2026-11-04

Recent multimodal agents have shown promising progress in embodied planning and web interaction. However, existing evaluations often rely on simplified tasks with limited combinatorial structure, leaving it unclear whether current agents can robustly integrate perception and reasoning for multi-goal planning under temporal, physical, and multimodal constraints. We introduce DualWorldBench, a unified benchmark for evaluating agent planning under controlled symbolic--temporal and grounded--topological complexity. Built around a shared pickup--delivery interface, DualWorldBench consists of three complementary components: TextWorld, which isolates symbolic planning under varying task compositions and release dynamics; SimWorld, which evaluates grounded path planning under controlled topological variation with text, image, or multimodal inputs; and DualWorld-VQA, a diagnostic suite spanning perception, scene understanding, and planning. Experiments reveal a common gap between local understanding and globally consistent planning. In TextWorld, agents struggle to derive optimal task sequences under interacting temporal constraints; in SimWorld, grounded observations are not reliably converted into feasible action sequences. DualWorld-VQA further confirms that perception and topology understanding do not necessarily translate into constrained planning. Together, these results identify the central challenge as composing symbolic or grounded evidence into globally feasible and efficient multi-goal plans. These findings establish DualWorldBench as a challenging testbed for agents with tighter perception--reasoning--planning integration. All datasets and code are publicly available at our project website https://dualworld.netlify.app/.

Multi-agent LLM systems assign different responsibilities to different agents: one plans, another retrieves evidence, another verifies, another executes. This division of labor is only meaningful if each agent stays within the functions it is authorized to perform. Over many turns, however, an agent can produce a locally reasonable message that exercises authority belonging to another role: a verifier that begins planning, a retriever that asserts unsupported conclusions, an executor that changes state no one approved. We formalize this as role drift, a turn-level breach of an agent's role contract. We introduce Agent Trajectory Control (ATC), which makes each role's authority explicit as a contract and enforces it at the point where a message would enter the shared conversation. ATC scores a proposed output for out-of-role deviation, leakage into the functions of other roles, impermissible speech acts, and logical contract breaches, then computes a leakage-regularized target that constrains regeneration; the revised message is re-checked before admission. Enforcement acts on the message rather than the generator, so ATC governs agents whose internal representations are unavailable. Across homogeneous and heterogeneous teams, ATC improves task performance and role consistency while lowering role-drift energy relative to strong baselines spanning prompting, memory, workflow control, and activation editing, with the largest gains under explicit role-confusion stress tests and in larger, more diverse teams.

Real-world geospatial decision support requires more than single-step retrieval or code generation. Disaster-response questions often involve multi-source workflows that combine spatial databases, map-derived features, satellite imagery, raster products, and reusable geospatial artifacts, where correctness depends on preserving spatial scope, temporal context, and intermediate execution state. While large language models and coding agents offer a promising interface for such workflows, they struggle with geospatial-specific failure modes, including handling large artifacts, selecting the wrong evidence source, losing spatial or temporal metadata, and applying invalid geometric operations. We introduce GeoFaham, a domain-specialized multi-agent system for geospatial disaster analytics. GeoFaham decomposes workflows into specialist agents for vector operations, map feature retrieval, imagery discovery, and raster analysis, coordinated by an Orchestrator with a dual-ledger control loop for dependency-aware planning, progress tracking, validation, and replanning. To support artifact-centric collaboration, GeoFaham uses a Response Store that persists full geospatial outputs and typed metadata outside the LLM context window while agents exchange compact identifiers and summaries. We also introduce GeoFaham-Bench, a 40-task benchmark grounded in 17 real-world disaster records, designed to evaluate multi-step geospatial reasoning, cross-source integration, raster-vector computation, artifact generation, and recovery under non-rigid ground truth. We evaluate GeoFaham on GeoFaham-Bench against strong general-purpose coding-agent baselines, including Codex, Claude Code, and OpenCode. GeoFaham achieves the best overall performance, scoring 0.940 versus the next-best 0.879, while using 7.5 times fewer tokens. Our code and benchmark are publicly available at https://github.com/microsoft/GeoFaham.

Roles in multi-agent large language model (LLM) systems allocate decisions as well as work. In a Planner--Executor--Reviewer team, the Reviewer's approval can determine when the team submits an answer and receives corrective feedback. We study both functions by varying role instructions while holding agents, topology, call budget, and ground-truth evaluator constant, using three open-weight models on Omni-MATH and two on SciBench. Removing roles largely eliminates functional specialization without lowering first-pass accuracy. After a failed first answer, however, submission decisions matter: a role-structured Reviewer often withholds approval, so its team submits less often and receives fewer evaluator verdicts. In an exploratory comparison with feedback available to both teams, the role-free team recovers more often across all five model--benchmark settings. Two further analyses support the importance of feedback opportunities. In a post hoc comparison at equal numbers of verdicts, the pooled recovery difference is small (role-structured minus role-free: $+0.009$, 95% interval $[-0.014,+0.033]$). In an exploratory intervention on $400$ Omni-MATH problems with gemma-4, keeping the role prompts but forcing submission in every cycle raises recovery from $0.198$ to $0.613$; the paired comparison detects no difference from the role-free team. The takeaway is to evaluate roles through recovery after failure as well as specialization. For systems with reliable external feedback, this requires measuring who may submit, how often the team receives feedback, and which agents can act on it.

Enterprises increasingly want LLM agents to automate standard operating procedures over their software stacks -- mapping each step of a procedure to the right call among thousands of internal APIs, without hand-authored step-to-API mappings. The core difficulty is API selection at scale: choosing and composing the correct tools over a long horizon, judged only by whether the overall task succeeds. Current tool-use benchmarks do not test this regime -- they expose only a handful of tools, evaluate single-step invocation, or grade multi-step behavior with subjective LLM-as-a-judge scores that penalize valid alternative trajectories. We introduce SLATE (Synthetic Large-scale API Toolkit for E-commerce), a large-scale benchmark that tests large-toolset, long-horizon API selection in a realistic enterprise domain. SLATE couples (i) queries decomposed into hierarchical, tool-grounded procedures over a ~1,000-tool e-commerce library, (ii) a deterministic, context-aware simulator that returns recorded outputs for matching calls and default responses otherwise, and (iii) an end-to-end task-success metric -- a task succeeds when it reaches every reference outcome -- that credits diverse-but-valid trajectories while remaining fully automatic and reproducible. Experiments with six open decision models and the language models they are built on, from 3B to 12B parameters, show that even the best model completes only 5.0% of the tasks, establishing SLATE as a challenging testbed for reliable enterprise tool use. We release SLATE, its simulator, and the evaluation harness at https://geshijoker.github.io/slate/

Multimodal large language models (MLLMs) increasingly address knowledge-intensive tasks by integrating evidence scattered across multiple document pages into a final answer. Because such answers are synthesized from many pages, the link between the answer and its supporting visual evidence can become obscured, hindering verification and trust. Cross-page visual attribution is therefore essential: it requires identifying the supporting pages and localizing fine-grained evidence regions within them, thereby making multi-page reasoning transparent and auditable. Existing approaches either process multiple pages jointly or iteratively refine evidence within a single trajectory, risking cross-page context pollution and, for interactive methods, substantial sequential overhead. We introduce \textbf{AttriSwarm}, an agent-swarm framework that assigns these concerns to separate roles: a leader handles global cross-page reasoning, while subagents perform page-wise attribution in isolation and in parallel. To support training and evaluation, we construct \textbf{AttriDoc}, a cross-page QA dataset linking supporting pages to their annotated evidence regions, and use it for role-specific training of AttriSwarm, improving answer accuracy, page-level $F_1$, and region-level IoU over its training-free counterpart by an average of 18.94, 29.34, and 23.04 percentage points, respectively, across three open-source backbones and three datasets.

Computational drug discovery offers a growing range of tools for proposing and evaluating molecules, but combining them and deciding what to do next still requires substantial expert effort. Here, we introduce El Agente F\'armaco, an AI scientist that uses modern large language model (LLM) technology to turn discovery goals expressed in everyday language into complex, scalable computational screening workflows. The agentic harness selects scientific tools, reviews their results, and adjusts subsequent steps to the target and the user's objectives. Our approach centers on a shared scientific memory: F\'armaco Atlas, which stores molecules, structures, and results together with their provenance -- a record of how they were obtained. The agent uses targeted database queries to retrieve the evidence needed for each decision. A single request can launch calculations for an entire collection of molecules, without requiring a separate LLM decision for every compound. This design limits the information the model must process as screening collections grow, while keeping the underlying results available for inspection and further analysis. The integrated tools support molecule generation, prediction of binding and molecular properties, and searches for compounds available for purchase. In a retrospective comparison with an expert-led campaign, the agent proposed chemically distinct candidates with competitive docking scores and higher calculated drug-likeness under explicitly imposed selection criteria. In a 22-task benchmark against plain tool calls and Codex, the agent achieved the highest observed mean model-judged score, with lower median cumulative token use and peak context. F\'armaco provides a foundation for flexible, scalable drug discovery workflows that connect scientific reasoning with an enduring, traceable record of computational work.

Students can use large language models to complete assignments, but a correct solution does not establish whether they understand the reasoning behind it. Individualized interviews can probe that understanding but require recurring staff effort. We introduce ProofPaths, an AI agentic framework that combines rubric-based grading with individualized and interactive interviews. Multiple grading agents assess a submission against instructor guidance, exchange critiques, and revise their judgments. Their scores determine which review supplies the grade and feedback. An interview agent asks students to justify steps in their submission and uses their answers to guide follow-up questions. Each question is linked to a retrieved submission excerpt, so instructors can compare the student's explanation with the cited passage. The system tracks assessment progress and stores grading reviews and interview transcripts for each submission. Several deterministic tests of the assessment controllers check source handling, score aggregation, and limits on correction attempts. Instructors can inspect all reviews and transcripts and override grades within the assignment workflow.

Large Language Models (LLMs) serve both as tools for aiding human decision-making in everyday life and as models of human agents in experimental simulation studies. Alignment on prototypical social decisions -- such as whether to cooperate on a collective action -- therefore has immediate practical consequences for researchers and for society at large. Here we systematically explore the impact of ``norm utility'' -- the non-material components of utility that account for an individual's desire to conform to social norms and to be internally consistent -- on the strategic decision-making of three open-weight LLMs in one-shot, two-player social dilemmas. We find that when a social dilemma is presented together with information about community beliefs -- but without any indication of how those beliefs should matter -- models tend to treat community belief as strategic information, so that in the Snowdrift game cooperation falls as community belief in cooperation increases. When we explicitly prompt the components of norm utility this relationship inverts, and cooperation increases with community belief, in broad agreement with the predictions of a simple norm utility model. We then show that prompting norm utility reduces the volatility of strategic decisions with respect to changes in the underlying social context, and generates more consistent reasoning habits across models. Taken together, our results suggest that prompting norm utility can help generate more human-like and consistent behavior across models and lower volatility within models.

Recent advances in large language models have enabled agentic systems that automate machine learning pipeline construction from minimal user input. However, existing agentic AutoML methods either depend on rigid templates that limit adaptability, target narrow task domains, or rely on free-form code generation that lacks reliability guarantees, leaving intermediate errors to propagate silently and degrade the final model. We present EvoML, an agentic framework for end-to-end ML pipeline synthesis built on expert-grounded skills that evolve across runs, executed through a validation-gated harness. The framework decomposes pipeline construction into four phases coordinated by three agents: an Orchestrator that enforces validation gates, advancing each phase only when its outputs pass structural, execution, and semantic checks, which precludes error propagation and keeps every artifact auditable; an Executor that runs expert-grounded skills, grounded in the nnU-Net workflow and pairing reusable templates with online adaptation; and a Curator that performs post-run evolution, consolidating the scripts and guidance recovered during a run into the skill library so it improves across datasets. We evaluate EvoML on eight datasets spanning six sophisticated, real-world ML task types, against agentic AutoML systems and command-line coding tools. EvoML achieves full automation and the best final model quality on all eight, spanning segmentation (up to 0.94 Dice), classification (0.99 accuracy), reconstruction (0.95 SSIM), and denoising (0.79 normalized PSNR), while transferring to broader ML tasks.

Long-horizon multi-agent systems suffer from three tightly coupled failures: task drift, role drift, and state staleness. We test the hypothesis that these share a common representational cause: agent hidden states move away from the directions encoding the active task, assigned role, and current shared state. A pilot study over 14,832 annotated turns supports this: drifted turns collapse near-uniformly across all three anchor subspaces (pairwise r = 0.72-0.78). Existing methods address these failures separately, operating at the surface-text level or treating control directions independently. We introduce ATLAS (Algebraic Task-role Locking with Adaptive Spectral Memory), a training-free intervention acting on hidden states before a token is generated. ATLAS builds a positive-semidefinite anchor kernel over task, role, and state subspaces and steers hidden states toward their shared high-coherence eigendirections; a low-rank Gram decomposition keeps this tractable. Across long-horizon memory, state maintenance, tool use, least-privilege coordination, and embodied stress tests, ATLAS outperforms prompting, retrieval, memory, and activation-steering baselines. Against A-MEM it improves memory macro-average accuracy by up to 5.7 points and reduces write/read failures by up to 31.7%/24.8%; against MemoryArena-Agent it gains up to 4.8 points. It gains 10.7-12.1 points under perturbation, reaches 93.5% role consistency, and transfers to a 72B backbone (macro-average 61.0 vs. 56.2). Ablating the joint task-role-state span costs 7.2 points against 0.8 for a matched-rank random span, indicating these directions are functionally implicated, not incidental; supplying them by prompt or retrieval recovers 4.3 points less. ATLAS is therefore a representation-level controller for open-weight multi-agent systems, complementary to output-level protocols for black-box agents.

Visual agents need to coordinate reasoning and action while their environment evolves during inference. We introduce LOLPlay, a benchmark in native \emph{League of Legends} with 36 champion-specific tasks spanning resource management, combat, cooperation, and objectives. Agents interact through screenshots and mouse--keyboard actions; automated initialization and 82 executable subgoal checks distinguish partial progress from completion. Across 324 task evaluations of three models at three reasoning-effort settings, GPT-6 Astra achieves the highest task success rate of 25.0\% at low and medium effort, and the highest subgoal success rate of 40.3\% at medium effort. However, high effort improves neither metric over low effort for any model. For Astra, longer requests leave fewer interaction rounds, suggesting a cost to additional reasoning in real-time control. Trajectories reveal incomplete subgoal composition and actions executing after the observed scene changes, motivating agents that combine visual reasoning with timely action.

Deep Research (DR) systems extend large language models into autonomous agents that decompose complex information needs, iteratively search the web, and synthesize evidence to answer challenging questions. However, their multi-stage execution may introduce robustness risks, as variations in user questions or intermediate search inputs may propagate across research stages and lead to incorrect answers. Existing evaluations focus on task-solving capabilities or robustness against external evidence disturbances, leaving intrinsic weaknesses in DR execution insufficiently explored. Thus, we conduct an empirical study of 1,932 execution trajectories from three representative DR systems and identify nine recurring defect categories across subquestion formulation, online search, and final reasoning. Guided by these findings, we propose EMBERS, an empirically motivated, defect-guided fuzzing framework for robustness testing of DR systems. Starting from correctly answered seed questions, EMBERS generates stage-targeted test inputs using defect-guided mutation, web-search-assisted fidelity assurance, and diversity-aware seed maintenance. Experiments across 16 framework--model--dataset settings show that EMBERS achieves 3.21 times the failure-discovery throughput of the strongest baseline under identical testing budgets. Human and LLM-assisted evaluations demonstrate that most generated inputs preserve the original information need and necessary contextual constraints. Moreover, model fine-tuning experiments show consistent-answer performance improvement, highlighting their value for robustness testing and model improvement.

Web navigation agents built on open-weight language models fail often and fail late, with most of the tokens spent after the trajectory has become unrecoverable. Recent work has shown that early termination is a viable cost-reduction strategy, but existing approaches produce opaque termination decisions that do not support diagnosis or human oversight. In this work we apply closed sequential pattern mining (BIDE+) to K-step prefixes of agent execution traces to produce a small, human-reviewable failure-signature library that organizes into a four-category failure taxonomy with associated repair recommendations. We use a multi-variant evaluation protocol that explicitly removes contamination artifacts (zero-step crashes and trivially separable single-step successes) so that reported detection accuracy reflects behavioral discrimination rather than trace-length distributions. We evaluate our methodology on 1,544 MiniWoB++ traces collected from three open-weight language models, achieving the best F1 at K=3 on the Exclude-errors configuration across eight methods (0.717), recovering 88% of test-set failures within three steps, and supporting an early-termination policy that saves 26.3% of tokens at 92% precision.

AI agents can make incorrect decisions or waste resources even when their tools and infrastructure work normally. Their nondeterministic behavior makes these problems difficult to reproduce and diagnose. We developed Signals to identify such problems from agentic traces and suggest concrete remediations. We built the method around a taxonomy of 45 problem types across eight families, with definitions and evidence requirements for each type. The method distinguishes realized failures, risk exposures, and inefficiencies, groups related findings into engineering issues called Signals, and assesses their severity from the supporting evidence. Each Signal includes actions a developer can take and checks for evaluating the change. Further executions can add evidence to an existing Signal or reveal a new problem. We evaluated the method on public agent executions and achieved 91.43% F1 when matching Signals to reference issues, about 18 percentage points higher than a one-pass LLM control that used the same model and taxonomy.

The deployment of AI agents in multi-agent settings enables inter-agent influence, whereby agents can strategically steer each other's behaviour towards an intended outcome. Such capabilities pose novel risks, as malicious agents could direct others towards harmful actions. In this work we investigate three such capabilities - persuasion, deception, and coercion - across five realistic evaluation environments and a variety of agents instantiated with frontier language models. We observed significant inter-agent influence capabilities; in oversight environments, all tested models shifted overseer decisions from rejection to approval through persuasion, deception, or coercion. We also observed successful influence in non-hierarchical 'peer-to-peer' settings through redirection of a peer's safety research trajectory. As multi-agent deployments become increasingly commonplace, these results point to a pressing need for work on mitigating the risks from inter-agent influence. Code and data are available at https://github.com/Cooperative-AI-Foundation/inter-agent-influence.

LLM agents often give conflicting answers to the same question, and many deployed agents are variants of one base model, so their errors are correlated. We study truth discovery among such agents: inferring the true answer and each agent's reliability without labels. Classical truth-discovery and crowdsourcing estimators reward an agreeing bloc and can fall below majority vote. We propose BP-TD, which adds per-agent decorrelation weights to mean-field variational inference on a one-coin Dawid-Skene model, so that agents sharing a base model count together as about one vote. Its first iteration is exactly a weighted vote; later iterations estimate agent reliability. With 13 local agents on MMLU, BP-TD beats majority vote without any labels (0.716 vs. 0.624 on 160 questions; 0.707 vs. 0.674 on 10,000) and ranks agents by their true accuracy, which classical estimators invert. The gain depends on the pool: over 2,852 sub-pools of four benchmarks, decorrelation helps when the largest correlated bloc is weaker than the other agents and hurts when it is the strongest, and we found no label-free statistic that tells the two cases apart. Labels change the picture quickly: with 20 labelled questions, selecting the single best agent is more accurate than every aggregator we tested on MMLU, ARC and TruthfulQA.