Agent 进展

Multimodal Large Language Models (MLLMs) are strong perceivers of images and video. We ask how far that reach extends into acting: dropping an MLLM directly into a drone's control loop, with its entire action space declared solely in the prompt. Recent systems approach this setting but increasingly narrow the model's decision-making. We widen it back. We introduce DroneCATS-Agent, an architecture where the MLLM is a swappable component, and DroneCATS, a benchmark treating the model as the independent variable. Beyond merely flying toward a pixel, our agent entrusts the model to yaw and search, deliberate when unsure, and self-declare arrival---all without fine-tuning or function-calling schemas. Evaluating frontier and open models across four core capabilities---approaching a visible target, tracking a moving one, searching outside the initial view, and commanding a multi-drone fleet---reveals that even the simplest embodied settings are far from solved. Crucially, to identify what breaks first at the edge, our roster scales down to 2B parameters. The findings expose a stark paradox: it is not the flying that fails. Small open models often navigate into the success radius more reliably than frontier models, yet lose the episode by declaring arrival prematurely or not at all. Multi-drone commanding amplifies this divide, with small models failing by blindly copying a single coordinate across distinct views. Viewed as vision-language-action agents, the models' spatial perception holds up, but their action protocol does not. What separates a deployable edge model from a frontier model is not navigation, but the discipline to sustain a declared protocol and emit the correct terminating action. The open problem is closing this gap at onboard compute costs---yielding a fast model that plans persistently and knows exactly when it is done---and DroneCATS is built to measure that distance.

Text-to-image (T2I) models remain vulnerable to jailbreak attacks that elicit Not-Safe-For-Work (NSFW) content, despite increasingly being guarded by heterogeneous, multi-layer safety stacks combining text filters, image classifiers, and cross-modal detectors. Existing jailbreak studies either optimize against individual filters or query the complete pipeline with aggregate feedback, making it difficult to identify the active constraint and adapt to conflicts across safety layers. In this paper, we introduce the Detection Surface, a unified geometric framework that characterizes the decision boundaries induced by heterogeneous T2I safety filters and their joint effect on the jailbreak search space. This formulation reveals that successful evasion is governed by a sparse and non-convex region shaped by cross-layer conflicts, where mutations that bypass one filter may increase exposure to another. Motivated by this analysis, we propose CRACK, a multi-agent debate framework for adaptive jailbreak search that decomposes jailbreak search into exploration, diagnosis, and arbitration. CRACK coordinates an Attack Agent, a Defense Agent, and a Judge Agent to iteratively generate prompt mutations, obtain layer-specific diagnostic feedback, and optimize mutation strategies through reward-guided refinement. Through repeated rounds of debate, CRACK adapts its search direction to the evolving cross-layer constraints while preserving the original harmful intent. Extensive experiments across multiple T2I models, datasets, and safety configurations show that CRACK achieves Attack Success Rates (ASR) of up to 99.63% under composite defenses, while requiring fewer queries than existing methods and maintaining semantic fidelity.

Computer-use agents are emerging as a paradigm for agentic human-AI interaction, combining language reasoning with multi-modal interface grounding to operate GUIs. Yet their effectiveness for blind screen-reader users in real-world desktop workflows remains unclear. We present a three-week diary study with 8 blind users using OLLA, a screen-reader-accessible CUA prototype, collecting 1,258 commands across 12 applications with screenshots, UI trees, model responses, and action traces. We evaluate GPT-5 during deployment and re-execute the same commands with four additional models. GPT-5 achieved the highest success rate at 52.5%. Trace analysis reveals grounding, planning, constraint-tracking, and termination failures, while interviews reveal beyond-automation needs.

Large Language Model (LLM) agents increasingly rely on external skills, yet standard evaluations obscure whether retrieving these skills actually helps. Aggregate metrics often compare retrieved versus non-retrieved tasks, introducing severe selection bias and failing to isolate the true effect of skill use. To measure this actual-use capability-which we formalize as Skill Following (SF)-we introduce the Retrieval-Invoked Actual-Use Effect (RAE). RAE computes the same-task outcome difference between matched skill-enabled and skill-disabled executions, conditioned exclusively on tasks where the agent actively retrieved a skill. Evaluating 17 LLMs across coding and mathematical domains, we uncover a stark evaluation paradox: models frequently show positive aggregate retrieval lift but negative RAE. On MBPP+, multiple models that appear to benefit system-wide actually harm their own performance on the exact tasks where retrieval occurred. These findings demonstrate that aggregate averages can create a misleading illusion of tool-use proficiency, whereas RAE directly measures whether the retrieval-to-answer pipeline genuinely rescues more outcomes than it harms.

Prospective memory means carrying out a deferred intention at the right future cue while other work continues. Benchmarks now isolate it as an agent skill, yet frontier LLMs still struggle: the best published PM-Bench scaffold reaches only 65.1% Set-F1. We argue that this loop is schema-constrained state tracking rather than open-ended reasoning, and that small models can execute it when the action space is typed. We propose the Prospective Intention Store (PIS) that puts lifecycle logic in code and scoped language work on the model. The scaffold is agentic and training-free: no selector fine-tuning and no trajectory distillation. On PM-Bench, DeepSeek-Chat with PIS reaches 82.9% Set-F1. On Gemma-E2B, Set-F1 is only 4.2% without a store and at most 6.6% under seven retrospective memories, while PIS reaches 66.2%. PIS further reaches 70.1% Set-F1, where retrospective memory methods stay at most 54.4%. PIS sets a new state of the art on this benchmark and enables small models to surpass the published large-model scaffold.

Training open-ended agents via reinforcement learning (RL) is hindered by the lack of verifiable gold answers and scalable rubrics. Moreover, even near the model's capability boundary, long-horizon open-ended agentic tasks often yield brittle and unstable rewards, resulting in weak or noisy rollout contrast that obscures fine-grained optimization signals for group-based policy learning. To address these challenges, we propose ARISE-RL, a novel full-cycle self-evolution framework that couples a task/rubric Generator and a reasoning Solver through rubric-mediated co-evolution. The Generator grounds tool-related rubric criteria in real tool observations and is rewarded for producing valid, intermediate-difficulty tasks aligned with the Solver's evolving capability boundary. The Solver, in turn, learns from fine-grained rubric satisfaction signals through multi-step reasoning and tool use. We further introduce Reward-Gated Self-Evolution Distillation (RG-SED), which selectively distills a memory-augmented variant of the same policy back into itself only when the memory yields empirical reward improvement, thereby reducing distribution mismatch and avoiding blind imitation of noisy guidance. Finally, to support rigorous evaluation, we present ECR-Bench, an expert-calibrated rubric benchmark suite covering single-tool deep research and multi-tool travel planning. Extensive experiments demonstrate that ARISE-RL consistently achieves robust and stable overall state-of-the-art performance across all evaluated benchmarks.

Creative generation tasks, such as narrative writing and scientific ideation, demand both high-quality outputs and distinct responses across independent runs to maximize exploration. Multi-Agent Debate (MAD) has shown strong quality gains on factual and reasoning tasks, making it a natural candidate for creative generation. However, we find its convergence-driven design actively suppresses output diversity across independent runs, creating an inherent trade-off with creative tasks. We theoretically show that preserving diversity among agents within each debate session is a necessary condition for achieving diverse outputs across independent runs. Building on this finding, we propose Creative-MAD, which introduces two synergistic interventions to sustain agent divergence. Specifically, Cognitive Lens Assignment counters identity drift by anchoring each agent to a distinct and persistent cognitive mode, while Embedding-based Peer Selection counters majority pull by limiting each agent's context to its most semantically distant peers. Experiments across four creative benchmarks demonstrate that Creative-MAD significantly enhances both lexical and semantic diversity while maintaining MAD's output quality.

Vehicle Routing Problems (VRPs) are fundamental combinatorial optimization problems with widespread applications in various scenarios. The advanced optimization solvers can effectively solve such problems. However, modeling complex VRP variants for solvers often requires substantial domain expertise, which limits the accessibility of advanced optimization technologies. In this paper, we propose Reinforcement Learning Enhanced LLMAgents(RLEA), a multi-agent framework designed to automate the modeling of complex VRPs. RLEA introduces a lightweight neural Planner trained with Soft Q-learning to efficiently orchestrate the actions of LLM-based agents. In addition, we equip the system with an evolutionary memory module and retrieval-augmented generation, enabling the agent to leverage both accumulated experience and external solver knowledge during program generation and refinement for solving VRPs. We evaluated 48 distinct VRP variants across various solvers. The experimental results demonstrate that RLEA outperforms the previous state-of-the-ar method, achieving a 16.67% higher success rate while significantly reducing runtime errors. These results validate that integrating reinforcement learning with LLM-based reasoning is highly effective for automated optimization modeling. The appendix is available at: https://doi.org/10.5281/zenodo.19134435.

Large Language Model (LLM) agents based on the ReAct paradigm have demonstrated remarkable capabilities in tool use and task execution. However, ReAct suffers from a fundamental efficiency problem: every query triggers a complete reasoning loop from scratch, and similar queries repeat identical steps without leveraging historical experience. We propose TRIAGE,a three-level routing framework that reduces token consumption by reusing historical execution trajectories. Its core innovation is TaaS (Trajectory-as-a-Skill), which abstracts historical execution trajectories into reusable skills, realizing 'experience as a service'. TRIAGE classifies queries into three levels: (1) Direct Reuse-identical queries, 0 tokens; (2) Skill Substitution-similar queries, 0 tokens via deterministic parameter substitution; (3) Full ReAct-novel queries, automatically stored for future reuse. In large-scale experiments on 1,007 security monitoring queries, TRIAGE achieves 62.3% token savings, with 56.0% of queries at Level 2 and 5.5% at Level 1, both executing at zero cost. Cross-domain validation on ToolBench (15 domains, 345 queries) achieves 76.3% token reduction, confirming the generalizability of semantic routing. An online learning experiment demonstrates cold-start-to-mature evolution: the L2 hit rate rises from 0% to 57% within the first 100 queries, and the average token cost drops from 198 to 74.7. We also propose an automatic Skill extraction mechanism that distills high-frequency trajectory patterns into deterministic Skills, creating a positive feedback loop of 'the more you use it, the more efficient it becomes'.

The growing rate at which LLM agents interact with one another raises key questions about language evolution in multi-LLM-agent settings, with implications for safety and monitorability as well as for linguistic accounts of LLMs. To address these questions, we introduce GlossoGen, a novel platform for studying multi-agent language evolution in complex scenarios. Within GlossoGen, we build the SaveVeyru scenario, which requires agents with partial information to communicate under pressure. We find that language evolution does occur between LLM agents, that the resulting languages are compositional and morphologically productive, and that they deviate from the LLMs' English prior in ways that render them incomprehensible to humans. Moreover, we identify several qualities essential to this evolution: pressure towards efficiency; the strength of the models backing the agents; and access to a "postmortem" stage in which agents can agree on linguistic conventions. Importantly, we observe that different conditions govern the transmission of language to new agents. Specifically, we find that agents learn new languages from usage alone, take an active role in this learning, and that while stronger models are required for novel language emergence, weaker models can learn an existing language once it has emerged. Taken together, our results indicate that current LLMs have the potential for cumulative cultural evolution — previously attested only in humans — with mixed populations of agents developing capacities that go beyond their lowest common denominator.

Prompt optimization can improve multi-agent LLM systems, but the prompts being optimized often serve two entangled roles: generating task-relevant content and specifying execution-critical protocols, such as message routing, output formatting, and termination signals, on which the underlying code relies. As a result, a prompt edit intended to improve content generation can inadvertently corrupt the protocol and cause the entire agent pipeline to fail. Our key observation is that these two roles have different representations: execution protocols are typically structured, while task-relevant content is usually expressed in unstructured language. Based on this, we propose control-data flow separation, where execution-critical control is represented as typed, validated program objects, while task-relevant language remains the optimizable data flow for agent communication. This design allows optimizers to improve multi-agent behavior without exposing the routing or formatting interface to prompt drift. Across synthetic reasoning, collaborative review generation, and insurance rating workflows, our framework empirically achieves 100% eventual protocol validity while consistently improving task performance.

Remote sensing visual models have continuously advanced various interpretation tasks. However, the research process behind model improvement still heavily relies on manual expertise, requiring extensive trial-and-error iterations in model design, data processing, and performance diagnosis. Existing agent-based approaches mainly focus on task execution and workflow orchestration, while lacking the capability of autonomous research iteration for continuous performance optimization. To address this issue, we propose RingMoClaw, an experience-inspired self-evolving multi-agent framework for remote sensing visual interpretation. RingMoClaw integrates a research branch, a quality-control branch, and a dual-stream dynamic experience bus to establish a closed-loop optimization process covering strategy generation, experiment execution, independent review, and experience accumulation. The heterogeneous Critic mechanism provides stage-wise diagnosis and feedback, while the dual-stream experience bus incorporates external knowledge and internal experimental experience to guide strategy evolution and eliminate ineffective searches. Extensive experiments on four remote sensing downstream tasks, including object detection, scene classification, semantic segmentation, and change detection, demonstrate the effectiveness and generalization of RingMoClaw. Compared with the corresponding baseline models, RingMoClaw improves performance by 1.84% mAP$_{50}$ on object detection and achieves consistent gains across the other three tasks, while reducing the required evolution steps by over 40% compared with existing research automation frameworks. These results suggest that RingMoClaw offers a feasible route from task execution toward continuous research driven model evolution in remote sensing.

Large language model (LLM) agent failures often contain multiple related errors rather than a single mistake. Existing attribution methods usually identify a responsible agent, step, or root cause, but do not explicitly model dependency between errors. We introduce EDGE, an Error Dependency Graph-guided multi-Error attribution framework. EDGE constructs an error dependency graph from observed error events and validates a reliable causal subset through counterfactual rollout. The inference graph guides a two-stage LLM-as-judge detector for error attribution, and the intervention-validated subgraph provides a more reliable basis for explanation and repair analysis. Experiments on TRAIL and MAST show that EDGE improves category-level multi-error attribution across most evaluated models and settings. Experiments with adapted Who&When-style prompts show that the graph helps across prompting strategies. These results suggest that dependency structure is a useful diagnostic prior for agent failures beyond isolated root-cause prediction.

Reinforcement learning is a natural way to post-train LLM agents for long-horizon interactive tasks judged only by end-of-task verification, yet a shared belief holds that outcome-only RL soon hits a ceiling on small open models. Recent work therefore compensates around the training with denser rewards, SFT priors, skill libraries, curated memory, or multi-agent orchestration. We argue the ceiling is an artifact of two failures of common practice. Signal starvation: group-relative RL with sparse outcome-only rewards yields a gradient only when a task's rollout group mixes successes and failures, so under-scaled exploration silences exactly the hardest, most instructive tasks. Policy drift: squeezing many updates out of a small task pool degrades the policy itself, as an unanchored objective lets the sampling distribution collapse exactly when saturation has already made informative groups rare. We present CANOPY (Coverage-ANchored On-PolicY RL), a minimalist protocol attacking both directly: scale same-task exploration until the natural signal reappears, keep every update on-policy, KL-anchored, and confined to the agent's own action tokens, then cash in an enlarged interaction budget at test time. On AppWorld, a long-horizon interactive coding benchmark, a Qwen3-14B policy trained with CANOPY through environment interaction alone--without task-specific supervision, auxiliary credit signals, or elaborate agent scaffolding--topped the public leaderboard (Feb. 2026; Test-Normal TGC 86.9, Test-Challenge 67.6), and the same design principles lift Qwen3.5-9B on SWE-bench Verified by 16.6 points. Agentic RL alone thus internalizes long-horizon capability directly into a small open model; we plan to release the complete training stack at https://github.com/AlibabaResearch/SignalCoverageRL.

Safe agents can fail together. Multi-agent LLM systems (MAS) move information, state, decisions, and authority across principal boundaries, creating failures that local checks may miss. Without an execution-level view, a multi-agent setting can easily be mistaken for evidence of a genuinely multi-agent security effect. We thus systematize MAS security through an execution-centered analysis of 197 works, covering six interaction interfaces, four adversary positions, seven system-level risks, and eight recurring attack paths. We introduce an A-I-R framework that organizes attacks by adversary position, interaction interface, and resulting system-level risk, unifying otherwise fragmented attack mechanisms across MAS. We organize defenses through a five-part contract covering path target, observation, intervention, trust boundary, and recovery, and identify path closure and recovery as key challenges. We audit 44 evaluation and benchmark works and identify open challenges in isolating interaction effects, designing comparable and diagnostic metrics, supporting reuse across MAS designs, and evaluating open-system operation. Together, these findings motivate an interaction-aware view of MAS security: trace attacks end to end, test whether defenses close those paths, and evaluate system-level effects with appropriate counterfactuals.

Retrieval over visually rich documents has a representation problem: important content often lives in tables, charts, figures, and layout relations that plain OCR linearizes, corrupts, or omits. ColPali-family visual retrievers address this with patch-level multi-vector indexes and late-interaction scoring, keeping image-derived retrieval on the query-time serving path. We introduce MIDR (Multimodal Indexing for Document Retrieval), a training-free framework for enrichment-augmented indexing that shifts multimodal reasoning to index time. During ingestion, a multimodal LLM converts rendered pages into verified textual fields that are indexed with BM25F and optionally fused with dense retrieval, enabling text-centric serving over multimodally grounded evidence. On ViDoRe V3, MIDR Hybrid achieves 0.6219 average nDCG across five English domains, a 23.0% relative gain over BM25, remaining competitive with ColQwen2.5. On two French-document domains, enrichment bridges English queries and French page text, lifting BM25 from 0.1532 to 0.5448 nDCG and outperforming ColQwen2.5. Across all seven domains, MIDR leads ColQwen2.5 on four while using approximately 9x smaller index memory and approximately 2x lower query latency. These results establish index-time multimodal reasoning as a compelling accuracy-deployment alternative to serving-time visual late interaction.

Multi-agent combinatorial optimization problems are notoriously challenging due to their NP-hard nature. Recent parallel autoregressive neural solvers improve inference efficiency by allowing agents to make decisions simultaneously, but their performance often degrades on large-scale instances. This is largely attributable to weak modeling of local geometric structures and the fact that conflicting task selections are handled only after action generation. To address these limitations, we propose GeoPAR, a geometry-guided parallel autoregressive reinforcement learning framework for scalable multi-agent combinatorial optimization. GeoPAR integrates three key components: (1) a projection-window sparse geometry mechanism that builds lightweight local candidate neighborhoods through multi-directional projections, (2) sparse edge-biased attention that injects these geometric relations into node representations, and (3) cache-guided conflict-aware assignment that reuses the geometric cache during decoding to suppress duplicate selections of exclusive tasks. Experiments on heterogeneous vehicle routing and open multi-depot pickup-and-delivery problems show that GeoPAR improves large-scale zero-shot generalization while substantially reducing rollout steps and maintaining efficient inference.

LLM coding agents are crossing from task-scale to project-scale: single assisted efforts that run for months, across repeated context compactions, on codebases far larger than any context window. We report operational experience from one such effort (a research project under a single continuous Claude Code session line since January 2026, driving a 633,000-line codebase, with its memory subsystem continuously instrumented since July 2026) and describe SIx Harness, the open-source memory and continuity infrastructure that emerged from keeping it coherent. The harness combines per-project long-term memory (hybrid lexical-vector retrieval over SQLite, fully local), precision-gated context injection, anti-recurrence stores for decisions and dead-ends, conventions engineered to survive compaction, and, the part we argue is missing from operational practice, reliability engineering for the memory subsystem itself: a session-start health gate with discriminated failure modes, heartbeat telemetry designed so that no enumerated failure mode can pass unrecorded, and alert-fatigue budgeting borrowed from SRE practice. We present what we believe is the first months-scale instrumented operational record of a persistent agent-memory subsystem in production development use: 78,933 hook invocations; 85 recorded failures, none silent: 84 in the subsystem's first three weeks, one since, none in the final 20 days; an injection layer whose ten-day precision instrument shows zero false fires against an intact denominator; and three production incidents traced from instrument reading to structural fix. From the record we distill seven design principles, state our limitations plainly (N=1, no control arm, self-reported), and publish a tagged pre-registered ablation protocol that any team can run with the released MIT-licensed kit.

Web agents have achieved significant success in automating complex internet tasks but deploying them in real-world environments requires continuous online adaptation. Given that deploying powerful proprietary models remains commercially cost-prohibitive, practitioners must rely on lightweight local models that evolve post-deployment via online teaching from a stronger teacher. However, standard interactive feedback imposes prohibitive costs. We show that conventional trajectory-level preference optimization wastes budget on both unresolvable episodes and redundant execution turns. To resolve these inefficiencies, we propose Score-Guided Online Teaching with Budgeted Trajectory Trimming, a budget-aware framework that systematically orchestrates when and what to teach. Specifically, our framework integrates a solvability-aware teacher gate to dictate when to query the teacher model and a score-guided turn selection mechanism to decide what informative turns to retain. Extensive experiments on MiniWoB and TimeWarp demonstrate that our method achieves comparable first-pass success while reducing teacher calls by 22.6% and student training compute by 52.1% on average. Our code is available at https://github.com/zjw131f1fc/budgeted-online-teaching.

Despite the strong capabilities of multimodal large language models (MLLMs), their parametric knowledge remains incomplete and difficult to update, motivating multimodal retrieval-augmented generation (RAG) to ground responses in external text and images. However, existing benchmarks face two major limitations: (i) they typically emphasize single-hop retrieval or reasoning over a small set of provided contexts rather than open-domain evidence discovery; and (ii) they provide fragmented coverage of cross-modal reasoning paths, leaving complex multi-hop and multi-image reasoning underexplored. In this paper, we introduce CrossModalQA, an open-domain benchmark for evaluating multimodal retrieval and reasoning over heterogeneous corpora. CrossModalQA contains 1,863 question-answer pairs constructed from 4,987 Wikipedia articles and 4,431 Wikimedia Commons images. It covers five complementary reasoning paths: vision-to-text, text-to-vision, vision-to-text-to-vision, multi-image intersection, and image-set reasoning. Every question requires retrieving and composing distributed textual and visual evidence, with an average reasoning depth of 3.50 hops. We construct the benchmark through multimodal knowledge graph-guided subgraph sampling and apply rule-based consistency checking and LLM verification to ensure multimodal dependence and traceable evidence. Extensive experiments demonstrate that existing multimodal RAG systems struggle to recover complete evidence chains and can underperform closed-book models when incomplete retrieval introduces distracting context. Further analysis reveals that complete cross-modal retrieval contributes more to answer accuracy than generator scaling, while multi-image retrieval and reasoning remain the primary bottlenecks limiting end-to-end performance.

Traditional finance relies on experts to hand-craft factors through a principled process grounded in economic rationale. Recent LLM-based multi-agent systems have automated this process, scaling factor mining far beyond manual effort. However, these automated approaches optimize directly for returns and rarely check whether a generated factor still expresses the economic hypothesis that motivated it. We identify this inconsistency between mathematical form and economic meaning as a structural failure mode of return-oriented automation. The resulting factors blur the line between real signals and spurious correlations and break down across regime shifts. We propose FaVOR (Factor Validation through Observable Reasoning), an agentic framework that restructures factor mining around hypothesis-level evidence rather than return outcomes. In place of the standard hypothesis-to-formula leap, FaVOR enforces a three-stage consistency loop tying mathematical form to economic rationale throughout. (1) Decomposition splits a broad economic hypothesis into independent observable conditions. (2) Validation checks whether each factor reflects its intended condition. (3) Integration merges them into a composite whose structure remains interpretable. On the CSI 500 and S&P 500 in 2025, FaVOR outperforms existing baselines while remaining effective across regimes. FaVOR shows that hypothesis-grounded factor discovery produces signals that are interpretable by construction, regime-robust, and economically faithful. The code is available at https://github.com/damilab/FaVOR.

Large language model-based multi-agent systems are evolving from fixed interaction topologies toward dynamically orchestrated Agent Swarms. However, existing benchmarks are still largely based on single-agent or general-purpose agent tasks, making it difficult to systematically evaluate key orchestration capabilities. We propose SwarmBench, a benchmark that evaluates model performance from multiple perspectives, including accuracy, efficiency, cost, and process quality. Experimental results show that current models exhibit substantial differences in orchestration capability. These differences are reflected not only in final accuracy, efficiency, and cost, but also in the overall quality of the orchestration process itself. Based on these findings, we further propose SwarmExp, a simple yet effective method based on experience extraction and experience replay, which consistently improves the orchestration performance of large language models.

This paper investigates distributed online optimization for multi-agent dynamical systems with constrained inputs and time-varying cost functions. While online convex optimization offers a principal framework for sequential decision-making, existing online learning and optimization algorithms typically require accurate system models, limiting their applicability in practical settings. To overcome this challenge, we propose a distributed bandit online feedback optimization algorithm that relies solely on real-time input-output data. The algorithm employs a smoothing zeroth-order one-point estimator to construct local gradient approximations directly from cost evaluations. Additionally, to enforce input constraints effectively, we integrate a projection-free conditional gradient update, making the algorithm well-suited for online and large-scale settings. Furthermore, we establish a sublinear dynamic regret bound that depends on a temporal variation measure of system non-stationarity. Finally, numerical simulations demonstrate the effectiveness of the proposed algorithm.

Large language model (LLM)-based multi-agent systems tackle complex reasoning by orchestrating how multiple agents are configured and how they collaborate. A central challenge is to adapt orchestration to the evolving collaboration state. Routing from the query alone cannot adapt to intermediate progress or errors, which hurts accuracy. Routing from the complete execution history supplies this missing context, but forces later decisions to process every prior step, including redundant or low-utility ones. This creates an execution-history overload that inflates cost. Effective orchestration instead requires a compact state that captures useful progress without accumulating redundant context. We propose Gated-Memory Routing, which conditions each decision on the query and a learned execution memory. A learned Memory Write Gate commits only non-redundant reasoning steps, and a learned Retrieval Gate supplies each agent a compact, relevant subset, so every decision conditions on a clean, informative state. At each step, the system selects the next role and backbone from this memory, while an Adaptive Halting Controller stops execution once the memory contains sufficient evidence for answering. Across five reasoning and code-generation benchmarks, our framework is both effective and efficient: it attains the best average accuracy, exceeding the strongest baseline by 2.44 points, while reducing HumanEval inference cost by 31.9% relative to that baseline. Code is available at https://github.com/rajibrhasan/gated-memory-routing

Production deployments of large language model (LLM) agents remain unreliable on long, multi-step workflows even as benchmark success rates climb steadily. We argue this gap is largely an artifact of task horizon: benchmarks are dominated by short-to-medium horizons where success remains high, while production workloads demand an order of magnitude more dependent steps. We measure the effect directly, characterizing the shape of agent degradation and disentangling its cause across a large controlled study spanning nine models, six open models from 1.2B to 671B parameters, and three deployed proprietary systems; four task families, including a genuinely agentic tool-use loop; five horizons; and three context regimes. Task success follows a geometric law governed by a single per-step reliability parameter, which rises with model scale but saturates well below 1 even for the strongest models, guaranteeing eventual collapse at sufficiently long horizons. The effect is sharpest on the agentic task, where every model tested, including widely deployed systems, falls from near-perfect success to near zero within sixteen steps of (n=10,664 analyzed trajectories. Degradation is driven by step count rather than context length: bounding the context window steepens decay rather than easing it (logit slope -0.69 vs. -0.44), p=3x10-6), contradicting a lost-in-the-middle explanation and warning against a common production shortcut. Projecting measured reliability onto representative benchmark horizons quantifies a substantial gap between benchmark and production conditions, from 0.42 at GAIA-length horizons to 0.24 at hundred-step production horizons. For teams responsible for agent orchestration and reliability at scale, these results argue for horizon-aware evaluation and reliability budgeting in place of aggregate pass-rate metrics. Code, prompts, seeds, and raw trajectories are released.

End-to-end GUI testing is essential for verifying web applications, yet existing evaluations rely on predefined checklists and are confined to the data and frameworks of web generation benchmarks, leaving the bug-discovery ability of vision-language models (VLMs) systematically untested. We introduce Code-driven Agentic Testing (CAT), a paradigm in which the agent writes Playwright code to drive the browser, gathers feedback, and autonomously explores web applications to uncover bugs. We instantiate CAT with CATJudge, an agentic framework that unifies Browser-Use and Computer-Use tools within a single environment and CATTest, a benchmark of 102 AI-generated web applications with carefully annotated bugs, built through close human-AI collaboration to feature complex interactions and subtle defects. Experiments with mainstream VLMs show that all evaluated models perform poorly, revealing a clear gap between current VLM capabilities and the demands of real-world testing in AI web development. We release our code and data at https://github.com/SleepyWithoutCoffee/CATJudge.

Translation students need to learn both how to use translation technologies and how to judge the choices those technologies make available. This article presents LoopCAT, an Apache-2.0-licensed, local-first computer-assisted translation environment co-created with OpenAI Codex using GPT-5.5 and GPT-5.6, and proposes a framework connecting workflow competence, evaluative judgement, and technological agency. The account draws on repository history, implementation inspection, and the verification records of an identified development build. LoopCAT combines local project storage, translation memories, terminology, quality assurance, document exchange, and optional connections to local or hosted AI services. Its English, Catalan, and Turkish interface catalogs also make the application itself available as teaching material: students can translate English UI strings into another language, review the existing automatically generated target drafts, import their revisions, and test the interface. We organize these opportunities around four forms of participation: operating a workflow, evaluating outputs, inspecting and configuring mechanisms, and making or defending a bounded intervention. A six-session sequence, a UI-localization assignment, a placeholder example, and an assessment rubric specify how teachers could use the framework. The paper separates implemented capabilities from proposed educational benefits; it reports no new student-learning outcomes. It distinguishes the latest package checks from earlier regression evidence and sets out a protocol for classroom evaluation. LoopCAT provides an inspectable setting for teaching how translation decisions interact with data, interfaces, and software rules. Whether these activities improve judgement, transfer, or participation remains an empirical question.

For emerging scientific research domains, local Small Language Models (SLMs) are becoming more attractive, as they offer stronger privacy control and more stable deployment pipelines than Large Language Models. However, in practice, scientific question-answering on SLMs often operates under inevitable constraints: small literature collections, fragmented evidence, limited context window and reasoning abilities. We propose the Evidence-Grounded Typed Knowledge Graph (EGT-KG), a retrieval framework to improve information retrieval with local SLMs. We assessed three question-answering settings: a vanilla Retrieval-Augmented Generation (RAG) workflow and two EGT-KG workflows: an automatically generated relation schema (AS) and an expert-defined relation schema (ES). Our experiments were evaluated with a six-dimensional evaluation framework (S3CRF: Soundness, Correctness, Completeness, Conciseness, Relevance, Fluency) on a Biopolymer-bound Soil Composite literature benchmark, showing that EGT-KG outperforms the vanilla RAG method in most settings, with the best improvement from llama3:8b: a Final Score of 70.37 (+14.67%) and 68.82 (+12.14%) by AS/ES EGT-KG variants.

Command-line coding agents (e.g., Claude Code, Gemini CLI) can already read and write files and sustain long sessions, yet end-to-end research still fragments across chat tools, IDEs, terminals, and writing environments, and the decisions that make it auditable are rarely preserved. We present Dr. Claw, an open-source workspace that wraps existing coding-agent executors in a controllable and auditable human-in-the-loop workflow rather than introducing another autonomous agent. Persistent state objects, a reusable skill library, and multi-executor coordination link human decisions to AI execution, turning planning, execution, and writing into one traceable, recoverable loop. We demonstrate Dr. Claw through an interactive three-view scenario and a failure-recovery walkthrough, and evaluate it against a bare command-line agent sharing the same backend executor, so the comparison contrasts the whole orchestration layer (task graph, state objects, and skill library) with the agent it wraps. Holding the executor fixed, Dr. Claw scores higher on research completeness while persisting an auditable, recoverable process trail. Demo access: repository https://github.com/OpenLAIR/dr-claw, released under AGPL-3.0 with GPL-3.0 upstream components.

Giving an agent a file about a named expert can supply hard-to-find material, produce a recognizable persona, or change what the agent decides. These are different claims. We test each one. mimeo is an open-source tool that finds a person's public work, checks each extracted quotation against the cached source text, and writes a file an agent can load. Eight logged builds averaged 38 model calls; the check rejects 13.2% of extracted quotations. We tested four expert files with one coding-agent harness. Knowledge access was clearest: mimeo answered all 20 obscure, quotation-heavy questions; no closed-book condition answered more than 10. Keyword search (BM25) over the same pages answered 15-17, a gap this sample cannot resolve. Grounding showed one clear benefit: personas written from model memory misstated a documented position on 1-4 of 20 answers under every grader; the plain agent and mimeo never did. Every persona was easy to spot on short open prompts, and adding task material lowered identification by 18-23 points. mimeo was no more identifiable than a from-memory profile. Judgment transfer remained unresolved because both tests hit their ceiling: every condition found 94-97% of the problems planted in engineering tasks and scored 94-100% on 16 new application scenarios. An AI-judged "sounds like the expert" score changed with the judge: two of four preferred answers based on a model's stereotype, while two found no difference on the same text. That is a caution against relying on a single AI judge. The evidence supports mimeo as a compact, inspectable reference on a person, not as a demonstrated transfer of their judgment. Toolkit and expert profiles: https://github.com/K-Dense-AI/mimeo