Agent 进展

Scripted and rule-based non-player characters (NPCs) in combat video games often exhibit predictable behaviors that experienced players can exploit, while reinforcement learning (RL) agents typically retain a fixed policy after training and cannot readily adapt their strategy to different opponents. We investigate a runtime strategy-selection framework in which a large language model (LLM) guides a trained RL policy without modifying its underlying behavior. To demonstrate this, we train five NPC agents with a shared PPO policy in Unity and compare a baseline configuration, in which the policy acts independently, with an LLM-augmented configuration in which a locally hosted Mistral 7B model, accessed through Ollama, reads the live game state every five seconds and assigns one of four tactical tags. We evaluate both configurations against three scripted opponent types across 600 episodes and analyze outcomes using the Mann-Whitney U test. Against a Balanced opponent that changes tactics during an episode, the LLM-augmented agents more than doubled their win rate from 11% to 24% and produced significantly longer episodes. Against an Evasive opponent, the augmented agents achieved a higher win rate and faster kills, although their shorter episode duration did not satisfy the strict hypothesis definition. Against an Aggressive opponent, the LLM's near-constant preference for encirclement was counterproductive. Analysis of 2,430 strategy selections showed that Surround was selected in 83.8% of cases regardless of opponent type, indicating limited zero-shot strategic differentiation at this model scale. These results demonstrate both the potential and limitations of LLM-guided runtime strategy selection for adaptive multi-agent game AI.

This report presents a framework to help organisations, policymakers and researchers reason about the risks that emerge when AI agents interact with each other, how those risks change as interactions cross organisational boundaries, and the controls that may help address them. As organisations deploy AI agents, those agents will increasingly interact with each other: inside the organisation, with the agents of partners, customers and suppliers, and with unknown counterparties on the open internet. Failures can emerge from the interactions themselves, and once those interactions cross an organisation's perimeter, no single organisation can fully see, control or govern them. The report introduces three deployment tiers, defined by the minimum common governance binding any two interacting agents: singular governance, where one organisation governs every agent; federated governance, where multiple organisations deploy into a shared environment under agreed rules; and open environments, where agents operate with no central authority and shared standards are adopted voluntarily if at all. Within each tier, the report examines risk factors, failure modes and available controls. It identifies who is positioned to apply the controls, and where no actor is positioned to act, it characterises the gap and the collective action required to close it.

Additive manufacturing (AM) enables compact hydraulic components whose internal fluid channels can follow free-form 3D paths rather than conventionally drilled holes. A representative case is multi-group channel layout in rotary direct-drive servo valves: once a channel is placed it permanently occupies volume, so later channels must clear earlier geometry--unlike classical multi-agent pathfinding (MAPF), where agents free space after moving. We study this setting as multi-group pipe routing under permanent geometric occupancy. Our contributions are a problem formalization with geometric dual-witness conflicts, a constructive 3D benchmark (two corridor generators x two obstacle painters, controlled difficulty, feasibility witnesses), and baseline results for CBS, PBS, and priority planning adapted to this coupling. We evaluate by success rate under a fixed time budget. PlaneSlice hard instances clearly rank CBS above PBS and PP, while easier cells validate the pipeline. The goal is a reproducible problem definition, suite, and classical baselines--not a new optimal MAPF algorithm.

AI agents increasingly perform long-term reasoning, planning, tool use, memory integration, and autonomous decision making, yet erroneous intermediate states can propagate and cause inconsistent decisions and unreliable outputs. Existing reasoning approaches mainly rely on iterative planning, self-reflection, augmented memory, or verification, but rarely localize and selectively repair faulty reasoning. We present ORDDAR (Observation-Driven Reasoning for Distortion-Resilient Decision, Action, and Cognitive Recovery), a reasoning framework that models reasoning as cognitive state transitions, detects localized distortions, retrieves related reasoning from prior experiences, and repairs only the affected states. ORDDAR therefore performs recovery at the local reasoning-transition level rather than regenerating the complete trajectory. Experiments across mathematical, commonsense, multi-hop, and clinical reasoning benchmarks demonstrate improved reasoning quality, recovery ability, and interpretability over multiple evaluated reasoning baselines.

Multi-agent motion planning for high-degree-of-freedom robotics manipulators in shared workspaces remains a fundamental yet challenging problem. Centralized planners often suffer from poor scalability, while decentralized approaches face robustness and safety concerns. Game-theoretic formulations offer a promising approach for modeling agent interactions, potentially overcoming these limitations. However, their application to articulated multi-arm systems remains limited. This paper presents an iterative Linear Quadratic (LQ) game framework for multi-manipulator motion planning, where each manipulator is modeled as an independent agent optimizing its own objective while interacting with other agents based on shared global states and collision constraints. The method solves a series of local LQ games by linearizing the dynamics and approximating the cost around a nominal trajectory, with Riccati backward recursions yielding feedback Nash strategies. To address the challenges of articulated systems, we incorporate differentiable penalties for self-collision and inter-arm collision into the optimization pipeline, enabling coordinated, collision-aware trajectory generation. Experiments demonstrate that our framework produces smooth, safe, and efficient trajectories in high-dimensional settings, outperforming traditional methods. This highlights the effectiveness of differential game formulations for multi-robot manipulation.

Agentic AI has the potential to accelerate curation of biological databases and knowledge bases. However, uptake has been hindered by a number of challenges and obstacles, including access to agents and appropriate training. Here we describe how we have attempted to address and mitigate these challenges and obstacles through the deployment of a cloud-based agentic environment, and the development of an interactive training workshop for the Gene Ontology Consortium. Our cloud environment for agentic-assisted curation was based on the JupyterHub platform, and utilized Claude Code as a universal harness. This allows curators to interact with an agent session through a terminal running in the browser, and has additional benefits such as centralization of access through a single API gateway, removing the need for participants to manage subscriptions or install software locally. We created four training modules, walking participants through basic agentic tool use first and then working up to agentic biological pathway curation using the existing GO-CAM (GO Causal Activity Model) curation tool. Thirty-seven participants took part in the four-hour workshop. Our key takeaway from this workshop is that building community capability with agentic AI is primarily a problem of access, workflow design, and training. Removing technical barriers, introducing capabilities gradually, grounding exercises in familiar curation tasks, and giving curators direct experience evaluating agent output can provide a practical route toward building shared agentic AI capability in distributed scientific communities.

The hidden states of large language models (LLMs) are known to capture rich information relating to model knowledge and behavior that can be hard to extract from examination of input and output alone. As LLM-based systems increasingly interface with the external world, one area of concern is detecting incorrect or improper use of tools. Motivated by this, we study the effectiveness of using linear probes to detect incorrect tool-calls, measuring probe efficacy across 18 tool-calling LLMs evaluated on the Berkeley Function Calling Leaderboard. Overall, we find that probing is an effective means to catch a range of different tool-calling errors, including errors arising from using an argument that has the wrong value but the correct type, which might not be recorded by standard logging frameworks. Important factors in success include model size, probing layer, and model post-training type. We also show that probes are capable of generalizing to novel types of errors, which is critical in real world deployments.

In real-world software development, code review typically involves iterative interactions between developers and reviewers to improve software quality, making the process costly and time-consuming. Although recent work explores large language models (LLMs) for automated code review, most approaches oversimplify code review into a single-round, static decision task, which fails to capture the multi-round interactive nature and the complex problem-solving processes inherent in realistic review scenarios. To bridge this gap, we introduce MCR-Bench, the first defect state-aware benchmark designed for realistic multi-round code review. MCR-Bench covers five commonly-used programming languages and consists of 2,269 real-world multi-round code review tasks, each of which is annotated with fine-grained defect information and cross-round state labels. Each task in MCR-Bench is equipped with fine-grained defect metadata (e.g., description, type, severity) alongside dynamic state annotations, capturing the complete evolutionary trajectory of a defect throughout the multi-round process. We obtain several findings through extensive experiments on MCR-Bench with mainstream LLMs. (1) Limited overall capability: experiments reveal that mainstream LLMs exhibit limited overall performance in defect detection and defect lifecycle state tracking, with performance degrading significantly as the number of interaction rounds increases; (2) Defect-sensitive performance: LLMs' performance varies substantially across different defect types and severity levels, with semantically complex or low-salience defects being significantly more likely to be missed; (3) Underlying Failure Mechanisms: our in-depth error analysis dissects the distinct drivers of false positives and false negatives, revealing critical weaknesses such as cross-round temporal misalignment and inadequate long-range memory.

Long-horizon agent runs generate experience that can improve both the current run and future work. Most self-improvement methods process this experience only after execution ends, so they cannot redirect the active run or immediately apply and validate lessons learned from it. We argue that self-improvement should instead be live, using emerging experience both to redirect the active run and to update the persistent harness. Existing agent architectures do not fully support this goal. Single-agent self-correction combines task execution and trajectory assessment within one context, while subagent delegation separates execution but typically cannot redirect an active subagent. We present PILOT, a supervisor-worker harness for live self-improvement through two coupled mechanisms: (1) live steering lets a separate supervisor redirect or abort the active worker during execution; and (2) live self-evolution distils procedures and failure modes revealed during execution into reusable skills and memory. Across two frozen backbones and three benchmarks, PILOT ranks first in five of six configurations. On Terminal-Bench 2.0, PILOT outperforms counterpart harnesses by up to 9.8 percentage points. In the self-improvement setting, PILOT gains 14.6 points with GLM-5.1 and 12.4 points with Kimi-K2.6. Mean output tokens fall by 42.9% and 47.4%, while successful evaluations per million output tokens rise by 110.3% and 134.0%, respectively.

Recent advances in agentic large language models (LLMs) have enabled increasingly autonomous software engineering workflows, yet automatic machine learning (ML) paper-to-code reproduction remains a challenging long-horizon problem. Unlike conventional code generation, this task requires constructing and maintaining a fully functional repository whose state continuously evolves during execution. Existing systems typically rely on static upfront planning followed by sequential file-level generation, which often leads to inconsistencies as dependencies, interfaces, and execution feedback change over time. We propose DeepRepro, a state-aware framework for paper-to-code reproduction based on execution-state-aware subplanning. DeepRepro dynamically transforms evolving repository states and runtime feedback into fine-grained implementation subplans, keeping planning aligned with execution throughout repository construction. The framework further incorporates repository-aware orchestration and a lightweight process-aware interface for transparent monitoring of long-horizon reproduction. Experiments on PaperBench Code-Dev show that DeepRepro consistently outperforms strong scientific and commercial code-agent baselines.

Low-bit quantization offers a promising avenue for reducing the computational and memory demands of Multimodal Large Language Models (MLLMs). Recent hardware support for low-precision formats, ranging from MXFP8 to ultra-low-bit formats such as MXFP4 and HiF4, has accelerated research into efficient MLLM training and deployment. In this work, we present a systematic study of these quantization schemes in representative MLLMs that span both video generation and reasoning tasks. Our analysis shows that MXFP8 achieves near-lossless performance, whereas aggressive 4-bit quantization leads to significant degradation. Through extensive ablations, we identify activation quantization as the primary source of this performance loss, contributing substantially more than weight quantization. Motivated by this observation, we propose Residual Fallback Quantization (RFQ), a lightweight activation reconstruction framework that supplements the primary ulta-low-bit activation representation with an auxiliary quantized residual pathway. By explicitly modeling and compensating for quantization errors, RFQ improves activation fidelity while preserving the efficiency advantages of ultra-low-bit computation. RFQ requires no architectural modifications and incurs negligible computational overhead. Extensive experiments on Wan2.2 and Qwen3-VL demonstrate that RFQ consistently recovers a substantial portion of the performance lost under the quantization of MXFP4 and HiF4, significantly narrowing the gap to BF16 baselines across both generation and 4 reasoning benchmarks. Our findings establish activation quantization as the dominant bottleneck in ultra-low-bit MLLMs and highlight residual-based activation reconstruction as an effective and practical strategy for robust 4-bit deployment.

Reinforcement Learning with Verifiable Rewards (RLVR) and other RL-style post-training paradigms have been used for aligning large language models (LLMs) with reasoning standards. The resulting recent Reasoning Language Models (RLMs) such as DeepSeek-R1, o3, and Kimi k1.5 show that such RL-style post-training ("RL-for-LLMs") can substantially improve chain-of-thought reasoning, long-horizon planning, and self-correction. However, the computational footprint of these systems is massive: state-of-the-art RLM training requires millions of GPU-hours and tightly coupled multi-model pipelines that stress modern hardware far beyond classical supervised LLM training. This makes RLM training as much a parallel and distributed systems problem as an algorithmic one. In this work, to facilitate developing RLMs that are simultaneously high-performance, scalable, and cost-effective, we first systematize the RL-for-LLM paradigm and provide a compute-centric analysis of prominent post-training algorithmic frameworks: Proximal Policy Optimization (PPO), Group Relative Policy Optimization (GRPO), as well as their variants. Second, we develop a taxonomy of intra- and inter-model parallelism strategies for RL-for-LLMs, covering both traditional techniques (data, tensor, pipeline, sequence, context, and expert parallelism) as well as novel forms of parallelism and optimization techniques for multi-model RLM training, for example disaggregated placement, stage fusion, hybrid parallelism, and asynchronous execution. We harness the work-depth model of parallel computing to make our taxonomy and its insights rigorous and portable. Finally, we analyze existing RLM frameworks and we distill practical guidelines and outline open research directions for building scalable, fast, and cost-effective RLMs.

Electronic design automation (EDA) has advanced engineering productivity through successive generations of tooling that progressively automate synthesis, optimisation, and verification. Large language models (LLMs) extend this trajectory by enabling direct translation from design intent to hardware implementations. In most of the EDA literature, LLM-based solutions are typically assisting siloed design stages or tasks, however this obscured the drivers by which capability emerges and systems scale. In this Perspective, we instead define three hierarchical roles that reveal how capability accumulates: a Generator that produces design artifacts in a single pass, an Agent that refines outputs through iterative tool feedback, and an Orchestrator that coordinates decisions across EDA-stages. Across published systems, this reveals a syntax trap in which models are trained to produce plausible code rather than physically correct hardware, compounded by fragmented tools and loss of design context that obscure how decisions affect later stages. Comparisons across the three roles show that current approaches struggle to scale to industrial designs, motivating a shift towards a standardised, physics-aware orchestrator that connects tools and agents across the EDA flow for more reliable and accessible hardware design.

Effective human-robot collaboration in industrial settings requires robots to understand human intentions and assist with task planning, reducing workload. Recent works have explored the use of Multi-modal Large Language Models (MM-LLMs) for task planning in such data-scarce scenarios, leveraging in-context learning to interpret user actions and generate long-horizon action plans in natural language. However, MM-LLMs inherently lack an understanding of system states and do not track state transitions, often leading to hallucinated actions that deviate from the intended goal. Additionally, generating action plans in natural language tends to limit the generated plans to a high level, introducing ambiguity in action execution. To address these limitations, we propose the State-aware Task Estimator and Planner (STEP), which prompts a MM-LLM to explicitly estimate the state of the system and predict the state transitions resulting from executed actions. By forecasting future states alongside actions, STEP ensures task-convergent planning while also providing additional assistance parameters necessary for executing the predicted actions. We evaluate STEP in a simulated environment using a robot assembly task. Our approach outperforms the state-of-the-art by 32.8% in action executability and 14.8% in final-state error.

Tool-augmented LLM agents must rely on untrusted runtime Observations to complete open-ended tasks; however, when tool outputs no longer merely provide data but begin to specify concrete actions, they effectively become ``commands'' that can drive real-world side effects beyond user intent. We argue that this risk arises from conflating action induction with execution authorization. To address this distinction, we propose SARA, which treats action induction and execution authorization as distinct runtime roles and separates action provenance from execution authority. On the Observation side, a context-isolated Action Probe exposes action-inducing semantics and persistently records action-origin provenance across steps as a review signal; on the execution side, actual tool calls are authorized only against the user objective and audited evidence from authorized successful executions, while satisfying goal, execution-chain, and argument-level support. To preserve this separation across multi-step execution, SARA applies No-History-Promotion to prevent historical recurrence from laundering action origins into execution authority. Across AgentDojo and AgentDyn, SARA limits ASR to no more than \(0.63%\) across four primary evaluation settings while maintaining competitive task utility, and consistently reduces ASR across additional Agent backbones.

LLM agents increasingly rely on generated interaction data to learn how to interact with external environments. Agentic data generation must maintain consistency among environments, tasks, interactions, and success signals while producing experience that is useful rather than merely abundant. Existing work spans many agent domains, but domain-centered organization and heterogeneous evaluation often obscure common generation mechanisms and conflate candidate construction with verification and selection. This work develops a two-level framework for the field. First, we represent agentic data as a common factorized object $(E,q,τ,v)$, comprising an environment specification, task signal, interaction realization, and optional verifier. We organize generation paradigms by their primary anchor and dependency structure. Second, we formulate generation as constrained distribution design through the Accuracy-Complexity-divErsity (ACE) lens. Accuracy establishes the feasible support of grounded and internally consistent data. Within this support, Complexity places learning mass relative to the capability of a declared learner and execution configuration, while divErsity controls coverage and redundancy of data. Using this framework, we explore how prior work verifies generated experience, constructs and calibrates difficulty, and expands behavioral coverage. The literature reveals a shift toward execution-grounded accuracy, learner-relative complexity, and diversity beyond surface variation or dataset size. We further discuss broader directions and emerging trends in agentic data generation through the ACE lens, including their implications for scaling, data sources, training regimes and adaptive learning. Overall, the central challenge is not simply to generate more data, but to continually allocate valid, informative, and non-redundant experience as agents and environments evolve.

Multimodal large language models (MLLMs) are increasingly capable scientific assistants, yet they remain far from fully autonomous research. This transition requires models to actively inspect academic papers, build global evidence views, and make traceable judgments without prespecified issues or evidence. However, existing work provides limited task paradigms or training studies for such issue- and evidence-absent verification. We study this challenge through scientific error detection, where models must determine whether errors exist and justify them with evidence-based reasoning. To fill this gap, we present VERA-RL, a reinforcement-learning formulation for scientific error detection over academic papers. Following a Reason--Verify--Scan progression, we construct VERA-13K, a 12,900-sample dataset organized into 4,300 matched chains, covering 6 scientific-error categories across the research workflow and broad natural-science domains. We further introduce fine-grained rewards for reasoning completeness, evidence alignment, and error precision. Training Qwen3-VL-8B with VERA-RL substantially improves verifiable reasoning, approaching flagship MLLMs such as Gemini 3 Pro and Qwen3-VL-235B-A22B on Scan.

As electricity market participants increasingly adopt learning-based agents for their bidding strategies, electricity markets are becoming algorithmic. Evidence from algorithmic markets in other domains shows that tacit collusion can arise purely through independent learning. Moreover, electricity markets are typically oligopolistic and feature repeated interaction among a small number of participants, making them structurally susceptible to non-competitive behavior. In the face of these observations, this paper investigates the hypothesis that tacit collusion may emerge in electricity markets where participants' actions are controlled by autonomous learning-based algorithms. We model strategic bidding as a repeated game with imperfect public monitoring, and model the participants' emergent behavior using multi-agent reinforcement learning. We propose a multi-dimensional set of criteria (going beyond profit comparisons against Nash equilibria) to assess whether the resulting behavior constitutes tacit collusion. Our experimental results showcase that such a danger is realistic for electricity markets: there are cases where agents do learn to sustain supra-competitive outcomes that are supportive of tacit collusion indicators, even though the agents were never instructed to collude.

Organizing long-term memory for multimodal agents remains challenging because existing methods either suffer from expensive question-agnostic offline summaries or naive embedding similarity matching that introduces incomplete and redundant context. To address these issues, we propose GraphMemix, a combinatorial-optimization graph memory framework that models memory organization as query-aware evidence-forest construction. Specifically, our method consists of three key components:(1) candidate graph construction, which expands multi-view seed memories through schema and semantic relations to acquire query-aware original context; (2) evidence utility and activation costs, which decouples direct memory support from anchor-conditioned relation verification to suppress redundant or conflicting information; and (3) forest optimization, which jointly selects a forest-format memory context under a maximum evidence budget and its reliable relational structure. By organizing memory into a query-relevant subgraph, the method avoids substantial lifecycle cost and recovers low-similarity complementary evidence. Experimental results across four long-term multimodal memory benchmarks demonstrate significant improvements with different foundation models and establish a new Pareto frontier between accuracy and lifecycle cost.

LLM-based agents are increasingly deployed in product-level execution harnesses, where jailbreaks can trigger harmful tool use and persistent state changes, creating greater risks than unsafe text generation alone. Existing automatic red-teaming methods often rely on fixed attacks, while recent agentic attackers coordinate multiple jailbreak tools and show stronger potential through trajectory-based retrieval. However, such retrieval can reuse misleading experiences due to retrieval bias and unclear tool credit, and full trajectories add context overhead while reducing interpretability. We propose RedEvoAgent, a black-box red-teaming agent that distills cross-case attack trajectories into a concise, human-readable attack skill. The attack skill adaptively evolves through tool-effectiveness profiling and Deciding-Tool Attribution for skill updates, and a validation ratchet that retains only updates improving validation performance. Experiments on multiple benchmarks, target models, and target execution harnesses show that RedEvoAgent outperforms fixed and agentic baselines, improves tool efficiency, and transfers across attacker models and target execution harnesses.

Travel planning agents assist users in generating personalized travel plans by modeling their individual preferences. Existing agents either rely on explicit user instructions or engage in multi-turn clarification to elicit user preferences. However, both approaches overlook the rich behavioral signals latent in users' past behaviors, which implicitly encode their preferences. This over-reliance on active user input increases interaction burden and limits plan personalization. To bridge this gap, we introduce a new task, Behavior-Aware Travel Planning, which infers user preferences directly from past behaviors and generates personalized travel plans. To facilitate research on this task, we introduce Behavior2Trip, a benchmark constructed from one of the largest Chinese online travel platforms, comprising 11,400 instances. Each instance represents an average of 39.8 past user behaviors spanning 14 attributes across 5 preference dimensions. We further propose B2T-Agent, a reinforcement learning-based agent that leverages user behavior trajectories, interacts with external tools for preference-aligned retrieval, and maintains an internal memory module. Experiments on Behavior2Trip show that GPT-4.1 achieves a full-constraint pass rate of only 0.5% on the hardest tasks, while B2T-Agent built upon Qwen3-8B outperforms all baselines, highlighting the substantial challenge of this task. Moreover, Qwen3-8B trained with B2T-Agent also outperforms GPT-4.1 on the TravelPlanner benchmark, demonstrating strong generalization. Code and data are available at https://github.com/BUAA-IRIP-LLM/Behavior2Trip

Recent advances in SLAM have leveraged 3DGS for photorealistic reconstruction and novel view synthesis. However, most methods rely on RGB-D input, which is unavailable on consumer-grade smartphones, and few integrate 3DGS within a collaborative framework. Therefore, we present CGS-SLAM, a hybrid decentralized/centralized system enabling multi-agent 3DGS SLAM using only RGB and inertial data. Each agent performs local tracking with inertial data as a motion prior and reconstructs a scaled map using a metric monocular depth estimator (Depth Pro). Keyframe encodings are shared among agents, enabling dynamic keyframing in regions of spatial overlaps with other agents, enhancing submap alignment. Afterwards, a central server aligns submaps using VGGT as a view alignment model. This bidirectional communication keeps communication cost low during mapping and global reconstruction in difficult GNSS-denied environments. Experiments on multiple datasets demonstrate competitive tracking performance, improved rendering quality over state-of-the-art methods, and accurate submap alignment.

LLM-based agents are increasingly proposed for network fault diagnosis, but existing benchmarks evaluate them only on accurate tickets and always assume a fault is present, conditions rarely met in practice. We present FaulT-Bench, a benchmark of 200 troubleshooting scenarios across eight network topologies, five reimplemented from public practitioner labs, spanning genuine faults, false fault reports, incorrect device attribution, and incorrect root-cause claims. To isolate how ticket wording affects diagnosis, we further rewrite 72 false-premise tickets into five reporter personas that vary reporter confidence and verifiable detail one factor at a time, holding the network state fixed. Our automated harness deploys each scenario in Kathará, lets agents interact through the NIKA tool interface, and scores free-text diagnoses with an LLM judge across outcome, fix, and reasoning quality. Evaluating SADE, ReAct, and Claude Code, we find all three are near-saturated on accurate tickets and robust to misdirection, yet degrade sharply when the network is healthy and the ticket is wrong, probing until a benign condition can be promoted to a root cause rather than concluding nothing is wrong. Persona rewrites show that how a ticket is written matters more than what it claims: a confidently wrong report is handled about as well as an accurate one, while a vague, underspecified report degrades performance sharply. The three agents also fail differently, from constant over-diagnosis to unanswered runs, at very different cost. These results position FaulT-Bench as a benchmark for developing agentic systems that can reason reliably over the noisy, unreliable tickets of real-world network troubleshooting.

Multi-agent large language model systems are widely reported to beat single-model baselines, but the evidence is mixed, and comparisons are usually confounded: pipelines change token budgets, tool calls, and prompts simultaneously, so an aggregate gain rarely reveals what actually helped. We investigate the effect of introducing the manager-worker scaffold over a shared filesystem workspace, with no training and no per-benchmark tuning, measured against the same model answering in a single pass. Across nine models — five open-weight, spanning 9B to ~2.8T parameters, and four frontier closed models — on the 100 latest hard LiveCodeBench problems, the scaffold's benefit is real but conditional: large and statistically significant for some (Qwen3.8-27B +23.4, GPT-5.6-Luna +10.6 and GPT-5.6-Terra +8.0, each over five paired passes; Kimi-K3 +30.4 and Minimax-M3 +11.0 over five paired passes with reasoning off, both at $p < 10^{-4}$, and +42 and +12 in a single pass at a 128k cap) and null or negative for others (Qwen3.6-35B -1 to -9 with reasoning off). With the manager, Opus-5 achieves the highest score in the study at 91% in one pass. Running a manager roughly triples the token bill, but it buys accuracy more cheaply than moving to a larger model does: GPT-5.6-Terra with a manager nearly matches Fable 5's single-call accuracy (85.0 against 87.4, $p = 0.59$) at a fifth of the price (\$11.71 against \$61.11 per 100-problem pass, $p < 10^{-4}$), and the Qwen-27B arm does it for \$51.75 on weights anyone can self-host. Our transcript analysis finds several mechanisms behind the gains, of which two recur: context management, in which short worker calls and shared notes organize state and reduce truncation, and problem decomposition. Improvements are modest for large models with reasoning enabled, but larger for some models with reasoning disabled and for smaller models with reasoning enabled.

Cooperative multi-agent reinforcement learning (MARL) faces significant challenges in maintaining robust coordination under noisy observations. Although observation disturbances are often introduced independently across agents, their downstream effects on cooperative decision-making can become structured through underlying cooperation structures. We characterize this phenomenon as structured noise effects, where noise-induced decision effects exhibit local correlation among agents with stronger task-related dependencies while remaining globally heterogeneous across different agents and local structures. Existing robust MARL methods, however, rarely explicitly characterize or exploit such structure-dependent noise effects. To address this limitation, we propose SIGMA, a hierarchical collaboration framework that exploits cooperation structures to learn robust representations under noisy observations. SIGMA first organizes agents into adaptive local structures through density-based grouping and performs intra-group consensus aggregation to preserve shared task-relevant information while smoothing agent-specific representation deviations. Inter-group attention then adaptively integrates information across different groups to preserve global coordination while accommodating their heterogeneous contributions. Experiments on noisy-observation tasks in StarCraft II empirically validate the structured noise effects and demonstrate that SIGMA consistently improves robustness under observation noise while maintaining competitive performance in noise-free environments.

Large language model (LLM) agents increasingly operate over untrusted webpages, documents, tools, and persistent states while exercising authority over security-sensitive resources. Existing defenses typically protect either planning or individual tool interactions, but persistent agents face a broader threat: attacker-controlled data can alter control flow, enter security-sensitive tool arguments, or compromise later queries. We present SPA, a plan-first architecture that secures planning, execution, and cross-query state reuse. SPA invokes the planner once per query to generate a complete executable plan in a declarative domain-specific language, then applies dual-lattice information-flow control to track confidentiality and integrity across explicit data flows and control dependencies. To support persistence without re-exposing untrusted payloads to the planner, SPA stores execution results as labeled artifacts and reveals only semantic metadata during later planning. We evaluate SPA on AgentDojo and AgentDojo-MQ, which is our multi-query extension for measuring secure state reuse and delayed attacks. Under the 'tool_knowledge' attack, SPA with information-flow control reduces attack success to zero on AgentDojo and 0.2% on AgentDojo-MQ. Our results show that plan-first execution combined with label-preserving persistence can substantially strengthen persistent LLM agents, while revealing an important security-utility tradeoff introduced by strict integrity enforcement.

Large language model agents reason, call tools, and act autonomously over many steps, but their agentic skills-correctly sequencing tools, planning under dependencies, judging untrusted inputs, and grounding generated arguments-are hard to measure with accuracy-only leaderboards. We present BekchiAI, which addresses both sides: a benchmark for measuring agentic skill and a platform for observing and controlling live agents. The BekchiAI-Benchmark, a suite of 13 tool-using ReAct agents across 7 task categories (arithmetic, structured/SQL, security detection, URL grounding, planning, orchestration, and tool-policy), totalling 2,057 deterministic, committed test tasks. Every task is verifier-checkable gold answers are computed by running canonical SQL against a real database, computing the exact schedule of a directed acyclic graph (DAG), or evaluating closed-form lambdas including adversarial security samples paired with deliberately imperfect signature scanners so a score reflects the model's own judgment, not the copying of an oracle. We define a small set of behavioral metrics beyond accuracy-tool-call adherence, URL hallucination and source-match, and per-model token cost and report a four-model comparison (Qwen3.7-Max, gemma-4-31B-it, gemma4:26b, gpt-oss-120b) whose story is in the per-family spread, not the aggregate. The benchmark runs are executed using the provided evaluation scripts. BekchiAI-Platform is a complementary web-based observability and control layer for deployed agents, providing full token and latency telemetry as well as remote run termination. The benchmark, evaluation tools, and platform are publicly released.

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

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

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