NeurIPS 2026 Papers

Off-policy reinforcement learning (RL) for large language models bottlenecks at the rollout stage: every parameter update invalidates the key-value (KV) cache of in-flight agent trajectories, forcing either expensive recomputation or staleness that destabilizes training. Recent advances such as M2PO and asynchronous RLHF tolerate token-level staleness by reweighting the policy gradient with second-moment importance corrections, but they implicitly assume that a stale rollout is sampled from a well-defined behavior policy $\pi_{\theta_{\text{old}}}$. We show that this assumption is silently violated by every modern partial-rollout system: once a stale KV cache is reused under updated parameters $\theta_{\text{new}}$, the resulting token distribution is a hybrid $\pi_{\text{hyb}}$ whose attention projections consume stale keys and values while its MLP, output, and gating projections use the new weights. The token-level importance weight $w_t = \pi_{\theta_{\text{new}}}(y_t \mid x_{<t}) / \pi_{\theta_{\text{old}}}(y_t \mid x_{<t})$ corrects the sampling distribution but not this hidden-state-level mismatch, leaving an uncorrected source of policy-gradient bias that grows with both staleness and weight-update magnitude. We introduce Gradient-Coupled KV Drift (GCD), an exception-safe correction comprising (i) a high-probability concentration bound on the drift, with explicit per-layer terms for full-attention and Gated DeltaNet layers; (ii) a tail-event detector that flags inputs outside the bound’s regime of validity; (iii) a composite safety guarantee that treats the bias correction as trustworthy only when both the analytical estimate and the detector agree; and (iv) a variance-preservation theorem showing that GCD composes with the M2PO trust region without re-tuning. On Qwen3.5-9B trained with GRPO on MATH-500 and $\tau$-bench, GCD enables stable training at staleness 512 (a regime where M2PO collapses), reduces rollout wall-clock by 38\% versus full recomputation, and lifts $\tau$-bench retail pass$^4$ by 5.7 points over the M2PO baseline.

Many real-world coding challenges are open-ended and admit no known optimal solution. Yet, recent progress in LLM coding has focused on well-defined tasks such as feature implementation, bug fixing, and competitive programming. Open-ended coding remains a weak spot for LLMs, largely because open-ended training problems are scarce and expensive to construct. Our goal is to synthesize open-ended coding problems at scale to train stronger LLM coders. We introduce FrontierSmith, an automated system for iteratively evolving open-ended problems from existing closed-ended coding tasks. Starting from competitive programming problems, FrontierSmith generates candidate open-ended variants by changing the problems’ goals, restricting outputs, and generalizing inputs. It then uses a quantitative idea divergence metric to select problems that elicit genuinely diverse approaches from different solvers. Agents then generate test cases and verifiers for the surviving candidates, and these problems then join the seed pool for the next round. On two open-ended coding benchmarks, training on our synthesized data yields substantial gains over the base models: Qwen3.5-9B improves by +8.82 score on FrontierCS and +306.36 (Elo-rating-based performance) on ALE-bench; Qwen3.5-27B improves by +12.12 and +309.12, respectively. Moreover, FrontierSmith-generated problems elicit long-horizon code-agent behavior comparable to human-curated open-ended tasks, causing agents to spend substantially more turns and tokens than on closed-ended problems.

Engineering value is often created not by producing a single correct answer, but by improving a feasible artifact under hard constraints. We introduce **Frontier-Eng**, a benchmark for generative optimization agents that repeatedly edit executable engineering artifacts, receive frozen verifier feedback, and maximize the best feasible score found within a fixed budget. Frontier-Eng contains $47$ tasks across five engineering categories, spanning kernels, cryptography, quantum circuits, inventory and scheduling, robotics and control, optics and communication systems, structural design, reactions, and sustainable infrastructure. Each task provides a feasible starting solution, an editable artifact, and a read-only evaluator with verifier-parsed scoring, enabling cross-domain comparison without reducing engineering quality to binary pass/fail completion. We evaluate frontier models under OpenEvolve and compare search frameworks across the full task set. Current agents exhibit clear iterative improvement: **gpt-5.4** obtains the best average rank under OpenEvolve. Long-horizon runs reveal a dual **inverse-law pattern**: improvement events become rarer with iteration, and improvement magnitudes shrink with improvement count. At the same time, improvement is uneven across domains and remains sensitive to search depth, exposing a concrete frontier for engineering-oriented agents under realistic constraints and verifier-grounded feedback.

LLM-based coding agents are usually evaluated in familiar software settings: mainstream languages, common libraries, and public repositories. These benchmarks remain important, but they can hide how agents behave when the language itself is unfamiliar. We evaluate six contemporary coding agents on four esoteric programming languages using a sequential setup with file editing, local execution, and hidden-test grading. Our protocol exposes capability differences between these agents that mainstream coding and agentic benchmarks such as SWE-Bench Verified and Terminal-Bench 2.0 compress into much narrower bands. We observe that the strongest agents, Claude Opus 4.6 and GPT-5.4 xhigh, often avoid writing the target language directly. On Brainfuck and Befunge-98, they write Python programs that generate target-language code and debug those generators locally. Forbidding this metaprogramming strategy causes large performance drops. Text guidance distilled from this strategy does not materially improve weaker agents. In contrast, Opus-derived Python helper code for building generators, with no solved benchmark programs or hidden-test answers, sharply improves Sonnet 4.6 and GPT-5.4 mini on the same problems, while Haiku 4.5 remains low. More interpreter calls and output tokens improve stronger agents but leave weaker agents near their original performance, indicating that these resources amplify useful strategies rather than create them. Together, these results show that strong coding agents adapt to unfamiliar languages by using tools, feedback, and workspace state to build a working model of the target language. Metaprogramming is the clearest case, but the broader gap is constructing and debugging a strategy that works under the target language's rules.

Reasoning-oriented post-training enables large language models (LLMs) to solve complex tasks via multi-step inference. Supervised fine-tuning (SFT) on high-quality reasoning traces is a particularly efficient approach, but its effectiveness depends critically on careful data curation. Existing pipelines rely on a costly generate-then-filter paradigm: large pools of long reasoning traces are produced by multiple teacher models and subsequently filtered for difficulty and diversity using additional LLMs, often requiring weeks of computation. In this work, we propose a fundamentally different approach that bypasses this pipeline by directly predicting the quality of reasoning data. We show that a brief LoRA-based adaptation, combined with evaluating loss on only 0.1–1% of each reasoning trace, suffices to estimate both difficulty and diversity. Our method is grounded in a geometric view of post-training, where pretrained LLMs lie in a low-rank, anisotropic loss basin. By probing this basin via low-rank perturbations, we construct compact loss signatures whose clustering captures diversity, while loss at curvature inflection points provides a robust proxy for difficulty. We also provide a simple criterion for teacher selection. Empirically, our approach reduces data curation time from weeks to hours while maintaining or improving performance. Remarkably, on mathematical reasoning, 5K generated examples match the highly filtered 89K-example in OpenThoughts. We further introduce a high-quality physics reasoning dataset.

Machine learning (ML) methods have achieved notable success in enhancing individual components of mixed-integer linear programming (MILP) solvers. However, most existing approaches focus on single-component enhancement, neglecting critical interactions among components and often yielding diminishing returns when multiple approaches are integrated. The emergence of large language models (LLMs) provides a promising paradigm for jointly optimizing multiple solver components within a unified framework. In this paper, we propose Jolly-MILP, an LLM-guided framework to jointly optimize presolving, cut selection, and branching variable selection in exact MILP solvers. Our approach introduces a synergistic Coordinator to align component-level generators and a hierarchical Controller to produce adaptive, multi-stage strategies. Extensive experiments on nine MILP datasets demonstrate that Jolly-MILP consistently outperforms both ML-based approaches and existing LLM-based methods. These results show that explicit joint optimization, rather than isolated component-wise enhancement, is crucial for acquiring further gains in exact MILP solving, offering a new pathway for more advanced solver design.

Evaluating new large language models typically requires costly human annotation campaigns at scale. LLM-as-a-judge offers a cheaper alternative, but judge scores carry systematic errors — such as position bias, self-preference, or intransitivity — that can strongly miscalibrate the resulting rankings. We quantify the resulting judge--human disagreement at two complementary levels. At the local level, we estimate per-battle uncertainty from the judge's own score differences by propagating calibrated win probabilities rather than hard labels into the Bradley--Terry procedure. This alone provides a drastic improvement to Elo estimation accuracy, bringing LLM-derived ratings within $17.9$ Elo MAE of human-derived ones when averaged over 55 held-out models on LMArena. At the global level, we apply split conformal prediction to the residual gap between LLM-derived and human-derived Elo ratings across held-out models, producing prediction intervals with distribution-free marginal coverage guarantees that account for irreducible LLM--human disagreement. Together, these two layers yield a low-cost evaluation tool that provides developers with calibrated Elo estimates and honest uncertainty bounds, without access to large-scale human annotations.

The emergence of specialized, domain-tuned Large Language Models (LLMs) has demonstrated that smaller models can achieve expert-level performance in specific tasks, while struggling in out-of-domain settings. Current ensemble methods to combine their complementary expertise primarily rely on iterative re-prompting or cross-model refinement. These approaches suffer from high computational costs and latency because they require repeated LLM inference calls. Furthermore, naive aggregation often leads to anchor corruption, in which noise propagated from weaker models degrades the performance of the most accurate expert. To address these challenges, we propose a framework that integrates model predictions at the semantic layer using a bipartite factor graph. In this architecture, individual LLMs are represented as variable nodes, while a set of check nodes assess their consistency based on diverse epistemic criteria. We develop a message-passing protocol inspired by error-recovery systems to resolve disagreements iteratively. Furthermore, we introduce an asymmetric damping mechanism that protects high-reliability anchor nodes from being overridden by the ensemble majority. Unlike existing methods, our approach operates entirely on output distributions and requires no additional LLM calls during the refinement phase. Evaluating on four benchmarks, including MMLU, MMLU-Pro, GPQA, and MedMCQA, our method demonstrates a 97% reduction in token usage and up to a six-fold decrease in API calls, reducing inference time from several minutes to mere milliseconds while consistently outperforming leading multi-agent baselines. These results suggest that graph-based belief propagation is a robust, high-speed, and scalable alternative to the current multi-agent LLM systems. The full pipeline and code will be made public.

Recent work has shown that large language models form linear representations of emotion concepts that causally influence alignment-relevant behavior, including blackmail, reward hacking, and sycophancy. These findings raise a natural question: can such representations be used not only to characterize model behavior in controlled evaluations, but also as practical tools for monitoring and guiding models during deployment? We explore this possibility, investigating how emotion-vector activations can serve as interpretable signals of a model's internal state over the course of a conversation or agentic task. We discuss how these signals might support debugging, early detection of drift toward misaligned behavior, and lightweight steering interventions to nudge the model back toward desired behavior. Beyond the Assistant, the same representations apply to other characters in the model's context — including the user — suggesting a broader family of applications in which internal representations of emotional state inform how systems respond. We outline the opportunities and limitations of this approach and highlight open questions around validation, robustness, and responsible use.

Language agents increasingly improve by reusing \emph{skills}---structured procedural artifacts distilled from past experience. In particular, \emph{domain-level} and \emph{model-generated} skills are especially promising. They offer fast adaptation within a domain by encoding domain-specific recurring procedures, and they scale beyond labor-intensive hand-crafting. However, while extraction methods continue to proliferate, understanding remains limited, with no comprehensive study spanning the full skill lifecycle---\textbf{experience generation}, \textbf{skill extraction}, and \textbf{skill consumption}---to ask whether such skills actually work, when they work, and what makes them succeed or fail. To close this gap, we build a utility-grounded evaluation framework that provides systematic experimental results across extractors and target agents, covering five diverse agentic task domains. We find that model-generated skills are beneficial on average but exhibit non-trivial negative transfer, and that neither extractors nor targets behave uniformly. A model can be a strong extractor yet a weak consumer, or vice versa, with skill utility independent of model scale or baseline task strength. To explain these patterns, we then dissect each lifecycle stage in depth, analyzing how experience composition shapes skill quality, what properties characterize useful skills, and how the same skill transfers across different consumers. Finally, we translate these findings into a concrete \emph{meta-skill} that guides skill extraction toward the features tied to actual utility, which consistently improves skill quality across domains and substantially reduces negative transfer.

Supervised fine-tuning (SFT) on long teacher trajectories is the dominant method for instilling investigation and reasoning capabilities into open software-engineering (SWE) agents. Under SFT, every retained response is an imitation target, so the student inherits not only the trajectory's outcome but also any flaw in its intermediate steps, including ungrounded leaps and redundant loops. High-quality training data must therefore be jointly effective (each step is grounded and narrows the agent's epistemic gap to the correct fix) and efficient (each step is information-bearing rather than redundant or looping). Existing recipes filter or relabel teacher rollouts using only a binary terminal verifier, which does not directly target these axes and provides no supervision on instances where the teacher fails. Every real issue ships with a developer-authored reference patch $p^\star$ that implicitly testifies to the file paths, runtime behaviors, and conventions a fix presupposes, but the standard pipeline discards it. We propose P2T (Patches-to-Trajectories), which uses $p^\star$ as privileged information during curation, and frames trajectory construction as a bi-objective program over per-step effectiveness and trajectory length. A reverse phase distills $p^\star$ into a latent process graph $G^\star$ of contextual facts and solution milestones, encoding dense intermediate anchors in constructive ordering. A forward phase curate trajectories from blinded teacher continuations, scoring per-step progress against $G^\star$ under a leakage-blocking groundedness check and committing the shortest segments that retain effectiveness. Using only 1.8k curated SWE-Gym instances, P2T improves both axes simultaneously over outcome-filtered SFT and its tool-error-masking variant: on SWE-bench Verified, it lifts Pass@1 by up to +10.8 points while cutting per-instance inference cost by 15%, with consistent gains on SWE-bench Lite and across two teachers. A size-matched ablation and qualitative analysis further isolate per-trajectory quality from data scale.

Large language models (LLMs) are driving autonomous agents toward real-world environments, with the Web emerging as one of the most representative interaction scenarios. However, training Web agents with naive exploration remains challenging: successful trajectories are rare, and the sparse terminal rewards make it difficult for agents to obtain useful feedback from failed trajectories. Inspired by skill acquisition theory, we propose Skill-Composed Progressive Exploration (SCOPE), a short-to-long learning framework for long-horizon Web agents that decomposes complex task exploration into a gradual expansion process from local skills to complete task-solving capability. Specifically, SCOPE iteratively explores around local skills: Skill Discovery extracts reusable local skills and failure experiences from historical rollouts; Skill Orchestration composes discovered skills into high-level skill chains; andSkill Expansion leverages failure experiences aligned with the current skill chain to guide the agent toward exploring new subskills. Through this process, SCOPE progressively constructs task-completing skill chains across rollouts and internalizes them into the policy parameters through agentic on-policy distillation. Extensive experiments on WebArena, WebVoyager, and WebShop show that SCOPE consistently improves exploration dynamics and achieves 3.9\%--18.6\% relative gains over the strongest baselines in each setting. Our code is publicly available at https://anonymous.4open.science/r/SCOPE-14D3.

Large language model (LLM) agents are increasingly used to operate browsers, files, code and tools, making personal assistants a natural deployment target. Yet personal agents face a privacy-cost-capability tension: cloud models execute multi-step workflows well but expose sensitive intermediate context to external APIs, while local models preserve privacy but remain less reliable. Both settings also pay repeatedly for long skill prompts and growing histories. We propose constant-context skill learning, a context-to-weights framework for recurring agent workflows: reusable procedures are learned in lightweight task-family modules, while inference conditions only on the current observation and a compact state block. A deterministic tracker renders this state block from task progress and supplies aligned subgoal rewards, so each module can be trained with step-level SFT and refined through online RL. Across ALFWorld, WebShop, and SciWorld, our agents achieve strong performance across Qwen3-4B, Qwen3-8B and Llama-3.1-8B. With Qwen3-8B, SFT+RL reaches 89.6\% unseen success on ALFWorld, 76.8\% success on WebShop, and 66.4\% unseen success on SciWorld. They match or exceed strong published agent-training results while reducing prompt tokens per turn by 2-7$\times$ relative to controlled ReAct prompting baselines, showing that procedural context can be moved from prompts into weights.

Existing approaches to optimization modeling using large language models (LLMs) treat tasks through from-scratch construction, generating variables, objectives, and constraints without leveraging reusable structural knowledge. However, in real-world scenarios, expert modelers typically begin by identifying the canonical type or core structure of a problem, followed by iterative refinements based on specific task requirements. To incorporate this expert incremental logic, we propose a novel two-stage framework. In the first stage, we conduct Logic-Anchored Data Synthesis starting from 17 canonical prototypes, preserving critical bottleneck constraints and generating synthetic problem–formulation pairs. These pairs, together with the OR-Instruct dataset, are used for supervised fine-tuning (SFT) to initialize a policy. Separately, for each generated problem, we produce multiple modeling trajectories that follow a prototype-grounded order with variables preceding expressions. These trajectories are ranked according to criteria of accuracy and efficiency, and subsequently used to train a Logical Reward Model (LRM). In the second stage, we introduce Logic-Test-Time Group Relative Policy Optimization (Logic-TGRPO), which leverages the LRM during test-time reinforcement learning to reward accurate prototype identification and disciplined structural adherence while penalizing illogical patterns. Evaluated across seven benchmarks, our 8B parameter model achieves an average accuracy of 79.9%, outperforming comparably scaled baselines and competing effectively with heavyweight multi-agent systems with far fewer parameters. Strong performance on out-of-distribution tasks confirms that the model exhibits genuine incremental reasoning rather than mere memorization. Ablation studies further validate that both stages in our framework are essential.

Agents based on large language models are increasingly expected to autonomously handle negotiation and transactions. However, existing benchmarks predominantly focus on text-only, bilateral, zero-sum price haggling, failing to capture the complexity of real-world commerce. To address this gap, we introduce Agenticpay, a unified framework and benchmark for evaluating how well multimodal agents reach high-welfare, multi-dimensional agreements in realistic markets. Built around four core components ( Environments, Tasks, Agents, and Metrics), Agenticpay comprises 160 multimodal tasks spanning 4 real-world business scenarios (E-commerce, Food Delivery, Ride-hailing, and Apartment Rental) and 8 market topologies, scaling from 1-to-1 bargaining to many-to-many (N-to-N) competitive markets. Beyond price haggling, agents must read product images, infer each opponent's hidden preferences, and trade off multiple binding contract terms (e.g., price, lease duration, return policy, delivery speed). Agents communicate through multi-round natural-language dialogue, with each turn proposing or revising a full contract, and outcomes are scored by a utility-based framework that rewards agreements maximizing joint welfare. Evaluations on state-of-the-art proprietary (GPT-5.4, Claude Sonnet 4.6, Gemini 3.1 Pro Preview) and open-weight (Qwen3-VL-32B-Instruct, InternVL3-38B) multimodal models reveal substantial gaps in non-zero-sum value creation and market reasoning: even the strongest agent (Gemini 3.1 Pro Preview) reaches a GlobalScore of only 42.3/100, and performance consistently degrades as markets scale from bilateral to multi-sided, with the average GlobalScore dropping by 5.7 points and open-weight models suffering the largest declines (up to 11.5 points). These findings establish Agenticpay as a foundational testbed for multimodal agentic commerce. Code and dataset are available at: https://anonymous.4open.science/r/AgenticPay-4BBB/README.md.

Large language model (LLM) agents are highly vulnerable to memory-poisoning attacks. Existing defenses primarily rely on external modules for either static memory isolation or continuous online auditing, resulting in low memory efficiency and high computational overhead. To address these limitations, inspired by neuroscientific mechanisms that weak reactivation can destabilize memories and induce decay in the absence of reinforcement, we propose an agent-intrinsic training-free memory repair framework, \textbf{DREAM} (\textbf{D}ynamic \textbf{R}eactivation for \textbf{E}ngram \textbf{A}ttenuation in \textbf{M}emory), that enables selective functional forgetting of poisoned memories. Specifically, DREAM implements a three-stage pipeline: perturbation-induced implicit reactivation to reveal the structural fragility of malicious memories, adaptive anomaly diagnosis based on multi-dimensional activation patterns to detect topological anomalies, and a dynamic memory repair module that selectively suppresses harmful memories while reinforcing benign ones. Extensive experiments across diverse attackers and real-world tasks demonstrate that DREAM reduces attack success rates by over 95\% against backdoor poisoning while preserving strong benign utility. DREAM also achieves competitive robustness against injection attacks. In terms of runtime and token consumption, DREAM achieves up to a 2.13$\times$ speedup and a 37.7\% reduction in token consumption compared with A-MemGuard. Furthermore, DREAM also achieves a task success rate of 92.96\% on poisoned multi-agent systems.

Agentic forecasting is important for decision-making in dynamic environments, but it remains challenging because agents must reason from incomplete, time-limited evidence and produce calibrated probabilities before outcomes are resolved. Memory provides a natural mechanism for transferring experience from resolved forecasts to future prediction tasks. However, existing agent-memory methods are not tailored to forecasting, as they typically store past interactions, reflections, or factual associations without explicitly representing reusable predictive factors or calibration knowledge. We propose ForecastCompass(FoCo), an adaptive factor-based memory framework for agentic forecasting. FoCo organizes forecasting experience with a hierarchical forecasting-task taxonomy, enabling retrieval task-relevant forecasting knowledge. It maintains two complementary memory components: factor memory, which captures reusable predictive dimensions, and reasoning memory, which encodes probability updating, uncertainty handling, and calibration principles. Using retrospective analyses as learning signals, FoCo iteratively revises memory through a verbalized memory-revision procedure, enabling the agent to accumulate transferable forecasting knowledge over time. Experiments on Prophet Arena and FutureX with GPT-5-mini and Gemini-2.5-Flash show that FoCo improves both probabilistic accuracy and calibration.

LLM benchmarks are usually consumed as scalar leaderboards, where adjacent ranks are read as adjacent capability differences. We argue that this display should be complemented rather than simply trusted or discarded: benchmark scores are useful summaries, but a sorted score table can overstate which model pairs are actually distinguishable. This limitation is especially clear for accuracy- and success-rate leaderboards, whose averages hide task-level structure, and it also appears in pairwise systems such as Elo-style ratings when pairwise evidence is compressed into a total order. We propose LLM tier ranking as a task-level reporting framework for benchmarks. Given benchmark outputs, we construct pairwise wins, losses, and ties between models on each task and aggregate them into ordered tiers using Selective Preference Aggregation (SPA). Cross-tier comparisons are reported only when supported by sufficient task-level agreement; models in the same tier are left unresolved, not declared equal. Across AIME 2025, $\tau$-bench airline, xbench DeepSearch, GDPVal, SWE-Bench Verified, and Chatbot Arena, tier ranking often preserves useful broad capability classes while abstaining from unsupported fine-grained distinctions among frontier models. In Chatbot Arena perturbation studies, tier outputs are more stable than Elo-style total orders, expressing weaker evidence through later tier separation rather than rank reshuffling. We recommend reporting tiered leaderboards, solution paths, top-tier size, singleton-winner thresholds, and pairwise resolution statistics alongside scalar scores.

Cooperative multi-agent reinforcement learning (MARL) under sparse rewards remains fundamentally challenging because agents often fail to concentrate their influence, leading to insufficiently coordinated exploration. To address this, we propose the Focusing Influence Mechanism (FIM), a framework that encourages agents to focus their influence on under-explored parts of the state space through an entropy-based criterion, while leveraging eligibility traces to enable multiple agents to consistently align and sustain their influence on the same parts of the state space when beneficial, thereby promoting coordinated and persistent joint behavior. By emphasizing under-explored regions of the state space, FIM facilitates more efficient and structured exploration even under extremely sparse rewards. Across diverse MARL benchmarks, FIM consistently improves cooperative performance over strong baselines.