that are stable across runs (ordering, variables, loops, and branches) in executable code, and delegates observation-dependent decisions, such as grounding and state checks, to neural models. We learn these policies with neuro-symbolic policy iteration: starting from one agent trajectory, it executes the policy, diagnoses failures with task-completion and step-level judges, and revises the code with a coding model informed by an agent's continuation from the point of failure, without access to the benchmark evaluator. Iterating on generated parameter and initial-state variants makes the policy reusable, and a pre-action verifier guards each state-mutating step at deployment. On OSWorld-Verified and ScienceBoard, the learned policies achieve the highest Pass^3 of all methods in all four settings, 3.6-15.8 points above the base agent, while cutting per-run cost by 15-217$\times$ and latency by 3.4-5.1$\times$. On OSWorld-Verified, policies built only on variants transfer to the held-out original tasks, exceeding AutoRPA by 8.6-17.5 points in Pass^3.

On-policy distillation (OPD) trains multi-turn language agents with dense teacher supervision on student-generated responses. However, standard think-then-act rollouts require lengthy reasoning before each short action, delaying environment transitions and experience collection. Generating actions directly reduces this delay but can degrade rollout quality. To address this, we propose ActFirst-OPD, an act-first, reason-later training framework that decouples environment interaction from full-response generation. The student infers and executes actions through reference-conditioned inverse dynamics using its current interaction context and a reference next observation, and switches to autonomous next-action prediction when the resulting transition deviates from the reference trajectory. From the collected interaction contexts, the student asynchronously generates full think-then-act responses for token-level teacher supervision. Experiments across 0.6B-, 1.7B-, and 4B-parameter Qwen3 students show that ActFirst-OPD achieves average wall-clock training speedups of $2.3\times$ on ALFWorld, $1.8\times$ on WebShop, and $4.9\times$ on ScienceWorld over Vanilla OPD. It matches or exceeds all compared OPD baselines in mean task success rate across eight of nine benchmark-model settings. These results demonstrate that reasoning need not block acting during multi-turn agent distillation.

Large language model (LLM) agents have demonstrated strong performance on complex web navigation tasks, yet they remain brittle in real-world settings where user intentions are underspecified and preferences are heterogeneous. In practice, users rarely provide explicit profiles, requiring agents to infer latent preferences from implicit signals. Despite its importance for deployment, this problem setting is largely underexplored in existing benchmarks. To address this gap, we introduce AdaptArena, a benchmark for evaluating test-time personalization of web agents via implicit preference inference. AdaptArena consists of 480 tasks, featuring both single-preference and double-preference scenarios. Each evaluation task must be solved by retrieving and leveraging the most relevant historical user trajectory that implicitly encodes the target preference. In addition, we introduce AdaptiveAgent, a retrieval-based framework for standardized evaluation of implicit preference inference. Experiments reveal a substantial performance gap: while oracle agents with access to ground-truth preferences achieve an 82.92% success rate, the evaluated LLM agents using our framework reach at most 15.62%. Furthermore, we find that correctly inferring user preferences is necessary but not sufficient for task success, as execution failures in downstream web interactions remain a significant bottleneck even when agents align with the target preference. These findings highlight implicit preference inference and robust action grounding as key challenges for deploying reliable, user-facing web agents. We release our code: https://github.com/McGill-NLP/web-agents-test-time-adaptations

Recent studies report that LLM-based multi-agent systems (MAS) fail at rates of 41%-87%, yet to our knowledge, no benchmark to date supports systematic anomaly detection (AD) for them. Building MAS AD benchmarks is hard because they must remain fresh as LLM systems evolve: tasks may leak into training data and thus be memorized by LLMs, traces and anomaly patterns expire as backbones evolve, and labels must be provided reliably for each refresh. To address these challenges, we present MAADBench (MA: multi-agent; AD: anomaly detection), the first refreshable MAS AD benchmark designed for diverse, evolving LLM backbones underlying the agents. MAADBench combines (1) sampled-and-coupled generative tasks over an approximately 10^37-task space to mitigate task leakage, (2) refreshable trace generation under configurable LLM backbones, and (3) automated provision of cost-free, deterministic step-level labels for fine-grained AD evaluation. Beyond offering the paradigm itself, we run MAADBench with five state-of-the-art LLM backbones and release the MAADBench-Full dataset with 5,200 step-labeled traces. Benchmarking 25 AD methods on the MAADBench dataset reveals substantial limitations in current approaches: they rely heavily on supervision, struggle with subtle MAS-specific anomalies, and lack robustness across LLM backbones. These gaps point to a rich research agenda for MAS-specific anomaly detection, with MAADBench providing a systematic and refreshable testbed for method development and evaluation. We open-source MAADBench-Full at https://huggingface.co/datasets/hww123/MAADBench-full.

Large language models (LLMs) enable agents to solve long-horizon tasks by generating a plan and then executing it in an environment. However, successful planning requires two distinct capabilities: selecting an appropriate plan for the task and executing it faithfully. Existing planner--executor systems can fail at either stage, while final task success alone cannot distinguish selection from execution failures. We therefore study the Plan Declaration--Execution Gap and introduce Planning-as-Routing, where an LLM declares one of four planning modes: Predefined, Sequential, Hierarchical, or Search, and a deterministic router dispatches the task to the corresponding pattern-specific executor. Across four benchmarks and three LLMs, we find three consistent patterns. First, generic Plan+ReAct often fails to preserve declared planning structure, especially for longer plans: across three benchmarks, only (22)--(45%) of trajectories preserve it, whereas pattern-specific executors enforce the intended structure. Second, planning-mode effectiveness varies across environments and models: Search performs best on ALFWorld, Hierarchical on SWE-bench, and the strongest pattern can vary across models within the same benchmark. Third, the largest gains come from execution: pattern-specific executors improve task success from (0.48) to (0.92) on ALFWorld and from (0.36) to (0.44) on SWE-bench Verified over Plan+ReAct. Current LLMs, however, do not reliably select the strongest mode for each task, although few-shot examples improve selection in some benchmark--model combinations. Overall, reliable agent planning requires both effective mode selection and faithful execution: routing substantially closes the execution gap, while task-specific mode selection remains open.

Advancing spatial intelligence in Multimodal Large Language Models (MLLMs) is bottlenecked by the scarcity of complex, scalable 3D question-answer (QA) data. While manual annotation is labor-intensive, directly utilizing LLMs to synthesize these QA pairs often fails due to their inherent deficiencies in spatial and geometric computation. We introduce Exemplar2VQA, a scalable exemplar-driven visual question answering generation framework that rapidly synthesizes large-scale spatial QA pairs in simulated environments via multi-agent coding. By equipping collaborative agents with a meticulously designed library of geometric utilities, Exemplar2VQA bypasses LLMs' spatial reasoning flaws through deterministic code execution. Crucially, the framework exhibits remarkable versatility: taking diverse static object-centric spatial query templates as exemplars, it seamlessly and autonomously scales them into massive, high-fidelity synthetic datasets. Fine-tuning Qwen2.5-VL (3B/7B) exclusively on Exemplar2VQA-generated synthetic indoor data yields significant performance improvements across various diverse benchmarks. Furthermore, its effectiveness is not limited to in-domain indoor datasets but also robustly extends to outdoor and mixed-scene benchmarks. These results establish Exemplar2VQA as a scalable and powerful paradigm for bridging the sim-to-real gap in Embodied AI. Our code is at https://github.com/yingjiayu12/Exemplar2VQA

Multimodal GUI agents have achieved impressive results on general software benchmarks, yet their ability to operate professional scientific software remains largely unexplored. In materials science, sparse domain-specific web data, specialized interfaces, and tacit workflow conventions create blind spots that general-purpose pretraining cannot readily bridge. We present MatToolBench, the first real-environment benchmark for evaluating multimodal GUI agents on professional materials science software, comprising 204 tasks across 10 tools in three modalities: GUI operation, OriginPro scripting, and code-based database queries, all executed inside a Windows 11 VM. Each task is decomposed into fine-grained sub-criteria by domain experts, enabling interpretable partial-credit scoring; the GUI component of our multi-level evaluation pipeline achieves an average F1 of 0.98. For OriginPro figure-generation tasks, we further conduct a human-LLM agreement study to validate the use of a multimodal judge for secondary aesthetic assessment. Our experiments show that strong performance on general benchmarks does not transfer to professional scientific workflows, and that this gap is not a visual-grounding problem alone: failures arise from domain-specific operational knowledge, sparse pretraining coverage of scientific software, weak cross-tool artifact handoff, and critical states exposed only visually. Even the best model reaches only 25% success rate on GUI tasks and 45% on code tasks. MatToolBench therefore serves as a challenging diagnostic benchmark and real-environment testbed for data-scarce, knowledge-intensive scientific workflows.

Multi-agent LLM systems are increasingly structured like organizations, with roles, critics and shared memory, yet they are evaluated on tasks that last minutes. We ask how such an organization behaves when the ground it stands on keeps moving. Frontier Autolab is a long-horizon testbed in which one simulated firm, voiced by sixteen role personas and a dedicated Red Team, must re-found itself in nine technology eras from 1990 to 2040. Each era is temporally gated: the firm decides from a dated briefing, a historian-judge then reveals what happened and scores the decision on a five-dimension rubric, and lessons enter a persistent Playbook. Six eras are scored against history, one against the live market and two are open forecasts. Across four trajectories (36 era decisions, 180 subscores) we find a consistent foresight-commitment gap: in all 24 historically scored eras the judge rated the firm's recognition of the coming shift above its choice of where to build (mean gap 1.9 points on a 10-point scale), because boards chose the layer their existing assets could reach. Organizational design shaped long-run character. A Red Team armed with numeric kill gates produced fifty years of gated pilots and no product, and the rubric rated this firm highest; firms whose memory stored market-structure lessons pivoted every era, while a firm whose memory stored only validation procedure kept one method throughout. We also show why such results are hard to trust. Scores rise across eras in every run while the judge's own hindsight subscore falls (within-run r = -0.58), so apparent learning is confounded with recall of history, and we trace further distortions to self-judging, briefing selection and score aggregation. We release all records and an API harness, and specify fictional and post-cutoff eras that would turn the testbed into a benchmark.

Driven by the rapid advancement of large language models (LLMs), LLM-based multi-agent systems (MAS) have emerged as a powerful paradigm for collaborative reasoning over complex tasks. A key design element of MAS is the communication topology, which governs information flow among agents and often encodes proprietary knowledge about the system architecture. However, recent work has shown that such topologies can be inferred even in black-box settings by exploiting semantic dependencies in observable reasoning traces, posing significant risks of intellectual property leakage and exposure of system vulnerabilities. To address this threat, we propose MIRAGE, a topology-concealment framework that preserves the genuine communication topology for task execution while shaping adversary-facing semantic evidence toward a carefully constructed phantom topology. Specifically, MIRAGE operates in three stages: (1) phantom topology synthesis, (2) semantic edge realization, and (3) protected MAS execution. It constructs a phantom topology structurally distinct from the genuine one, materializes phantom edges as plausible semantic dependencies, and suppresses source-specific cues that could reveal genuine edges absent from the phantom topology. Extensive experiments across three topology optimization frameworks and four benchmark datasets demonstrate that MIRAGE substantially reduces the effectiveness of topology inference attacks while largely preserving the task utility of the protected MAS.

Indirect prompt injection makes an LLM agent treat untrusted retrieved text as instructions. We present CounterSteer, an inference-time defense that suppresses this behavior inside the model. Per model, a five-step recipe fits a residual-stream direction from paired episodes differing only in whether an embedded instruction is followed, and retains it only if it passes pre-specified causal and capability gates. At deployment, the direction is subtracted from every tool-result token during prefill. The edit is always on--there is no detection decision to evade--and requires no fine-tuning, auxiliary model, or added tokens, only white-box serving and tool-result span boundaries. Across five open-weights models (8B-106B, five vendor lineages), held-out attack success falls from 0.21-1.00 undefended to 0.00-0.17 defended, and AgentDojo compromise rate from 0.10-0.49 to 0.006-0.079, at 93-100% typography-normalized benign utility, with larger task-dependent costs when reasoning over steered content. A benchmark-level adaptive attacker reaching 0.67-0.73 undefended is held to roughly a quarter of that on the two most deeply evaluated models. Among the defenses we measured on capable models, those achieving lower compromise rates either lost 22-89% of benign utility or fine-tuned the served weights. White-box gradient attacks through the deployed vector compromise at most 2 of 52 episodes, and none of 2,052 replayed human red-team attacks succeeds. CounterSteer largely neutralizes instructional takeover: a black-box framing search cracks 3 of 18 development samples. Parameter manipulation--attacker-chosen arguments in otherwise legitimate calls--is only partially resisted (13 of 18); the decision becomes linearly readable at argument emission but not at the examined pre-generation sites, and is not removed by the tested prefill- or decode-time steering, motivating argument-provenance controls.

Compiler backends are expensive to build and maintain as programming models, workloads, and accelerators evolve. We investigate whether large language models can replace the conventional optimizing and lowering pipeline, a process that we call AI lowering. We study AI lowering from Triton to NVIDIA PTX: an LLM agent translates Triton kernels directly into PTX. We build an environment that evaluates candidate PTX, and an agentic harness in which an LLM translates Triton kernels into PTX. Across twelve common kernels on Ada, Hopper, and Blackwell GPUs and ten kernels from recent ML papers, AI lowering achieves 0.83x-3.34x the performance of autotuned Triton. The largest gains come from transformations that Triton's lowering pipeline does not perform, such as decoding packed binary weights directly into Tensor Core operands (3.34x on BitDelta), assigning each thread a complete softmax row in tensor memory (1.37x on FlashAttention), and reusing overlapping convolution windows (up to 2.23x). These results rely on a robust evaluation harness with comprehensive verification support. We build on Volta, an existing PTX verifier, and substantially extend it to support modern GPU architectures by introducing support for Blackwell's tcgen05 Tensor Core interface. This requires modeling three architectural features: managed tensor memory, descriptor-based operand layouts, and asynchronous execution coordinated through commits, waits, memory barriers, and proxy fences. We discuss the challenges involved in formalizing them, as well as the current limitations. Our results suggest an emerging future in which AI compilers replace custom-written intermediate representations and checkers, reducing the time and engineering effort required to bring up software for new general-purpose and custom chips.

LLM-based multi-agent systems (MAS) increasingly use latent collaboration to avoid the information loss and repeated encoding-decoding overhead of natural-language communication. However, directly forwarding all sender latents makes the receiver-side context scale with both the number of agents and the reasoning length, increasing computation, memory usage, and collaboration latency. A natural solution is latent compression. But we find that cross-agent redundancy remains unresolved in existing latent compression approaches, which typically compress each sender independently and then concatenate the results. We propose LatCom, a cross-agent latent compression framework for efficient multi-agent latent collaboration. LatCom maps multiple sender latents into a fixed number of receiver-readable and task-relevant slots. Rather than reconstructing all sender hidden states, it optimizes the compressed latents for receiver-side task utility. LatCom trains the compressor in two stages: single-sender readability learning first establishes a latent interface interpretable by the frozen receiver, and multi-sender fusion learning then trains the compressor to fuse complementary evidence and remove redundancy across agents. Experiments on multiple benchmarks with Qwen3-4B show that LatCom achieves an average 2.46x inference speed-up over LatentMAS and reduces output token usage by 70.3% while maintaining comparable average accuracy.

No single way of parallelizing attention serves large language models well under all loads. Low concurrency favors tensor parallelism, many independent requests favor data-parallel attention, and long prompts favor context parallelism. Reasoning, agentic, and RL-rollout workloads make a fixed choice untenable: a batch that begins as many short requests ends as a few very long ones, so the best layout changes while the same requests run. Serving engines nevertheless fix one layout at launch, because changing it has meant draining requests and restarting workers. We present SPLASH, a serving system that switches the parallel layout of attention while requests are running. It builds on one observation: modern attention, with few or no KV heads, decouples where a request's KV cache lives from how attention weights are sharded. This has two consequences. First, layouts differ only in who owns the weights and the cache, and most of that state already sits where the next layout needs it; SPLASH reuses it, moves the rest in the background of ongoing inference, and hands off at a batch boundary, making a switch nearly free: its median overhead is under 0.51% of the step it runs in. Second, the decoupling exposes a layout that existing engines lack: Decoupled Ownership Parallelism (DOP) shards attention weights as tensor parallelism does while keeping each request's cache on a single owner as data-parallel attention does. DOP replicates neither, offers 27-60% more KV capacity than data-parallel attention, and gives the scheduler a choice when KV memory limits admission. A transition-aware scheduler follows the best of the four layouts as load changes. On B200 GPUs serving GLM-5.3, SPLASH improves end-to-end serving throughput by 1.3-1.73x over fixed-layout deployments, and the same layout regimes appear with DeepSeek-V3.2 on H200 and GLM-5.3-Flash on DCU.

Agentic sessions driven by Large language models (LLMs) often alternate between model inference and tool use, accumulating long histories across successive rounds. Serving these sessions efficiently requires reducing attention computation and retaining history key-value (KV) caches to avoid recomputation. Recently, frontier open-source LLMs adopt sparse attention to reduce computation by selecting only part of the history, while SSDs provide a cheaper alternative to CPU DRAM for storing KV caches. However, sparse KV selection depends on the ad hoc intermediate values during model inference, so it forces SSD reads to lie on the inference critical path. These reads are further slowed by fragmented accesses and read-write interference in SSDs. To address these challenges, we present Janus, an agentic serving framework for sparse attention LLMs with SSD-centric KV storage. Janus focuses on append prefill, which processes each round's newly added inputs and accounts for most history KV loading. To move SSD reads out of the critical path, Janus runs the model's own KV selection module on earlier intermediate values, predicting KV demand without additional training. The predicted reads overlap with model computation, and any prediction misses are fetched before attention executes to preserve model outputs. To improve SSD efficiency, Janus coalesces adjacent reads, packs scattered KV pages into sequential writes on the CPU, and limits background writes while reads are active. Across three models and three agentic traces, Janus outperforms existing works by up to 1.57-3.69 times (1.22-1.85 times on average) in terms of the time to first token latency, while maintaining decode efficiency.

Large language model (LLM) agents repeatedly load reusable content, such as skills, documents, and memory entries, into the current context. Re-encoding this content for every request wastes computation. Position-independent caching (PIC) alleviates this by encoding each artifact independently and reusing its key-value (KV) states at arbitrary positions, but it incurs a quality loss relative to full-context prefill. Existing methods repair this loss by restoring global position IDs or recomputing selected tokens. In this work, we isolate the source of the loss, finding that the positional mismatch has minor effect, and independently cached artifacts retain faithful representations: reading a provided artifact stays largely accurate, and performance degrades only when the model must select among multiple artifacts. Moreover, replacing PIC's attention scores with full-prefill scores recovers performance with the cached KV unchanged, localizing the failure to the attention rather than KV recomputation. Motivated by this, we propose \textsc{Attuner}, a query-side adaptation method that learns to read a frozen artifact cache. \textsc{Attuner} inserts low-rank adapters into the query projections and is trained by distilling full-prefill distribution into the student. It trains fewer than 0.05% of the model parameters and, at inference, requires neither cache recomputation nor a full-context reference. On Qwen3-4B and Qwen3-8B across seven benchmarks covering skills, documents, memory, and code, \textsc{Attuner} substantially outperforms prior PIC baselines in both in-domain and out-of-domain settings, matches full-context prefill quality while providing up to $3.73\times$ speedup.

Users of language-based assistive agents often make ambiguous requests. In response, an assistant can either directly act on its interpretation of the request --- risking misalignment with the user --- or ask a clarifying question. Which option is the most safe and helpful? A common approach is to ask questions that minimize uncertainty about the user's intent until a threshold is reached. However, this neglects the impact of uncertainty reduction on downstream performance, the costs of asking versus acting immediately, and the possibility that users may provide corrections without being asked. To navigate these trade-offs, we introduce Rational Enquiry via Value-of-Information Reasoning (REVOIR). REVOIR makes clarification decisions via inference-time reasoning about the value-of-information of a question, which captures the expected improvement in task reward due to the answer received. In two assistive tasks --- ambiguous question answering (CondAmbigQA) and preference-aligned household task planning (ADAPT) --- we show that REVOIR achieves greater success with fewer questions than approaches based on prompting, chain-of-thought, fine-tuning, or information gain, improving preference satisfaction on ADAPT by 13-15% over a fine-tuned clarification policy while requiring no training and asking five times fewer questions. Furthermore, when the assistant can receive cheap user corrections after acting, REVOIR naturally infers that asking questions is not always efficient, demonstrating the adaptivity of our approach. In contrast, we find that vanilla reasoning agents fail to adaptively clarify user requests, and request fewer clarifications as reasoning effort increases.

Skill evolution improves the capabilities of large language models by analyzing trajectories generated under a given skill and modifying the skill accordingly. Existing approaches typically generate a single trajectory per question. However, this provides insufficient optimization signals since it requires inferring effective skill edits from a solitary path. It is difficult to pinpoint which actions caused the failure in a failed trajectory, or to determine which actions in a successful one should be incorporated into the skill. Furthermore, they rely on a local batch of trajectories for analysis, making the optimization direction susceptible to noisy evidence. To address these, we propose SkillCome, a Skill-evolution method based on group Contrast optimization with dual memory. For each question, SkillCome generates trajectories and performs group contrast analysis to precisely identify key behavioral divergences between successful and failed trajectories, offering reliable optimization signals. The dual memory system further accumulates evidence from historical steps to track patterns shared across different groups, leading to more generalized optimization directions. Together, SkillCome builds a systematic optimization process that transforms experience from observed successful trajectories into reusable skills. Extensive experiments on six benchmarks spanning question answering, reasoning, and agentic tasks demonstrate the effectiveness of our method. SkillCome consistently outperforms baselines across five models of varying families and scales, with gains up to +5.69 points.

Large language models (LLMs) are increasingly deployed as social agents, yet credible human-like interaction requires more than fluent responses or persona consistency. Agents must autonomously decide whether, when, and how to communicate while adapting to evolving contexts, goals, and relationships. Existing research, however, lacks a unified approach to enabling, evaluating, and improving such capabilities in continuous, open-ended interaction. We introduce AnthroDial, a unified framework for developing anthropomorphic social agents from three complementary aspects: MindFlow, a lightweight interaction harness that enables autonomous, asynchronous, and adaptive communication through a dynamic Mind Buffer; CAPS-Eval, a theory-grounded framework for evaluating cognitive, affective, and behavioral dimensions of anthropomorphic interaction; and a scalable training paradigm that combines SEEDS for environment expansion with DiAPO for adaptive capability optimization. We further construct evaluation datasets covering everyday communication, game interaction, and long-horizon character interaction. Extensive experiments across diverse models and scenarios demonstrate improved interaction autonomy and naturalness, validate the reliability, discriminativeness, and agreement with human rankings of CAPS-Eval, and confirm the effectiveness of our training paradigm. Together, these components provide a unified framework for developing credible human-like social agents in open-ended interaction.

Incorporating rich task-relevant context, such as domain knowledge and external observations, is a key capability yet remains challenging for Bayesian optimisation (BO). Recently, practitioners have started to use large language models (LLMs) to generate and execute BO programs through coding harnesses. In such emerging practices, the posterior belief is shaped not only by Bayesian inference but also by LLM-generated model and data artefacts, offering a flexible route for task context to enter BO as executable code. To study whether and how LLMs can be harnessed to compile diverse contextual signals for BO, we formulate LLM-compiled BO as generalised-context decision making. We propose HarBO, a BO-specialised harness that compiles generalised context into the core artefacts of standard BO through a validated multi-stage workflow. Our theory analyses the regret under imperfect compilation and the effect of adding new context. Across synthetic functions and real-world benchmarks, we find that LLM harnesses can effectively compile context into standard BO, achieving competitive performance with specialised LLM-embedding-based and direct LLM-in-the-loop BO methods. General coding harnesses can be effective in familiar domains such as hyperparameter optimisation, but fall short in unfamiliar, context-rich domains. Together, these results establish LLM harnesses as a promising, but not automatically reliable, route for making rich task context usable in BO.

LLM-based multi-agent systems (MAS) have shown promise in complex problem solving. As MAS methods diversify, systematic evaluation becomes increasingly challenging. However, existing benchmarks largely focus on final outcomes, leaving unclear how collaboration gains arise, are preserved, or are lost. To address this limitation, we introduce MASTraceBench, a benchmark for diagnosing collaboration gains through proposal trajectories in MAS. Across six cooperative and competitive tasks, MASTraceBench tracks and grades proposal trajectories and provides a multi-layer metric suite covering Task Score, Collaboration Gain, proposal-trajectory indicators, and Token Cost. Using MASTraceBench, we systematically compare representative MAS methods not only by final performance, but also by how agent proposals evolve and are aggregated into the final answer. This analysis reveals a recurring pattern: final MAS answers rarely surpass the strongest initial proposal; interaction often lifts initially weaker proposals toward it, while strong initial proposals are seldom further improved and may regress. To reduce this risk, we propose CLEARS, which replaces whole-proposal exchange with claim-level evaluation across agents to guide reliable synthesis. CLEARS more often preserves or improves upon the strongest initial proposal and achieves the highest Collaboration Gain on five of the six tasks.

Long-horizon competition tests agents' ability to coordinate business decisions under uncertainty and adapt to changing rival strategies. We introduce CEO Arena, a benchmark that uses matched replacement evaluation to assess operating returns alongside an agent's effects on rivals and the market. Each CEO agent is compared with a reference policy in the same company under the same economic seed, holding other agents' identities and assignments fixed while all agents adapt. In a shared eight-company market spanning 500 simulated days, CEOs make sequential decisions on pricing, procurement, marketing, research and development, and service using private company information and noisy market signals, under resource constraints and delayed feedback. We evaluate eight LLM-based CEO agents in 27 main runs and 26 robustness runs. In the main evaluation, most agents have negative mean returns, and private gains can accompany market losses. Robustness analyses suggest that aggregate patterns extend beyond the original rule-based baseline; four of the 56 directed pairs show relatively stable effects. Memory, action, and accounting traces suggest demand capture and rivals' pricing and spending responses as possible explanations. CEO Arena provides a controlled testbed for studying long-horizon agent competition, strategic interaction, and market externalities.

Multimodal large language models (MLLMs) are increasingly envisioned as core components of scientific and engineering agents, yet their ability to interpret physical fields remains poorly understood. Existing physics benchmarks largely emphasize textbook problem solving or intuitive physical reasoning, leaving open whether MLLMs can infer physically meaningful information from continuous field observations. We introduce PhysFieldBench, a benchmark comprising 24 tasks and 1,160 evaluation examples across controlled equation fields, simulated physical fields, and observed physical fields. The tasks assess three forms of inference: identifying physical mechanisms, comparing latent control variables, and predicting outcome properties. Across representative open-source and proprietary MLLMs, zero-shot performance is low: the best model achieves a chance-normalized score of 29.3, while several open-source models remain near chance. In contrast, a task-specific supervised vision transformer performs substantially better, demonstrating that the inputs contain learnable physical information. To diagnose these failures, a structured self-explanation analysis attributes most errors to missed visual patterns and incorrect visual-to-physical mappings. Further, to explore whether post-training can improve physical inference and generalize to unseen tasks, we compare supervised fine-tuning with final answers or chain-of-thought supervision and reinforcement learning. Final-answer supervision performs best overall but transfers less effectively, whereas reinforcement learning after chain-of-thought supervision achieves the best generalization. Together, these findings highlight the need to improve visual-to-physical grounding and cross-task generalization for MLLMs to reliably interpret physical fields in scientific and engineering workflows.

Large language models (LLMs) are increasingly used to compute clinical risk scores from free-text notes. Notes are often incomplete, and treating undocumented findings as normal can silently misclassify patients. We test whether separating three-state extraction (present, absent or unknown, by an LLM) from decision logic (deterministic code computing score bounds over unknown inputs) lets a system ask only questions that can change the decision. On 1,200 synthetic emergency cases across six calculators (HEART, CURB-65, qSOFA, PERC, Wells, Cockcroft-Gault), with a simulated clinician answering questions, we compared this bounds policy with asking for every missing input, a missing-equals-normal schema, and an end-to-end LLM agent (Claude Opus 5.5). With Claude Haiku 4.5 as extractor, the bounds policy matched ask-all accuracy (99.4% vs 99.4%) with half the questions (0.92 vs 1.78 per case) and no irrelevant ones. Treating missing as normal dropped accuracy to 91.2% and under-triaged 8.5% of patients (95% CI 7.1-10.2), and under-triage persisted under messy notes and a noisy clinician. The agent was equally accurate under ideal conditions (99.6%) but 9.5% of its questions were irrelevant; with a noisy clinician it was less accurate than the bounds policy (83.5% vs 87.0%, p<0.001) and committed prematurely in 2.7% of cases (bounds: 0%). A 9B local model as extractor reached oracle-level accuracy (99.8%). In 584 real case reports from MedCalc-Bench, only 52% contained enough information to determine the category (HEART 13%). Routing decisions through code that reasons explicitly about unknowns avoids premature commitment and irrelevant questions, halves the questions asked, and works with small local models.

As LLM agents increasingly collaborate on complex tasks, how to organize their interactions becomes a central design question. Existing multi-agent systems typically learn or adapt explicit roles, hierarchies, routing policies, or communication topologies. We shift the learning target to a reusable local law that can be shared across interchangeable agents and adapt coordination as populations or interaction conditions change, without redefining a global organization. We introduce Waggle, a shared anonymous policy over bounded local views that jointly selects task actions, semantic communication, and local commitment updates. Repeated execution of the same law allows coordination to form, persist, and reorganize online without explicit roles or global topology. To learn this law across interchangeable agents and evolving coordination, we develop Swarm-Consistent Distillation (SCD), combining anonymous-orbit consistency with rollout-grounded prediction of the next local coordination field, with no added inference-time components. Across diverse coordination settings, the same learned law remains effective as populations and interaction budgets change, retains over 96% of substrate-specific oracle quality, and transfers without retraining; SCD further improves reorganization after counterevidence. Together, these results show that LLM-agent organization can emerge and adapt through repeated execution of a learned local law.