NeurIPS 2026 Papers

Frontier AI models and multi-agent systems have led to significant improvements in mathematical reasoning. However, for problems requiring extended, long-horizon reasoning, existing systems continue to suffer from fundamental reliability issues: hallucination accumulation, memory fragmentation, and imbalanced reasoning-tool trade-offs. We introduce **STAR-Math**, a multi-agent framework that systematically addresses these challenges through meta-level supervision and structured Reasoner-Verifier interaction. STAR-Math is structured as an orchestrated state machine with nested challenge-step-replan loops, governed by a reasoning-free Python orchestrator that separates control from inference and bounds error propagation through trace-back and re-planning. Our key innovation is a **persistent Meta-Strategist** that maintains cross-attempt memory and exercises meta-level control by issuing high-level strategic guidance or mandatory directives, so the system can escape unproductive loops rather than stagnate or over-rely on tools. STAR-Math achieves state-of-the-art results on all eight top-tier competition benchmarks: AIME 2025-2026, MathArena Apex Shortlist, MathArena Apex 2025, Putnam 2025, IMO 2025, HMMT February 2026, and USAMO 2026. It obtains perfect scores on AIMEs, Putnam, and HMMT, and shows its largest margin on Apex 2025, scoring 93.75% compared with 80.21% by the strongest baseline GPT-5.5. Ablation studies show that the gains arise from the framework's orchestration rather than from model-level diversity since removing key components or substituting in mixed backbones consistently weakens performance.

AI agents are increasingly used to diagnose and mitigate failures of production systems, known as agentic Site Reliability Engineering (SRE). Current SRE benchmarks are limited to oversimplistic SRE tasks and are unfortunately hard to extend due to bespoke designs. We present SREGym, a high-fidelity, interactive benchmark for SRE agents. SREGym exposes a live system environment built atop real-world cloud-native system stacks, where high-fidelity failure scenarios are emulated through fault injectors. SREGym models the complexity of production environments by simulating (1) a wide range of faults across the stack, (2) various ambient noises, and (3) diverse failure modes such as metastable failures and concurrent failures. SREGym is architected as a modular, extensible framework that orchestrates fault injectors and event emulators across stacks; this framework is key to curating high-fidelity, challenging SRE problems. SREGym shows that the effectiveness of frontier models and agents varies significantly (40+%) in mitigating different kinds of problems. SREGym is actively maintained as an open-source project and has been used by many researchers and practitioners.

How fast *could* a deep-learning model run on target hardware, and how far is today’s implementation from that limit? These questions are central to software, hardware, and algorithm optimizations. Speed-of-Light (SOL) analysis answers them by computing a workload’s theoretical minimum execution time on a given architecture. Yet deriving SOL bounds remains manual, error-prone, and disconnected from rapid model development. To close this gap, we introduce SOLAR, a framework that automatically derives validated SOL bounds from Pytorch and JAX source code. SOLAR leverages both *generative* and *deterministic* components in its flow: an LLM frontend translates any source programs into an executable Affine Loop IR, validated by output comparison; a deterministic flow lifts the IR into an einsum graph; and an analytical backend computes unfused, fused, and cache-aware SOL bounds. SOLAR provides comprehensive operator and language coverage, produces validated bounds with zero observed SOL violations, and offers multi-fidelity analysis that tightens bounds and surfaces optimization insights. We evaluate SOLAR across KernelBench, JAX/Flax models, and robotics workloads. These experiments demonstrate four use cases: headroom analysis at multiple fidelity levels, identifying optimization opportunities, cross-platform exploration, and inverse-roofline hardware provisioning.

As agentic AI systems become increasingly capable of generating and optimizing GPU kernels, progress is constrained by benchmarks that reward speedup over software baselines rather than proximity to hardware-efficient execution. We present SOL-ExecBench, a benchmark of 235 CUDA kernel optimization problems extracted from 124 production and emerging AI models spanning language, diffusion, vision, audio, video, and hybrid architectures, targeting NVIDIA Blackwell GPUs. The benchmark covers forward and backward workloads across BF16, FP8, and NVFP4, including kernels whose best performance is expected to rely on Blackwell-specific capabilities. Unlike prior benchmarks that evaluate kernels primarily relative to software implementations, SOL-ExecBench measures performance against analytically derived Speed-of-Light (SOL) bounds computed by SOLAR, our pipeline for deriving hardware-grounded SOL bounds, yielding a fixed target for hardware-efficient optimization. We report a SOL Score that quantifies how much of the gap between a release-defined scoring baseline and the hardware SOL bound a candidate kernel closes. To support robust evaluation of agentic optimizers, we additionally provide a sandboxed harness with input correctness checks, isolated compilation and execution, CUPTI-based GPU activity timing, and defenses against common reward-hacking strategies. A third-party post-release evaluation by Cursor reported a $1.38\times$ geometric-mean speedup over a single-agent optimized PyTorch baseline and an overall SOL Score of 0.56 across the full benchmark. SOL-ExecBench reframes GPU kernel benchmarking from beating a mutable software baseline to closing the remaining gap to hardware Speed-of-Light.

Large language model (LLM)-powered multi-agent systems have been shown to effectively solve real-world end-to-end problems through collaboration among diverse agents. However, as LLM capabilities advance and user demands continue to grow, the range of problems that agents can address expands dramatically, while the population of available agents and tools continues to increase, giving rise to open and evolving ecosystems of autonomous agents. In this paradigm, agents are likely to become increasingly specialized, each focusing on particular capabilities or problem types. In such a setting, a central challenge is how to efficiently assemble an appropriate team of agents to take on different responsibilities from a massive agent pool for a given task. In this work, we study this problem in a controlled setting with 300 task-oriented agents. SMART (Scalable Multi-Agent Role-conditioned Teaming) first plans a task-conditioned collaboration topology and candidate role set, then constructs a compact role-conditioned candidate pool, and finally uses a LLM-free Monte Carlo Tree Search (MCTS) to assign agents to roles online via metadata-based rollouts. This design enables high flexibility without incurring prohibitive inference costs. Across multiple experiments, our approach achieves superior performance on GSM8K (93.64\%), MATH (61.73\%), HumanEval (93.89\%), and MBPP (87.68\%) using gpt-4o-mini under a limited budget.

Large Language Model (LLM) agents are increasingly deployed under structured generation, where grammar constraints enforce output format at inference time. Whereas existing compression methods target unconstrained generation, we first identify a fundamental mismatch: under structured generation, grammar constraints reduce the model's output distribution to valid tokens, and the logits matter most at the small set of positions that decide the agent's action. Motivated by this finding, we propose SKIM (Selective Knowledge-Informed Masking), the first pruning and distillation framework for LLM agents under structured generation, with both stages guided by token role. SKIM partitions token positions into context, decision, and format categories that reflect the input the model reads, the constrained options it chooses among, and the format realized by the constraint mechanism. Specifically, this partition drives two complementary axes: (1) category-aware pruning saliency scores that preserves parameters most relevant for decision positions; and (2) an on-policy distillation objective that aligns student and teacher on valid tokens, paired with category-aware feature matching. We evaluate SKIM on tool-calling benchmarks and within multi-step agent loops, across multiple model scales and sparsity levels. SKIM consistently improves accuracy over baselines, with its superiority becoming most striking in high-sparsity regimes where conventional methods struggle, demonstrating the potential of category-aware compression for deployment-time grammar-constrained agents.

Frontier LLMs are increasingly capable of expert-level scientific reasoning, but using them as reliable scientific agents requires more than general reasoning ability. Scientific simulation is a central test case: modern simulators are configured through executable interfaces such as XML decks, input scripts, and namelists that function as domain-specific languages tied to the simulator's internal API. Translating a researcher's natural-language intent into a runnable configuration remains a recurring expert bottleneck. We study how much of this bottleneck a general-purpose LLM coding harness, Claude Code, can absorb when wrapped with a **Simulator-Interface Grounding Adapter (SIGA)**: a package of skills, tools, and workflow control flow that grounds the agent's outputs in simulator documentation, schema, and example libraries. We instantiate SIGA for GEOS, an open-source multiphysics simulator used in CO₂-storage and induced-seismicity research. A Resolution-IV factorial surfaces three benefits over vanilla Claude Code. First, SIGA improves **reliability**, reducing across-seed variance by roughly 40× by preventing unparseable or empty decks on a hard tail of compound multi-physics tasks. Second, it improves **quality**, raising mean structural similarity by about +7 percentage points on the same hard tail. Third, a self-evolved variant matches the best hand-designed cell with roughly 16% fewer tool calls, suggesting that automatic SIGA discovery is tractable. A preliminary human baseline finds that geoscience-domain-expert volunteers new to GEOS take between 8 and 36 times as long as the agent on a representative task. An explicit human-consultation tool is used in only about 3% of under-specified trials, with the agent instead relying on the on-disk example library as a cheaper retrieval substitute. A small OpenFOAM transfer study indicates that the recipe is not specific to GEOS XML: the same stop-hook component again dominates reliability, and the best SIGA cell outperforms both vanilla Claude Code and a constrained Foam-Agent lint-only baseline. We close with SIGA-design recommendations grounded in failure modes that our best configuration still does not fix.

Most video anomaly understanding (VAU) systems still infer anomaly judgments from evidence collected in a feed-forward way under predefined rules, typically sampled frames, or events clustered from such frames. This protocol fails when the collected evidence is incomplete under imperfect predefined rules, because once reasoning begins the model can no longer search for the missing evidence. We address this limitation with SEEK-VAU, which reframes VAU as an agentic search-and-verify interaction protocol under a bounded search budget: instead of passively reasoning over pre-collected evidence, the policy actively gathers evidence across the three stages of an anomaly event chain (precursor, trigger, and aftermath) and verifies its sufficiency before finalization, turning event-chain completeness into an explicit design rather than an outcome of predefined rules. Although search-and-verify spans multiple interaction turns, the bounded search budget caps both the per-turn input tokens and the cumulative visual context, keeping inference efficient. To make this behavior learnable, we introduce evidence-faithful counterfactual verification (EFCV), which rewards selected evidence that remains supportive, compact, and necessary under counterfactual verification. We further introduce SEEK-Bench, featuring video-level episodes with temporal interval annotations, semantic QA and event-chain stage labels. Together, SEEK-VAU and SEEK-Bench establish a strong foundation for evidence-faithful and actively verifiable VAU. The code and data will be released.

The exponential growth of machine learning submissions has strained the traditional peer review process, resulting in slow feedback loops for authors and an immense burden on reviewers to rigorously audit technical soundness and verify literature. To address this, we introduce ScholarPeer, a multi-agent framework designed to operationalize the rigorous auditing workflow of a senior researcher. Rather than attempting to replace human judgment, ScholarPeer serves as a co-scientist: acting as a mentor for rapid author iteration prior to submission, and as an active verification assistant that augments human reviewers. The framework structurally decouples contextualization from critique by deploying a sub-domain historian to synthesize the field's trajectory, a baseline scout to proactively hunt for omitted state-of-the-art comparisons, and a multi-aspect Q&A engine that deeply audits technical soundness—scrutinizing internal logical consistency, experimental validity, and mathematical rigor—while cross-referencing claims against top-tier academic venues. We comprehensively evaluate ScholarPeer on $\sim$1,800 ICLR submissions spanning 2020 through 2025. Our results show that ScholarPeer achieves significant win-rates against state-of-the-art fine-tuned models and search-augmented agentic baselines.

Compositional implicit surface representations model scenes as collections of objects, each encoded by a Signed Distance Field (SDF). A fundamental limitation of this approach is that multiple SDFs can produce geometries that interpenetrate, violating physical plausibility. Existing mitigation strategies rely on soft penalty terms that reduce but do not eliminate intersections, and require careful loss weighting. To truly prevent interpenetration, we propose a hard constraint on vector-valued SDFs and introduce S2MDF, a lightweight plug-and-play module that enforces the constraint on any object-compositional SDF representation without architectural modifications. It introduces \fs{negligible} computational overhead and is compatible with linearly-interpolated standard meshing algorithms such as Marching Cubes. It can be applied during training or as a post-processing step. Experiments on multiple state-of-the-art compositional methods show that S2MDF reduces intersections to numerical precision while preserving reconstruction quality, outperforming existing mitigation strategies.

Runtime monitoring of autonomous systems traditionally relies on mapping continuous sensor observations to discrete logical propositions defined over low-dimensional state variables. This abstraction breaks down in perception-driven settings, where such mappings require additional learned modules that are often computationally expensive, brittle, and semantically misaligned. In this work, we propose Embedding Temporal Logic (ETL), a temporal logic that performs monitoring directly in learned embedding spaces. ETL defines predicates through distances between observed embeddings and target embeddings derived from reference observations. This formulation allows specifications to capture high-level perceptual concepts, such as similarity to visual goals or avoidance of semantic regions, that are difficult or impossible to express using traditional predicates. By composing these predicates with temporal operators, ETL naturally expresses temporally extended and sequential perceptual behaviors. We introduce ETL monitors for evaluating specifications over bounded embedding traces, along with a conformal calibration procedure that provides reliable and safety-oriented predicate evaluation. We evaluate our approach across multiple manipulation environments to show that ETL achieves strong empirical agreement with ground-truth semantics, including accurate monitoring of temporally composed behaviors.

Game balancing is a longstanding challenge requiring repeated playtesting, expert intuition, and extensive manual tuning. We introduce RuleSmith, a framework for automated game balancing that couples multi-agent LLM self-play with Bayesian optimization over a parameterized rule space. As a proof of concept, we build CivMini, a simplified civilization-style game with two asymmetric factions governed by 12 tunable parameters. LLM agents read textual rulebooks and game states to generate legal actions, enabling fast evaluation of balance metrics such as win-rate disparity. To search the rule space efficiently, we use Bayesian optimization with acquisition-based adaptive sampling: candidates with high Expected Improvement receive more evaluation games, while exploratory candidates receive fewer. RuleSmith consistently finds near-balanced configurations and produces interpretable parameter adjustments. It provides a scalable alternative to manual playtesting. Code will be available upon acceptance.

LLM routing aims to choose, for each prompt, which model in a portfolio is most likely to answer correctly. This only makes sense when the benchmark contains enough prompt-specific variation for different models to be best on different inputs. On coarse-scored LLM benchmarks, however, many models often tie for the best score, and the single best model may already leave little room for improvement. We therefore propose a routeability-first audit, instantiated as RAP and RouteabilityBench, a reviewer-runnable evaluation package. Across four benchmarks and 66 LLMs, generic routing enhancements such as larger embedding sets, dimensionality reduction, selective abstention, clustering, and stronger regressors do not consistently outperform the single best model. The clearest positive evidence is concentrated in portfolio-dependent Omni-MATH settings, while GPQA, MMLU-Pro, and IFEval show different ways in which routing can remain hard or predictor-sensitive. Our practical recommendation is that routing gains should be reported together with routeability diagnostics, tied-best analysis, portfolio construction details, and family-corrected statistical tests.

The ability to reason, adapt, and creatively solve problems under unexpected challenges is essential for robots operating in real-world environments. However, current robotic benchmarks primarily emphasize skill-level execution and provide limited insight into such cognitive reasoning capabilities. We introduce RoboWits, a bi-manual robotic benchmark designed to systematically evaluate cognitive reasoning, creative tool use, and robustness to unexpected conditions. To enable scalable construction of high-quality reasoning-centric unexpected scenarios, we propose an automated task generation pipeline formulated as a multi-agent cooperative framework, comprising agents for seed task generation and verification, metric generation, scene generation, and task mutation. Using the pipeline, we curated 30 diverse seed tasks and 208 tasks with mutations and graded difficulty across geometry, material, and assembly-based reasoning. We benchmark popular robot policies, pre-trained VLAs, and oracle-state planners. Our results reveal a significant performance gap: while pre-trained VLAs exhibit preliminary success on seed tasks after single-task fine-tuning, they struggle to perform on mutated tasks, implying their brittleness in manipulation tasks requiring reasoning, strategy adaptation, and robustness to deceptive or constrained environments.

Coding agents are increasingly deployed in real software development, where a single version iteration requires months of coordinated work across many files. However, most existing benchmarks focus predominantly on single-issue bug fixes from Python repositories, with coarse pass/fail evaluation outcomes, and thus fail to capture long-horizon, multi-target development at real engineering scale. To address this gap, we present ROADMAPBENCH, a benchmark of 115 long-horizon coding tasks grounded in real open-source version upgrades across 17 repositories and 5 programming languages. Each task places the agent on a source-version code snapshot and provides a multi-target roadmap instruction requiring it to implement the functionality introduced in the target version, with a median modification of 3700 lines across 51 files. We conduct a systematic evaluation on thirteen frontier models and find that even the strongest, Claude-Opus-4.7, resolves only 39.1% of tasks, while the weakest achieves merely 5.2%, in stark contrast to the high scores these models attain on existing bug-fix benchmarks, suggesting that long-horizon software development remains a largely unsolved problem.

Specialized agents collaborate by transferring memory: retrieved records, generated summaries, tool traces, and long-running state. In regulated domains this is a governance problem, not a retrieval problem: the receiver's task matters, but so does whether it was authorized to receive each item and whether the handoff exposed more than necessary. We formalize minimum-necessary context selection as a constrained submodular subset-selection problem with an additive exposure cost and a distributional task-success risk bound. Our algorithm, ScopeSelect, first hard-masks items whose layer or scope the receiver contract does not admit, then runs a cost-benefit greedy selector against a monotone submodular utility, attaining a $\\frac{1}{2}(1-1/e)$ approximation bound under a budget, and finally wraps threshold selection in Conformal Risk Control, giving a finite-sample distribution-free guarantee $\\mathbb{E}[L(S_{\\hat{\\gamma}})] \\le \\alpha$ on exchangeable calibration data. We evaluate on HealthHandoff-Bench, a synthetic benchmark with 21,600 healthcare-agent handoff tasks across four workflows, and Synthea-HH, a realism tier ingested from 340 Synthea-generated FHIR bundles. A three-source blinded label audit compares benchmark labels, model-generated labels, and clinical-informatics expert review, with high agreement for sensitivity labels and policy-boundary disagreements for necessity and authorization. Against authorized TF-IDF, embedding retrieval, a cross-encoder reranker, and a calibrated LLM-as-selector, ScopeSelect is the only method meeting $\\alpha = 0.15$ at competitive task success. The hard-mask safety property holds under 30% label noise, 100 adversarial attacks, and receiver-role shift; a live-LLM run with Claude Sonnet 4.5 exhibits a calibration-transfer gap that motivates per-workflow recalibration. We release the benchmark, Synthea-HH pipeline, CRC calibration harness, and a signed-handoff prototype.

Reinforcement learning (RL) shows promise for enhancing LLM agentic reasoning capabilities with external environments. However, sparse terminal rewards hinder fine-grained, state-level optimization. Although process reward modeling offers a promising alternative, training dedicated reward models often entails substantial computational costs, risks reward hacking, and faces annotation bottlenecks. To address these challenges, we introduce *RewardFlow*, a lightweight method for estimating state-level rewards for agentic reasoning. *RewardFlow* leverages the intrinsic topological structure of states within trajectories by constructing state graphs. This enables topology-aware graph propagation to estimate state-wise contributions to success, yielding principled, annotation-free state-level rewards. As dense rewards for RL optimization, *RewardFlow* substantially outperforms prior RL baselines across four agentic benchmarks, with average success-rate gains of +6.2\% on text-based benchmarks and +29.7\% on visual reasoning over the strongest baseline across three model scales, and over +10\% accuracy improvement on DeepResearch, while demonstrating superior robustness and training efficiency.

Multi-Agent Systems (MAS) built on Large Language Models (LLMs) require effective orchestration to coordinate specialized agents, yet training such orchestrators is hindered by limited supervision and high computational cost. We propose **Orch**estration **R**eward **M**odeling (**Orch-RM**), a self-supervised framework for evaluating orchestration quality without human annotations. Orch-RM leverages intermediate artifacts from multi-agent executions to construct win-lose pairs for Bradley-Terry reward model training. Unlike existing MAS test-time scaling and orchestrator training frameworks that rely on costly sub-agent rollouts, Orch-RM operates directly at the orchestration level, enabling efficient and high-performing reward-guided orchestrator training and MAS test-time scaling. As shown in Figure 1, Orch-RM improves training efficiency by up to 10$\times$ in token usage while improving MAS test-time scaling performance by up to 8\% in accuracy. These gains consistently transfer across multiple domains, including mathematical reasoning, web-based question answering, and multi-hop reasoning, demonstrating orchestration-level reward modeling as a scalable direction for robust multi-agent orchestration. Data, code, and trained models will be released upon publication.

Recent advancements in multimodal diffusion language models (MDLMs) have exhibited strong global modeling capabilities through iterative refinement. We observe that mask uncertainty follows an overall downward trend during iterative denoising, while a few denoising steps exhibit localized entropy spikes closely associated with hallucinations. We argue that anomalous steps can be identified by measuring the entropy deviation of each denoising step from its neighboring temporal window. We hypothesize that discrete token commitment discards the distributional information at masked positions, limiting the continued interaction between candidate semantics and the context during entropy spikes. With this goal, we present StepRefine, a plug-and-play decoding strategy that leverages semantic context to achieve reliable denoising. The core of our method lies in a step-aware switch between discrete denoising and latent refinement. The model employs probability-weighted embeddings at entropy spike steps to preserve diverse reasoning hypotheses and performs multiple rounds of continuous refinement. Meanwhile, a refinement controller monitors distributional stability to reduce unnecessary loops and switch the model back to standard discrete denoising. Through extensive experiments, StepRefine demonstrates significant hallucination mitigation across different MDLMs on multimodal benchmarks.