NeurIPS 2026 论文列表

Evaluating large language model (LLM) agents in multi-turn interactive environments is expensive and risky, as it requires online environment interaction. We propose \textbf{ADWM} (Autoregressive Diffusion World Model), an evaluation framework that estimates the performance of a new LLM agent policy purely from pre-collected trajectories. The core idea is to learn a latent diffusion world model that simulates how the environment responds to the evaluation policy, without ever executing it in the real environment. Existing diffusion-based OPE methods guide full trajectories in a single pass by jointly diffusing states and actions, an assumption that breaks down for LLM agents whose actions are discrete text that must be sampled from the policy after observing the environment. Unlike autoregressive world models that suffer from compounding errors, ADWM models each transition as an independent denoising process, enabling reliable step-by-step rollouts where the world model and agent alternate in causal order. Crucially, the LLM agent under evaluation directly guides the diffusion generation at each step via a policy-conditioned score function, ensuring that simulated trajectories accurately reflect its decision-making patterns. Empirically, ADWM achieves accurate value estimates and evaluation reliability across diverse multi-turn agent tasks, demonstrating its promise as a practical framework for offline LLM agent evaluation.

Real-world digital environments are highly diverse and dynamic. These characteristics cause agents to frequently encounter unseen environments and distribution shifts, making continual learning in such environments essential for computer-use agents (CUAs). However, a key challenge lies in obtaining high-quality and environment-grounded training data without relying on costly human annotation. In this work, we introduce **ACuRL**, an **A**utonomous **Cu**rriculum **R**einforcement **L**earning framework that continually adapts agents to specific environments with zero human data. The agent first explores an environment to acquire initial experiences. During subsequent iterative training, a curriculum task generator leverages these experiences together with feedback from the previous iteration to synthesize new tasks tailored for the agent's current capabilities. To provide reliable reward signals, we introduce CUAJudge, a robust automatic evaluator for CUAs that achieves 93% agreement with human judgments. Empirically, our method effectively enables both intra-environment and cross-environment continual learning, yielding 3–29% absolute performance gains on the target environments without catastrophic forgetting on others. We also show that it can mitigate performance degradation under environment changes (e.g., version updates, platform migration, and resolution shifts). Further analyses show highly sparse updates (e.g., only 20% parameters), which helps explain the effective and robust adaptation.

The performance of AI coding agents is highly dependent on their underlying coding rules. However, existing coding rules are typically hand-crafted, making the process labor-intensive and often suboptimal. In this work, we propose RuleRefine, a framework for automatic coding rule refinement. RuleRefine maintains a pool of candidate coding rules and iteratively improves them. In each iteration, it employs an LLM-powered mutator module to generate variants from existing candidates, and then uses a judge module to evaluate these variants and update the pool with the best-performing ones. Extensive evaluations across two coding-agent frameworks, four backbone LLMs, and three benchmarks demonstrate that RuleRefine outperforms both manual engineering and existing prompt optimization baselines in terms of functional correctness of the generated code, code length, and/or generation cost (e.g., tokens used).

Scientific research proceeds through iterative cycles of hypothesis generation, experiment design, execution, and revision, often requiring researchers to explore multiple competing directions as evidence accumulates and priorities shift. LLM agents can automate parts of this process, but existing agents either concentrate reasoning within a single research thread or coordinate through a central planner with fixed objectives. As a result, they struggle to sustain parallel exploration across research directions or reorganize as promising and unproductive directions emerge over time. We introduce AutoScientists, a decentralized team of AI agents for long-running computational scientific experimentation. Rather than following decisions from a central orchestrator, agents independently interpret a shared experimental state, self-organize into teams around research directions, critique and filter proposals with a discussion phase before committing experimental compute, and exchange both successful and failed findings across teams to avoid redundant exploration. Under matched experimental budgets, AutoScientists outperforms prior agentic systems across biomedical machine learning, language-model training optimization, and protein fitness prediction. On BioML-Bench, spanning biomedical imaging, protein engineering, single-cell omics, and drug discovery, AutoScientists achieves a mean leaderboard percentile of 74.4% across 24 tasks, improving over the strongest prior biomedical agent by +8.33%. On GPT training optimization, AutoScientists reaches a target validation bits-per-byte 1.9x faster than autoresearch and continues discovering improvements from a stronger starting champion where the single-agent approach finds none (7 vs. 0 accepted improvements). On ProteinGym fitness prediction, AutoScientists discovers a method for ACE2-spike binding that improves over current state-of-the-art model by +12.5% Spearman correlation. Applied without modification to all 217 ProteinGym assays, the same method improves over the prior state of the art by +6.5% in Spearman correlation. Code: https://anonymous.4open.science/r/autoscientists_anonymize-D10C

LLM agents remain unreliable on long-horizon tasks, where small local failures can compound over extended interactions and lead to overall task failure. Although external harnesses can substantially improve robustness, harness design remains a manual and expensive process that requires searching over a large space of prompts, tool configurations, and control logic. We propose AutoSaddler, an automatic harness optimization framework that formulates harness improvement as an offline learning problem and iteratively updates the harness using failure signals from mini-batches. AutoSaddler combines failure-trace diagnosis, structured patch generation that treats the harness as code, and validation-based update selection. Experiments on GAIA2 and Terminal-Bench 2.0 show that AutoSaddler substantially improves agent performance over the corresponding base harnesses, achieving gains of 9.0 and 10.0 percentage points, respectively. Ablation studies further suggest that effective harness optimization benefits from three ingredients: deep debugging rather than shallow reflection, targeted modifications rather than unconstrained editing, and generalization-aware selection rather than trajectory-specific repair. Together, these results suggest that automatic harness optimization is a promising path toward more performant and reliable agent systems.

Manifold embedding for data visualisation is dominated by a small set of hand-designed state-of-the-art (SOTA) algorithms. Their design targets are fixed at development time, resulting in inflexible algorithmic behaviour that cannot be readily steered toward user-specified structural priorities. Hand-designing new embedding algorithms to address user priority requires highly specialised expertise and is time consuming. To address autonomous algorithm design tailored to user preference, we introduce an agentic algorithm generation pipeline AutoManifold. It composes new manifold embedding algorithms from a constrained vocabulary of affinity, cost, and optimisation primitives, supported by multi-agent large-language-model (LLM) orchestration. AutoManifold conditions every stage of its design on user-specified structural preservation preferences, and iteratively refines the algorithm configuration through an LLM-guided, metric-grounded iterative loop. We compare the generated algorithms against three strongest and most frequently used SOTA (t-SNE, UMAP, and PaCMAP) on real-world datasets spanning different difficulty regimes under identical evaluation infrastructure. The results demonstrate that LLM-driven, priority-conditioned algorithm synthesis can move beyond hyperparameter tuning, producing genuinely new algorithms that outperform traditional hand-designed methods, in their ability to steer towards user-specified properties without compromising much inherent structure in the original data.

Post-training data pipelines are traditionally hand-designed by researchers who orchestrate generation, verification, and diversity balancing. While Large Language Models (LLMs) can already execute individual subtasks, the architectural composition of these pipelines remains a manual bottleneck, excluding the most substantial engineering step in post-training from automation. We introduce AutoDataBench, a benchmark evaluating whether LLM agents can autonomously design post-training data-synthesis pipelines. Each task gives the agent only a description of a target capability and asks it to produce a complete, executable pipeline whose generated corpus is then used to fine-tune a fixed student model and scored on the evaluation. Each task is delivered as a natural-language instruction and a sandboxed workspace given to the agent, paired with an evaluation protocol. We instantiate AutoDataBench on five capabilities: instruction following, competition math, deep search, repo-level software engineering, and terminal automation. For each task we pair the evaluation with a strong human-designed pipeline whose downstream score serves as a reference point for the agent. Experimental results reveal that frontier agents close most of the gap to the human reference on text-centric tasks (e.g., Opus-4.7 nearly matches humans in instruction following), yet a noticeable performance gap persists in tasks requiring executable-environment construction (SWE and Terminal). Furthermore, we observe a decoupling between downstream task-solving strength and pipeline-construction quality, where strong reasoning models like GPT-5.5 may still produce suboptimal pipelines. These findings suggest that autonomous environment synthesis remains the primary bottleneck for self-improving data loops, distinct from mere task-solving competence.

Language agents increasingly operate over streams of related tasks, yet existing memory systems struggle to convert accumulated experience into reusable knowledge. Retrieval-augmented and structured memory methods record per-session observations effectively, but often couple acquisition and consolidation into a single online process, leaving the agent without a global view across sessions to discover recurring patterns, abstract shared procedures, or prune redundant entries. Inspired by complementary learning systems theory, we propose Auto-Dreamer, a learned offline consolidator for language-agent memory. Auto-Dreamer decouples fast per-session memory acquisition from slow cross-session consolidation. Given a region of a typed memory bank, the consolidator performs bounded tool-use to search memory, inspect candidate entries, and trace them back to raw source trajectories, synthesizing provenance-grounded replacement memories that supersede the original region. We train Auto-Dreamer via GRPO, using end-to-end agent performance as the reward signal to learn how to consolidate memories acquired through fast online experience. Trained on ScienceWorld trajectories alone, Auto-Dreamer transfers without retraining to held-out ALFWorld and WebArena, improves task success over fixed, RL-trained, and prompted memory baselines in both continual-memory deployment and fixed-bank consolidation, and does so with an active memory bank an order of magnitude smaller than competitive baselines.

NeurIPS now publishes a broad mix of algorithmic, theoretical, empirical, dataset, benchmark, systems, and evaluation-infrastructure work. Yet we lack systematic evidence on how the composition of this mix has changed over time. We audit titles, abstracts, and metadata for 19,361 accepted NeurIPS papers from 2015--2024. We define semantic axes for formal-theoretical evidence, benchmark/resource evidence, empirical-performance evidence, artifact-release orientation, scale/capability framing, claim strength, and hedging; score papers with contrastive embedding measures; and validate the constructs with blinded LLM-agent ratings, lexical probes, encoder triangulation, length checks, and topic controls. In raw decade trends, accepted papers shift toward benchmark/resource evidence (+0.842 SD), empirical-performance evidence (+0.750 SD), and scale/capability framing (+0.606 SD), and away from formal-theoretical framing (-0.872 SD). Topic composition explains much of this movement, but embedding-cluster residuals remain +0.246, +0.264, +0.247, and -0.300 SD, respectively. Independent lexical LDA controls preserve the same signs at smaller magnitudes. Artifact-release framing is a useful negative result: it rises modestly in the raw corpus (+0.226 SD) but is near zero or negative after topic adjustment. We release a reusable audit package that includes metadata, semantic and lexical features, validation artifacts, robustness scripts, and an interactive Evidence Norms Atlas for inspecting aggregate evidence-framing patterns by year, topic neighborhood, and track.

LLM-based peer review is now used at production scale, but rigorous evaluation remains limited. We introduce a benchmark for evaluating AI peer reviewers’ detection of injected errors and quantifying false-positive rates across error type, severity, prompting strategy, and reviewer architecture. We also present skepthical, a multi-agent AI reviewer that verifies citations by retrieving and reading cited papers, checks mathematics with computer algebra, audits numerical claims with code, and uses a cross-model ensemble for general scientific issues. On 15 astrophysics preprints containing 177 injected errors, we benchmark 12 reviewer systems: skepthical; six single-LLM reviewers from three providers, each run with free-form and structured prompts; two Claude Code reviewers using the same Claude model and tools, one with a free-form prompt and one with a structured prompt; and three external reviewers. Detection performance varies sharply by model, prompt, tool use, and error type. Under the same free-form prompt, three frontier LLMs differ by 48.6 pp in total detection rate; a structured prompt halves this spread by improving the weaker models. Within Claude Code, changing only the prompt from free-form to structured raises detection by 32.2 pp. skepthical achieves the highest overall detection rate, 70.6±6.2%, with the best performance on general issues, 82.2%, and citation errors, 46.7%. On numerical errors, it matches both GPT-5.5 configurations at 80.0%, behind structured Claude Code at 91.1%; on mathematics, its 73.8% rate overlaps with both GPT-5.5 configurations within confidence intervals. False-positive rates also vary substantially: free-form Opus 4.7 has the lowest overall rate at 1.0%, while skepthical is the only reviewer with no false-positive failure mode jointly across mathematics, numerics, and citations, yielding a 2.8% overall rate. skepthical is deployed publicly at skepthical-ai.org, and we release the evaluation set as a benchmark.

An attacker that strategically chooses when to attack is much harder to catch than one that attacks indiscriminately. AI control is a safety framework for deploying capable but untrusted AI agents under oversight from a weaker trusted monitor and a limited human audit budget. Control evaluations stress-test these protocols by pitting a red-team attack policy against the blue-team monitor, but current evaluations typically assume attackers that do not strategically select when to attack. However, a capable attacker would not attack indiscriminately. We study this capability, attack selection, in agentic settings by decomposing attack decisions into a start policy, which decides when an attacker should attack, and a stop policy, which decides when an attacker should abort an ongoing attack. Across two agentic settings, BashArena and LinuxArena, both policies substantially lower measured empirical safety without changing the underlying attack capability. At a 1% audit budget, our start policy reduces safety by 20pp on both BashArena and LinuxArena, and our stop policy reduces safety by 20pp on BashArena and 28pp on LinuxArena. These reductions in safety should be interpreted as upper bounds on the effect of attack selection. Existing control evaluations may therefore yield overly optimistic safety estimates against selective attackers. We recommend that future evaluations, system cards, and safety cases should elicit attack selection to produce more realistic safety estimates.

Reinforcement learning (RL) is increasingly used to improve the reasoning, coding, and tool-use capabilities of large language models, but agentic RL remains prohibitively expensive. Scaling RL to agentic LLMs requires supporting complex workloads, including multi-policy collaborative training, while efficiently using elastic, heterogeneous, and cross-region compute resources. Existing LLM RL systems support some of these capabilities, but each new extension often requires dedicated system engineering. This burden arises from trainer-centered control architectures and the lack of principled abstractions for RL system components. To address these limitations, we propose AstraFlow, a dataflow-oriented RL system that replaces conventional trainer-centered control with principled component abstractions. In AstraFlow, rollout services, dataflow management, and training are decoupled into autonomous components, enabling the system to natively support complex multi-policy agentic RL workloads and efficiently exploit diverse compute resources. We evaluate AstraFlow across math, code, search, and AgentBench workloads, showing that the same system supports multi-policy training, elastic scaling, heterogeneous cross-region execution, and composable data algorithms without system-level code changes. In multi-policy collaborative training, AstraFlow achieves comparable or better accuracy than existing RL systems while speeding up training time by 2.7x.

We study a stochastic bandit problem in which learning is split between two agents. In each round, a human observes rewards but can choose only between a pair of arms selected by an assistant, while the assistant observes the human's behavior but not realized rewards. We call this model the **assistive dueling bandit**. The human is modeled as a learning agent whose regret on any subset of arms is bounded by a function $g(T)$ unknown to the assistant. This is a natural model for recommender systems or AI assistants that must present a slate of options to a user who may be still learning about their own preferences. We provide a general reduction from assistive dueling bandits to the problem of **max-finding with imprecise feedback.** With it, we design assistant algorithms that cooperate with any $g(T)$-regret human to achieve a joint regret of $\tilde{\mathcal{O}}(K \cdot g(T/K))$, even without knowledge of $g(T)$. Crucially, when $g(T) \in \mathcal{O}(\sqrt{T})$, our method recovers the nearly optimal minimax rate of $\tilde{\mathcal{O}}(\sqrt{KT})$, implying that splitting up the responsibilities of learning in this way can be done without suffering any excess regret asymptotically. We complement our theoretical analysis with experimental evidence that this algorithm outperforms an assistant implemented via standard dueling bandit algorithms.

Recent advances in machine learning and large-scale biological data collections have revived the prospect of building a virtual cell, a computational model of cellular behavior that could accelerate biological discovery. One of the most compelling promises of this vision is the ability to perform in silico phenotypic screens, in which a model predicts the effects of cellular perturbations in unseen biological contexts. This task combines heterogeneous textual inputs with diverse phenotypic outputs, making it particularly well-suited to LLMs and agentic systems. Yet, no standard benchmark currently exists for this task, as existing efforts focus on narrower molecular readouts that are only indirectly aligned with the phenotypic endpoints driving many real-world drug discovery workflows. In this work, we present AssayBench, a benchmark for phenotypic screen prediction, built from 1,920 publicly available CRISPR screens spanning five broad classes of cellular phenotypes. We formulate the screen prediction task as a gene rank prediction for each screen and introduce the adjusted nDCG, a continuous metric for comparing performance across heterogeneous assays. Our extensive evaluation shows that existing methods remain far from empirically estimated performance ceilings and zero-shot generalist LLMs outperform biology-specific LLMs and trainable baselines. Optimization techniques such as fine-tuning, ensembling, and prompt optimization can further improve LLM performance on this task. Overall, AssayBench offers a practical testbed for measuring progress toward in-silico phenotypic screening and, more broadly, virtual cell models. Our benchmark is available at https://github.com/Genentech/AssayBench.

Mapping software issue descriptions to relevant code segments, a task known as code localization, is a primary bottleneck in automated software engineering. While recent attempts have automated it using code graphs that capture syntax-level dependencies among code elements, including functions and modules, their code-centric approaches often overlook the design rationales and prior fix patterns present in a codebase's development history, such as pull requests and commits. We propose Co-Locator, a collaborative multi-agent code localization framework that integrates code graphs reflecting functional dependencies and knowledge graphs constructed from triplets capturing design rationales and past code modifications scattered across the development history. During localization, a Code Graph (CG) agent explores functional hierarchies and consults a Knowledge Graph (KG) agent to gather insights into developer intents and prior bug fixes. Combined with a Retriever agent that assists CG and KG agents, Co-Locator leverages a holistic view of the code's functional and historical context, considering what the code implements and how it was changed, to identify locations to fix. Experiments show that Co-Locator significantly outperforms 12 state-of-the-art baseline methods.

Long-horizon AI agents execute complex workflows spanning hundreds of sequential actions, yet a single wrong assumption early on can cascade into irreversible errors. When instructions are incomplete, the agent must decide not only whether to ask for clarification but when, and no prior work measures how clarification value changes over the course of execution. We introduce a forced-injection framework that provides ground-truth clarifications at controlled points in the agent's trajectory across four information dimensions (goal, input, constraint, context), three agent benchmarks, and four frontier models (three per benchmark; one on a single benchmark only; 84 task variants; 6,000+ runs). Counter to the common intuition that ``earlier is always better,'' we find that the value of clarification depends sharply on what information is missing: goal clarification loses nearly all value after 10% of execution (pass@3 drops from 0.78 to baseline), while input clarification retains value through roughly 50%. Deferring any clarification type past mid-trajectory degrades performance below never asking at all. Cross-model Kendall τ correlations (0.78-0.87 among models sharing identical task coverage; 0.34-0.67 across the full 4-model panel) confirm these timing profiles are substantially task-intrinsic. A complementary study of 300 unscripted sessions reveals that no current frontier model asks within the empirically optimal window, with strategies ranging from over-asking (52% of sessions) to never asking at all. These empirical demand curves provide the quantitative foundation that existing theoretical frameworks require but have lacked, and establish concrete design targets for timing-aware clarification policies. Code and data will be publicly released.

A bottleneck in learning to understand articulated 3D objects is the lack of large and diverse datasets. In this paper, we propose to leverage large language models (LLMs) to close this gap and generate articulated assets at scale. We reduce the problem of generating an articulated 3D asset to that of writing a program that builds it. We then introduce a new agentic system, Articraft, that writes such programs automatically. We design a programmatic interface and harness to help the LLM do so effectively. The LLM writes code against a domain-specific SDK for defining parts, composing geometry, specifying joints, and writing tests to validate the resulting assets. The harness exposes a restricted workspace and interface to the LLM, validates the resulting assets, and returns structured feedback. In this way, the LLM is not distracted by details such as authoring a URDF file or managing a complex software environment. We show that this produces higher-quality assets than both state-of-the-art articulated-asset generators and general-purpose coding agents. Using Articraft, we build Articraft-10K, a curated dataset of over 10K articulated assets spanning 245 categories, and show its utility both for training models of articulated assets and in downstream applications such as robotics simulation and virtual reality.

Computer-use agents turn visual predictions into executable GUI clicks, so uncertainty estimates must support rejection, calibration, miss-severity ranking, and spatial safety regions. Existing post-hoc uncertainty quantification (UQ) evidence is fragmented across isolated model and dataset pairs, leaving unclear whether UQ rankings remain stable when the agent, benchmark, or observable interface changes. We present a cross-regime benchmark for post-hoc UQ in single-step executable GUI grounding, covering a 27-method, seven-family open-weight matrix over 4 GUI-grounding VLM agents and 4 datasets, plus an 8-method API-compatible closed-source matrix across 3 frontier vendors where logits, hidden states, and attention maps are unavailable. The main finding is selective transfer: UQ rankings are stable across datasets for a fixed model, but degrade across model classes and observable interfaces. Hidden-state and density methods form the most stable open-weight family, while CoCoA-1MCA, Focus, sampling-based scores, and verbalised self-assessment win in specific regimes. Ranking transfer is strongest within a fixed model across datasets, reaching Spearman $\rho=0.969$ and averaging $\rho=0.705$ over 120 open-weight pairs. In contrast, open-weight to closed-source transfer is the strongest break: cross-tier Spearman $\rho$ averages only $+0.08$ over 12 vendor x dataset pairs on the shared 8-method intersection, indicating that UQ recommendations do not survive removal of internal model signals. Model-class transitions further reshape UQ preferences: attention, verbalised, and VLM-native families lose AUROC on every dataset, while density methods remain stable; a scale-only baseline shows that logit-family degradation on ScreenSpot-Pro is partly confounded by scale rather than grounding fine-tuning alone. Finally, conformal click regions show that score-level discrimination is not enough for deployment: locally weighted disks can shrink radii by 40-60% when the plug-in UQ is calibrated, but coverage can degrade under calibration-test or interface mismatch. We release per-item records, calibration/test splits, UQ scores, and analysis scripts as a reproducible basis for regime-aware UQ selection in GUI agents.

Language agents are increasingly deployed in complex professional workflows, with tutoring emerging as a particularly high-stakes capability that remains largely unmeasured in existing benchmarks. Effective tutor agents require more than producing correct answers or executing accurate tool calls: a robust tutor must diagnose learner state, adapt support over time, make pedagogically justified decisions grounded in educational evidence, and execute interventions within realistic learning-management systems. We introduce EduAgentBench, a source-grounded benchmark for holistically evaluating tutor agents across the full scope of teaching work. It contains 150 quality-controlled tasks across three capability surfaces: situated multi-turn tutoring, professional pedagogical judgment, and Canvas-style teaching workflow completion. Tasks are constructed through a pedagogical-insight-driven pipeline and evaluated with complementary verification signals and human review. Across a comprehensive evaluation of frontier models, our findings reveal that current models are generally capable of bounded pedagogical judgment, but still fall short of professional teaching standards in situated tutoring and autonomous teaching-workflow execution. To our knowledge, EduAgentBench is the first theory-grounded and realistic benchmark for evaluating the holistic teaching capability of tutor agents, providing a measurement foundation for developing future tutor agents that can support realistic teaching work.