每天知道
你的领域
出了什么

Language model agents are becoming proficient executors at isolated, short-horizon tasks such as software engineering and customer service. Yet real-world challenges require a combination of sophisticated skills that remain largely untested in agents: (1) navigating long horizons amid uncertainty; (2) acquiring information in noisy environments; (3) adapting to a changing world; (4) orchestrating multiple moving parts toward a coherent goal. We introduce CEO-Bench, which evaluates these capabilities together by simulating a representative real-world task: operating a startup for 500 days. An agent manages pricing, marketing, budgeting, and many other aspects of a fictional company through a programmable Python interface, operating in the same environment and facing the same challenges as a human CEO. Success demands analyzing noisy, interconnected business databases, translating signals into sound strategy, and coordinating many decisions with programming. The strongest agents write sophisticated code that simulates customer cohorts to forecast future cash and mines negotiation history to uncover hidden customer preferences. Even so, most state-of-the-art models struggle in this environment. Only Claude Opus 4.8 and GPT-5.5 finish above the $1M starting balance, and neither consistently turns a profit. CEO-Bench takes a first step toward measuring the intelligence required to drive sustained, adaptive progress over time.

Tool-using language-model agents introduce security failures that go beyond unsafe text: they can disclose protected objects, write persistent memory, send messages, modify databases, or trigger harmful code and tool effects. Existing evaluations often collapse these stages into a single attack success rate, making it difficult to tell whether a model merely agreed with an attacker or actually produced observable harm. We introduce SafeClawBench, a staged benchmark for tool-using agent security with 600 controlled adversarial tasks across six attack families: direct and indirect prompt injection, tool-return injection, memory poisoning, memory extraction, and ambiguity-driven unsafe inference. SafeClawBench reports three separate endpoints: semantic attack acceptance, audit-visible harm evidence, and sandbox-observed tool/state harm. Evaluating five agent endpoints under four prompt-level policies, we find that these endpoints capture different failure modes. Without additional prompt protection, semantic failure rates vary widely across models, from 9.0% to 44.2%. Audited harm evidence is narrower than semantic failure, and under a separate executable protocol some matched task identities produce sandbox harm despite passing the Semantic Core call: in a 12,000-row matched analysis, 291 of 347 observed sandbox harms occur in rows that pass the semantic check. Prompt policies change endpoint outcomes, but their effects depend on both model and protocol. SafeClawBench provides a reproducible framework for comparing agent models and prompt-policy conditions without conflating textual compliance, evidence-supported harm, and executable state changes. The open-source dataset is available at https://huggingface.co/datasets/sairights/safeclawbench.

Enterprise intrusion response still depends on static playbooks and analyst-driven triage, creating delay between alert generation and containment. We present Agentra, a supervisable multi-agent Intrusion Response System (IRS) framework that converts alerts from IDS, EDR, and XDR platforms into structured incident response plans grounded in MITRE ATT&CK, MITRE D3FEND, and NIST CSF 2.0. Agentra decomposes response reasoning across role-scoped agents, validates proposed plans through a bounded Planner--Validator review loop, screens retrieved threat intelligence through a Moderator security gateway, gates actions through an Action Catalog and risk score, and records decisions in an append-only audit log. We evaluate Agentra against a static OASIS CACAO v2.0 cyber-playbook baseline on a 120-event corpus drawn from ThreatHunter-Playbook, Splunk BOTSv3, and DARPA OpTC. The strongest configuration improves FP-aware IRS F1 from 0.61 to 0.84 and restores the projected harmful-action rate to the static baseline level of 0.0% after Planner-only configurations introduce unsafe overreaction. These results indicate that multi-agent response planning can improve ontology-grounded IRS coverage while preserving analyst approval and auditability.

AI agents are moving from advisors to actors, booking travel, planning menus, and running procurement on behalf of users. Existing benchmarks for AI and animal welfare evaluate model text responses to question-answer prompts, leaving open whether the welfare reasoning surfaced in those responses transfers to agentic deployment where the model must take actions with tools. We introduce TAC (Travel Agent Compassion), the first agentic benchmark measuring whether AI agents avoid options involving animal exploitation when acting on behalf of users. TAC presents an AI agent with twelve hand-authored travel booking scenarios across six categories of animal exploitation, augmented to forty-eight samples to control for price, rating, and position confounds. We evaluate seven frontier models from four labs. Every model scores below the chance level of sixty-four percent, with the best performer (Claude Opus 4.7) at fifty-three percent. A single welfare-aware sentence in the system prompt yields gains of forty-seven to sixty-three percentage points in Claude and GPT-5.5, twenty-six points in GPT-5.2, and under twelve points in DeepSeek and Gemini. An auxiliary Inspect Scout audit of 288 base-condition transcripts from the top two performers, using Gemini 2.5 Flash Lite as judge, flags zero transcripts for evaluation awareness, suggesting the below-chance rates do not stem from the models recognising the evaluation. We discuss implications for category-level variation across cultural domains, the limits of text-response welfare benchmarks, and the EU General-Purpose AI Code of Practice systemic risk framework.

Computer-use agents (CUAs) approach human-level performance on standardised benchmarks but still struggle on long-horizon tasks and unseen software. Existing skill libraries address this with reusable skills, but represent the skill artifact as text only, despite the visual nature of GUI interaction. We propose VISUALSKILL: a hierarchical multimodal skill, tailored to each target application and organised as a central index over per-topic files, which the agent consumes through a load_topic MCP tool that fetches the relevant topic's text and figures on demand. We construct each skill with a two-stage pipeline that combines authored documentation with live-application UI exploration. On two CUA benchmarks, CUA-World and OSExpert-Eval, a Claude Code CLI agent backed by Claude Opus 4.6 reaches an average score of 0.456 with VISUALSKILL, a +15.3 point absolute lift over the no-skill baseline (0.303). Against a matched text-only skill that is generated from the same source content and differs from VISUALSKILL only in modality, VISUALSKILL yields a further +8.3 point absolute gain over the matched text-only skill (0.373 vs. 0.456), providing direct evidence that retaining visual figures in the skill artifact, rather than verbalizing them away, helps the agent both identify UI elements and verify workflow state after each action. Our code is available at https://github.com/XMHZZ2018/VisualSkills.

Personalized dialogue agents require continuous long-term memory to maintain coherent interactions across multiple sessions. However, deploying these capabilities on consumer-grade hardware (e.g., 8 GB VRAM edge devices) introduces severe memory and compute bottlenecks. Existing systems typically rely on isotropic cosine similarity for retrieval and heuristic rules for context compression. These approaches lack a unified theoretical foundation, frequently suffering from the hubness problem in high-dimensional retrieval and syntactic fragmentation during compression. To overcome these limitations, we propose CoreMem, a resource-efficient edge-cloud memory architecture fundamentally unified by information geometry. First, Riemannian retrieval replaces cosine matching with a locally adaptive Fisher-Rao metric, effectively penalizing hub memories via Mahalanobis distance with O(Ndr) Woodbury acceleration for real-time search. Second, Fisher-guided discrete token distillation (FDTD) introduces a hierarchical sentence-to-token compression mechanism. It derives sensitivity scores from Fisher information traces, providing a principled compression-KL tradeoff augmented with explicit structural syntax protection. Evaluated on the LOCOMO and LongMemEval-S benchmarks, CoreMem achieves strong accuracy improvements, yielding substantial gains in Open-domain (+4.51 pp) and Temporal (+4.17 pp) reasoning. Extensive profiling confirms that CoreMem operates seamlessly within a strict 8 GB VRAM budget, successfully bridging the gap between resource-constrained edge devices and the demand for theoretically grounded, lifelong memory agents.

The growing deployment of artificial intelligence (AI) necessitates robust deep learning (DL) compilers, such as TVM and ONNX-MLIR. These compilers take as input high-level AI models, lower them through multi-layer transformations, and specialize them to diverse hardware. Testing such compilers is uniquely challenging as correctness depends on implicit constraints embedded throughout the compilation stack. Existing testing approaches largely take type constraints to restrict input model generation and therefore emphasize type validation and monitor compilation crashes or coverage gains. This focus overlooks compiler-platform interaction bugs that arise from interleaved effects across compilation and execution environments. In this work, we propose a scalable, automated DL compiler testing framework for, in tandem, (1) finding compiler-platform interaction bugs and (2) enabling behavior equivalence partitioning. Our key insight is that these bugs are caused by violated assumptions arising from interactions across compilation passes and hardware platforms. Therefore, we move beyond constraining input generation and derive full-stack constraints. Our approach is three-fold. First, we design an automated approach to extract full-stack constraints that jointly guide model generation and characterize compilation behaviors. Second, we prioritize constraints that expose interaction-sensitive behaviors, so our generated models are capable of exercising deep compilation logic. Third, we enable behavior equivalence partitioning by automatically inserting assertions to monitor distinct compilation symptoms that coverage or pass/fail signals miss. We evaluated our tool, XCheck, on three widely-used DL compilers and found 2,034 bug-revealing cases, including memory overflows, integer overflows, and silent unexpected compilations that were rooted in compiler-platform interactions.

LLM serving exhibits extreme length variability, making size-based scheduling difficult in practice. Recent LLM schedulers approximate SJF/SRPT using predicted decode lengths or ranks and primarily report mean-centric metrics such as TTFT and TBT. We show that these prediction-driven policies can be fragile under distribution shifts, bursty arrivals, and GPU memory pressure, while offering limited control over the tail latency (P90-P99) that dominates user experience, even with perfect decode-length knowledge. We introduce a distribution-aware, prediction-free scheduling framework that replaces explicit length prediction with soft priority boosting driven by lightweight statistical signals. Our design co-optimizes scheduling and cache-aware preemption to account for memory-coupled decode dynamics across workload mixes. Evaluated on production and open-source traces, our method reduces P99 TTLT by up to 35-50% relative to SRPT with perfect length knowledge and reduces TTFT by 34-47% across workloads, including reasoning-heavy and chat-heavy tasks. These results demonstrate a robust alternative for optimizing tail latency in online LLM serving.

Deploying Large Language Models (LLMs) requires exploring a large configuration space spanning parallelization strategies, batching techniques, and scheduling policies. Exhaustive measurement across this space is impractical, making latency prediction essential for system optimization. While NPUs have emerged as accelerators designed for LLM inference, no prediction methodology has been established for them. Specifically, applying prior work to LLM inference latency prediction on NPUs faces three challenges: undisclosed microarchitecture of commercial NPUs, unpredictable compiler optimizations, and latency non-linearity induced by bucketing. We present LENS, a latency estimator that predicts NPU inference latency without information on the microarchitecture or compiler, and captures the non-linear latency induced by bucketing. LENS profiles each bucket with two end-to-end (E2E) measurements and composes the results to predict latency for arbitrary input-output length combinations. We validate LENS across NPUs from multiple vendors, several LLMs, and diverse workloads, achieving a mean prediction error of 2.15%. We further compare LENS against two methodologically related baselines, confirming the validity of its approach.

Speculative decoding (SD) accelerates autoregressive Large Language Models (LLMs) by drafting multiple tokens and verifying them in parallel, but it faces a scaling limitation: increasing the draft budget improves speed only when acceptance remains high and drafting overhead stays low. This ceiling has been difficult to break because prior head-based SD methods face a causality-efficiency dilemma. Autoregressive drafters produce path-conditioned candidates that are effective for tree speculative decoding with higher acceptance length, but their drafting cost grows with tree depth. Bidirectional block-diffusion drafters generate all positions in one pass, but their branch-agnostic marginals can form individually plausible yet mutually inconsistent trees, wasting budget and reducing acceptance. We propose JetFlow, a head-based SD framework that combines one-forward drafting efficiency with branch-wise causal conditioning. JetFlow trains a causal parallel draft head over fused hidden states from the frozen target model, producing candidate trees whose scores align with the target model's autoregressive factorization. This enables JetFlow to convert larger draft budgets into longer accepted prefixes and higher end-to-end speedup. Across math, coding, and chat benchmarks on dense and MoE Qwen3 models, JetFlow consistently outperforms bidirectional-head and tree-based SD baselines. On H100 GPUs, JetFlow achieves up to 9.64x speedup on MATH-500 and 4.58x on open-ended conversational workloads, with further latency gains demonstrated through vLLM integration under realistic serving loads. Our code and models are available at https://github.com/hao-ai-lab/JetFlow.

We introduce VEGA, an approach for training navigation VisionLanguage-Action (VLA) models from unlabeled egocentric navigation videos. Internet-scale egocentric videos provide a scalable source of navigation-relevant visual observations, capturing cluttered scenes, close-range obstacles, and natural human motion through real-world spaces. However, these videos are not directly usable for policy learning because they do not provide obstacle-aware trajectories conditioned on explicit navigation goals in the robot's coordinate frame. VEGA addresses this gap by reconstructing local scene geometry from monocular video, sampling navigation goals (represented as text, image, or spatial waypoints) and generating obstacle-aware trajectories using the constructed geometry. The resulting trajectory distribution is then used to train a flow-matching VLA navigation policy. By using geometry exclusively during training, VEGA distills obstacle-aware planning directly into a vision-based policy. Furthermore, we introduce VEGA-Bench, a benchmark containing 250k scenes and approximately 5 million navigation goals paired with scene geometry, designed to evaluate goal progress, collision avoidance, and obstacle clearance of VLAs. Our evaluation shows that VEGA achieves competitive goal progress while reducing collisions by 33.0% and improving obstacle clearance by 17.9% over the strongest baseline on VEGABench, while improving success by at least 150.0%, reducing collisions by at least 66.7%, and improving obstacle clearance by at least 60.0% in real-world trials. Ultimately, we demonstrate that video-derived geometric supervision provides a scalable and effective signal for training obstacle-aware navigation VLAs. The code and benchmark will be released at the time of publication.

Recent progress has shown promise in distilling multi-step video diffusion models into efficient few-step students. Among them, Distribution Matching Distillation (DMD) and its successor DMD2 achieved strong generation quality and fast convergence. However, due to the nature of the reverse Kullback--Leibler (KL) objective, these methods exhibit two persistent failure modes: a substantial drop in sample diversity, and visibly over-saturated outputs that deviate from real-video appearance. In this work, we propose Data-Forcing Distillation (DFD), a simple post-training framework that restores diversity and fidelity in DMD with only a single-line of code change. At its core is the teacher score discrepancy to guide the student toward the real-data distribution, pulling it to missing modes (mitigating mode collapse) and away from problematic modes absent in real data (avoiding over-saturation). We provide an in-depth theoretical analysis of our framework and validate our approach on text-to-video, image-to-video, and autoregressive video generation. With only 100--300 steps of finetuning, DFD effectively restores diversity and fidelity on both Wan2.1-1.3B and Cosmos-Predict2.5-2B model, resolving the over-saturation artifacts with significantly better video dynamics and appearance, and even outperforms the teacher model.

This paper develops a variational framework for regulated language generation. Starting from autoregressive token sampling, we derive the induced distribution over complete messages and relate it to an entropy-regularized Gibbs law. Regulation is modeled as an optimal discriminator whose convex-dual value is an f-divergence, and the generator-regulator interaction is formulated as a saddle-point problem. The framework applies to moderation, censorship, AI deception detection, compliance auditing, phishing defense, and manipulation control, where regulation concerns a distribution over possible messages rather than a single output. The equilibrium clarifies the tradeoff among utility, entropy, regulatory alignment, and finite-length detectability. Two finite-vocabulary case studies, censorship filtering and phishing defense, illustrate how the theory can be evaluated through utility, entropy, divergence, receiver-side scores, and detection probability.

LLM agents increasingly rely on external skills — reusable tool specifications — but real-world tasks often require composing multiple skills, not just selecting one. We formalize this as the Compositional Skill Routing problem: given a complex user query and a large skill library, decompose the query into atomic sub-tasks, retrieve the appropriate skill for each sub-task, and compose an executable plan. We present SkillWeaver, a decompose-retrieve-compose framework combining an LLM task decomposer, a bi-encoder skill retriever with FAISS indexing, and a dependency-aware DAG planner. To support evaluation, we introduce CompSkillBench, a benchmark of 300 compositional queries over 2,209 real MCP server skills spanning 24 functional categories, sourced from the public MCP ecosystem. Our experiments reveal that task decomposition quality is the primary bottleneck: standard LLM decomposition reaches only 34.2% category recall at the step level. To address this, we propose Iterative Skill-Aware Decomposition (SAD), a retrieval-augmented feedback loop that iteratively aligns decomposition with available skills. SAD improves decomposition accuracy from 51.0% to 67.7% (+32.7%, Wilcoxon p < 10^-6) in a single iteration; DA-conditioned analysis confirms that correct granularity is the prerequisite for effective retrieval (CatR@1 rises from 34% to 41% when DA=1). SkillWeaver reduces context window consumption by over 99%, and transfer experiments confirm generalization (+35.6% relative DA gain even when target categories are absent from the retrieval pool).

Fine-grained action recognition in egocentric video is challenging for Vision-Language Models (VLMs): actions often differ only in small visual cues, and a single model tends to be biased toward a subset of these cues. We propose Divide, Deliberate, Decide, a fully-local, zero-shot multi-agent framework in which (i) a VLM orchestrator chunks the video and proposes a top-k candidate label list per segment, (ii) an ensemble of heterogeneous VLM specialists, drawn from different open model families, engages in a structured deliberation that includes a peer-consultation round of questions, and (iii) agent rankings are aggregated with a Borda count and the orchestrator re-ranks its own prediction in light of the specialists' evidence. The entire pipeline runs locally with no fine-tuning. Experiments show that our method positively improves zero-shot action recognition performance over the baseline, highlighting the influence of a heterogeneous deliberation step, showing that the gain stems from decorrelated model priors rather than from additional compute.

LLM agents in interactive environments are highly sensitive to their prompts, yet prompt engineering remains a manual, task-specific process. We introduce an automated prompt optimization framework for LLM agents that decomposes the observation-to-action pipeline into a goal-conditioned descriptor agent and an action selection agent, and iteratively refines each module's prompt through an LLM-driven evolutionary loop guided by environment returns. We propose a behavior analyzer to attribute episode outcomes to specific prompt components, and a mutator to propose targeted revisions to the prompt, before validating them through environment rollouts. We evaluate on all five BabyAI tasks in the BALROG benchmark, comparing our pipeline against BALROG's RobustCoTAgent under both plain and guided prompt initializations. Optimization improves performance consistently across tasks and conditions, without requiring updates to the model weights. On PutNext, a multi-step coordination task where the RobustCoTAgent achieves 0% success, our framework reaches up to 72.5% success rate using the same underlying LLM with optimized prompts. These results suggest that a multi-agent framework, combined with automatic prompt optimization, enhances LLMs without the need for fine-tuning or extensive human supervision.

The LLM-empowered personal health agents with user health (sensor) metrics have offered a promising pathway to alleviate global disparities in healthcare access. However, large-scale clinical deployment remains constrained by an open-ended evaluation bottleneck: physician annotation is reliable but costly and unscalable, while LLM-as-a-judge evaluators are scalable but subjective, inconsistent, and sometimes clinically misaligned. We introduce RubricsTree, a scalable evaluation framework with an expert-aligned hierarchical taxonomy of over 100 atomic, clinically-verifiable Boolean rubrics, evolving from the insights of 4,000 real user queries through an iterative human-in-the-loop curation protocol with an expertise panel led by an experienced physician. A context-aware adaptive router activates only the relevant auto-weighted rubric subset per query, providing the throughput needed for scalable evaluation with expert-aligned quality. Through a systematic meta-evaluation, we show that RubricsTree (i) substantially exceeds a strong large-scale evaluation baseline in expert alignment on challenging open-ended queries; (ii) reliably penalizes contextually degraded responses; and (iii) when used as structured instructions, text feedback, or training rewards for performance optimization, yields up to ~66% relative gains on HealthBench for Gemini, GPT, and Qwen model families. RubricsTree thus provides a scalable, auditable, and evolving evaluation infrastructure required for the continuous optimization of product-level personal healthcare AI.

Self-evolving LLM-based agents improve mainly by changing their agent harness: the structured execution layer around a base model, including prompts, memory, tools, middleware, runtime state, and the model-tool interaction loop. Existing evaluations often reduce this process to isolated task scores or a single sequential curve, obscuring whether an update produces reusable improvement, overfits recent tasks, increases cost, or harms older behavior. We introduce SEAGym, an evaluation environment for measuring agent harness updates across training, validation, test, replay, and cost records. SEAGym turns Harbor-compatible benchmarks into dynamic self-evolution task sources with train batches, frozen update-validation, held-out ID and OOD transfer views, replay diagnostics, and saved snapshot and metric records. Instantiating SEAGym on Terminal-Bench 2.0 and HLE, we compare ACE, TF-GRPO, and AHE under a shared epoch/batch protocol. The results show that these evaluation views provide complementary signals about the evolution process: frequent updates may fail to improve held-out performance, useful intermediate snapshots may collapse later, and source diversity and model backend can affect harness reliability.

As LLM-based shopping agents enter production, existing benchmarks fail to capture how a shopper's requirements arrive: stated implicitly in the query, recorded in a profile, or revealed only when the right question is asked. Benchmarks that expose full intent upfront and grade only the final choice can neither pose this long-horizon challenge nor explain which requirement an agent missed. To address this gap, we introduce EComAgentBench, a benchmark of 662 tasks grounded in real Amazon products and reviews. Each task scatters these requirements across a visible query, a tool-gated profile, and scripted clarification; an agent must uncover hidden intent, verify candidates against attributes and review evidence, and commit to a single product within 100 tool calls. Moreover, typed, source-tagged rubrics grade every task, attributing each failure to a requirement and its source. Construction is automated yet reliable, with every answer fixed in code before any text is generated and every sample validated. Our evaluation of seven models reveals that even the strongest attains only 57.1% overall accuracy, and rubric satisfaction degrades from visible to hidden sources. Overall, we believe EComAgentBench will serve as a reproducible foundation for moving shopping agents from single-query search toward dependable assistance over long horizons.

Game generation is an emerging application of coding agents, requiring models to transform natural-language specifications into playable interactive systems. Unlike traditional coding tasks, game generation takes place within a game engine, where scripts, scenes, assets, rendering, and runtime interactions must jointly produce coherent gameplay. We formalize end-to-end game generation as the problem of producing a complete game artifact that realizes a specification through observable player-game interaction in a target environment. We argue that evaluating this setting requires three desiderata: Engine Grounding, Artifact Completeness, and Interactive Verification. We propose an interaction-grounded evaluation framework that assesses executable gameplay through replayed demonstrations and rubric-guided multimodal judging. We instantiate this framework as GameCraft-Bench, a benchmark comprising 140 Godot tasks across 15 game families. Evaluations of frontier coding agents show that end-to-end game generation remains highly challenging: the strongest agent achieves only 41.46%, and most agents score below 40%. Further analysis reveals that while agents often implement recognizable mechanics, they struggle to deliver complete games with sufficient content, functional visual feedback, and coherent presentation. See https://tongxuluo.github.io/gamecraft-bench-website for demos, code, and data.

Computer-using agents drive real software through the screen — clicking and typing — but they solve every task from scratch: asked to repeat a task, an agent re-reads the screen, re-reasons every tap, and pays the full cost again. We present PreAct, which lets such an agent get faster on tasks it has done before. The first time it succeeds, PreAct compiles the run into a small state-machine program-states that check the screen, transitions that act-and on later runs replays it directly instead of invoking the agent 8.5-13x faster, with no per-step language-model calls. Replay is not blind: at each step PreAct checks that the screen matches what the program expects before acting, and hands control back to the agent the moment something is off. PreAct applies the same discipline when deciding what to keep: a freshly compiled program enters the store only if, re-run from a clean state, an independent evaluator confirms it solved the task-catching programs that replay to their last step yet leave the task undone. Across a mobile, a desktop, and a web benchmark, this store-time check separates repeated runs that improve from ones that degrade as faulty programs accumulate, worth 1.75-2.6 tasks per benchmark, the same direction on all three; a fallback that explores afresh when no program fits brings PreAct level with a strong record-and-replay baseline. We also report what did not matter: prompt wording, runtime guardrails, and whether a language model or a plain embedding retriever selects which program to reuse.

Tool-using LLM agents increasingly use the Model Context Protocol (MCP) to answer from heterogeneous evidence sources, including search, APIs, databases, clinical records, and formulary tools. Standard factuality metrics usually test whether an answer is supported by pooled evidence, missing a provenance-sensitive failure mode: a claim may be supported somewhere while being attributed to the wrong source. We call this cross-source conflation. We introduce ProvenanceGuard, a source-aware verifier for MCP-grounded answers. It consumes captured MCP traces with stable tool IDs, source IDs, and raw outputs; decomposes answers into atomic claims; routes claims to source-specific evidence; checks support with NLI and a token-alignment proxy; compares stated attribution with the routed source; and returns per-claim verdicts plus an answer-level allow/block decision. Blocked answers can be repaired with retrieval-augmented answer revision and re-verified. We evaluate on 281 medical-domain MCP-agent traces. A 266-trace adjudicated subset yields 2,325 LLM-assisted claim labels split by trace; 361 held-out labels are human-verified. On the 40-trace held-out split, ProvenanceGuard achieves block F1 0.802 and source accuracy 0.858 over 260 source-eligible claims, outperforming source-blind baselines that do not emit claim-to-source IDs. On a harder multi-source benchmark it reaches block F1 0.846, while source-plus-relation accuracy drops to 0.229, showing that exact source ownership remains difficult with semantically close sources. Repair-and-reverify resolves all blocked answers in the full trace set, often via conservative fallback. In 50 controlled clinical conflation probes, ProvenanceGuard detects all injected attribution swaps with no retained wrong attribution. These results show that source attribution is an independent axis for factuality verification in MCP-based agents.

Evaluating the decision-making capabilities of large language models (LLMs) is a growing research priority, yet existing benchmarks focus on isolated cognitive tasks such as reasoning, knowledge retrieval, and economic rationality in stylized settings. These evaluations overlook the defining challenge of real executive decision-making: integrating conflicting recommendations from specialized stakeholders under information asymmetry, organizational constraints, and temporal dependencies. We introduce \textsc{CEO-Bench}, a multi-agent benchmark that evaluates LLMs on CEO-level strategic resource reallocation — the process of redirecting capital across business units in a multi-round, constraint-rich organizational environment. In \textsc{CEO-Bench}, LLM agents receive conflicting advice from four role-conditioned C-suite advisors (CFO, CTO, COO, CMO), each with private signals and distinct priorities, and must synthesize these into a concrete allocation plan evaluated along four dimensions: role integration, conditional boldness, history-sensitive judgment, and plan validity. Experiments across five frontier models on 13 scenarios reveal that all models achieve high structural validity but diverge sharply on strategic calibration — the hardest capability layer. We identify systematic failure modes including single-advisor capture, conservative default under ambiguity, and historical amnesia, and uncover a structural integration-boldness tradeoff: models that engage more deeply with conflicting perspectives tend to produce less decisive action. These findings delineate the current capability boundary of LLMs as organizational decision-makers and inform the design of future AI-assisted executive systems.

Agent skills are emerging as an important attack surface in LLM-based systems. Through an empirical study of existing skill scanners, we find that current defenses primarily rely on textual descriptions, manifests, and source code as the main signals for security analysis, which can leave visually conveyed malicious intent insufficiently examined. This creates a practical blind spot: harmful operational instructions hidden in images may bypass scanning while still being recoverable by multimodal agents during deployment. To systematically investigate this threat, we propose SkillCamo, a document-mediated multimodal instruction attack that conceals malicious instructions within images bundled with a skill while rewriting the surrounding documentation to naturally reference those images as part of the normal workflow. Thus, the attack does not rely on the image alone, but on the joint interpretation of textual guidance and visual payload at execution time. To defend against such attacks, we further propose ExecScan, an execution-grounded multimodal scanning module that performs intent extraction, behavior reconstruction, abuse assessment, and deliberative execution simulation over skill artifacts. ExecScan jointly analyzes documentation, code, referenced resources, and visual content to recover hidden instructions, reconstruct executable behavior chains, and identify downstream risks such as exfiltration, destruction, persistence, deception, and privilege escalation. Extensive experiments show that image-hidden malicious instructions challenge existing skill scanners, while ExecScan can improve the skill scanning performance.

Access to evidence-based psychotherapy remains limited worldwide, with long waitlists even in high-income regions. Recent advances in large language models (LLMs) offer potential for scalable mental health support when designed with clinical oversight and safety mechanisms. We present Mind Companion, an LLM-based embodied conversational agent integrating multi-layered psychological analysis with process-based therapy principles. The system performs real-time analysis of client statements across fact extraction, psychological flexibility process detection, emotion recognition, and safety monitoring. Analysis results are stored for supervising clinicians to inform therapeutic planning. Response generation incorporates retrieval-augmented generation from evidence-based therapeutic literature and context-aware prompting. Responses are delivered through an embodied avatar with synchronized speech synthesis and animation. We evaluated three LLM configurations (GPT-4.1-mini, GPT-5.2, Claude Sonnet 4.5) against therapist responses from real therapy sessions using automated LLM-judge assessment and expert evaluation with 11 professional psychotherapists. GPT-5.2 achieved higher ratings than human therapist responses across understanding, interpersonal effectiveness, collaboration, and therapeutic alignment in both evaluations, demonstrating the feasibility of LLM-based conversational agents as tools to complement clinical care.

Knowledge distillation transfers a teacher's competence to a small student but is brittle in the small-student regime: forcing the student to imitate logits from a much larger teacher concentrates it on the teacher's sharpest modes, hurting generalization on benchmark families beyond the training corpus. Reinforcement learning (RL) avoids logit imitation by training on the student's own rollouts. However, on questions where every rollout fails-yielding zero advantage and being silently discarded-injecting a stronger teacher's response into the policy gradient breaks the on-policy assumption and induces drift. We introduce Zone of Proximal Policy Optimization (ZPPO), inspired by Vygotsky's zone of proximal development, which keeps the teacher inside the prompt rather than the policy gradient. On hard questions, ZPPO constructs two reformulated prompts: a Binary Candidate-included Question (BCQ) pairs one correct teacher response with one incorrect student response as anonymized candidates the student must discriminate, and a Negative Candidate-included Question (NCQ) aggregates the student's wrong rollouts into a single prompt to surface their shared failure modes. A prompt replay buffer recirculates each hard question until it either graduates-the student's mean rollout accuracy on it reaches half- or is FIFO-evicted under finite capacity, amplifying BCQ and NCQ inside the student's current zone of proximal development. On the Qwen3.5 family at four student scales (0.8B-9B) with a 27B teacher, post-trained as vision-language models and evaluated on a 31-benchmark suite (16 VLM, 10 LLM, 5 Video), ZPPO outperforms off/on-policy distillation and GRPO, with the largest gains at the smallest scale.

On-policy self-distillation (OPSD) has proven effective for post-training large language models (LLMs), yet its application to diffusion LLMs (dLLMs) remains unexplored. Existing OPSD methods are inherently autoregressive-centric. They inject privileged information via left-to-right prefix conditioning with token-level divergence supervision, a design that fundamentally conflicts with the arbitraryorder generation of dLLMs. We introduce d-OPSD, the first OPSD framework tailored for dLLMs. Our approach makes two core contributions. First, we reframe self-teacher construction by using self-generated answers as suffix conditioning, enabling the student model to learn from "self future-experience" rather than privileged prefixes. Second, we shift supervision from token-level to step-level, aligning training with the iterative denoising process of dLLMs. Experiments across four reasoning benchmarks show that d-OPSD consistently outperforms RLVR and SFT baselines with superior sample efficiency, requiring only around 10% of the optimization steps by RLVR and opening a promising pathway for dLLM posttraining. The code is available at https://github.com/xingzhejun/d-OPSD.

Modern large language model (LLM) serving clusters distribute inference requests across multiple worker processes on different GPUs, but failures are prevalent at scale. When a worker fails, the cluster simultaneously loses the failed worker's GPU-resident key-value (KV) caches and serving capacity, leaving surviving workers to absorb the redirected traffic while re-running interrupted requests from scratch. Existing fault-tolerant systems either restart interrupted requests from scratch or restore KV caches from checkpoints stored on a fixed neighboring worker, but both approaches route recovery work without considering current cluster load and leave the recovering worker idle during model reload. We present LUMEN, a fault-tolerant LLM serving system that treats recovery as a load-aware coordination problem across three decision points: checkpoint placement before failures, interrupted-request distribution at failure time, and serving capacity restoration during model reload. We evaluate LUMEN using both prototype experiments and large-scale simulations and demonstrate significant improvements in serving and recovery times.

Heterogeneous LLM serving stacks split scheduling into two layers that optimize in isolation: model routers pick a model from quality and cost signals while ignoring instance load, and serving load balancers optimize queues while ignoring quality. We present RouteBalance, a serving-aware scheduling layer that fuses both into a single online assignment over concrete model instances, jointly trading off quality, latency, and cost. A batched in-process predictor stack and dead-reckoned instance state keep the joint decision cheap on the request hot path ($\approx$32 ms at 12 req/s). On a 13-instance, 28-GPU heterogeneous cluster serving four model sizes, a single deployed RouteBalance stack traces the upper region of the three-way quality-cost-throughput frontier. Sweeping one weight vector reaches both the highest routing-decision quality (DeepEval $0.419$, $+0.013$ over the strongest baseline, $95%$ CI $[{+}0.005,{+}0.022]$; the ordering holds when a second judge re-scores the actually served text) and, at its cost-priority corner, per-request cost that ties the cheapest baseline. With router engineering equalized against concurrent-scoring baseline variants we build, its balanced preset serves at $2.8$ s and $30$ req/s, leading $2.6$ to $4.1\times$ ahead of enhanced BEST-Route at high load. (Deploying those routers as published, one serial scoring call per request, makes them collapse $23\times$ under load, a deployment-architecture effect we isolate separately, not the routing result.) A four-arm isolation shows the benefit follows from pricing latency at model-selection time; the learned predictors contribute calibration and SLO headroom rather than the headline frontier. Code: https://github.com/AKafakA/route-balance

State Space Models (SSMs) such as Mamba-2 offer linear-time inference but their memory footprint limits edge deployment. Prior ternary SSM work (Slender-Mamba) trains from scratch on 150B tokens; we show a pretrained checkpoint suffices, reducing the marginal token budget by 1,000x. Using grouped quantization-aware training (QAT) with knowledge distillation from a frozen FP16 teacher, we compress Mamba-2 1.3B to 3.61x (2,687 to 744 MB) and achieve 48.1% zero-shot accuracy (7-task average) in just 102M tokens (4 GPU-hours, single H100) — approaching Bi-Mamba's 48.4% (within +/-0.9pp CI). This QAT-from-pretrained setting reveals zero-ratio collapse, a novel instability caused by learnable quantization scales that does not arise in from-scratch training. We further show that post-hoc correction strategies effective for Transformers fail for SSMs due to error accumulation through the recurrence. These results demonstrate that ternary SSMs do not require expensive from-scratch training: QAT from pretrained checkpoints with KD is a data-efficient alternative.