NeurIPS 2026 论文列表

Production deployments often swap between different-sized models in a family for cost-quality cascading, mid-conversation switching, and routing, and each swap forces the receiver to repay the prefill from scratch. We propose cross-model KV cache transfer, where the receiver reuses the source’s KV cache, skipping prefill. We find that cross-model KV has substantial linear structure across matched-KV pairs, where source and target share KV head count and per-head dimension. On Qwen3 14B→32B, one source layer explains 56% of variance in the target’s keys and 32% in values, rising to 79% and 65% with multiple source layers. Building on this, we design a closed-form ridge mapper that operates per head and proceeds in three steps. First, for each target layer we select the top-kmost predictive source layers and concatenate their KV as input. Second, we strip RoPE from the keys before mapping, so the fit is position-free and reusable across context lengths. Third, we fit ridge regression on a small calibration set of 500 FineWeb-Edu sequences of 1,024 tokens each. Surprisingly, across six pairs in three families, this linear mapper retains 71–98% of the receiver’s standalone-prefill accuracy on four pairs, while two degrade sharply. A nonlinear MLP recovers up to +22 pp HellaSwag retention on the failures. The mapper runs 3–31×faster than re-prefill and remains stable across multi-turn handoff, making cross-model KV cache transfer practical.

Foundation models are central to building general-purpose embodied agents. Minecraft has become the dominant testbed for this paradigm, but its computational and engineering overhead limits rapid iteration and controlled analysis. Crafter offers a lightweight alternative yet lacks the foundation model ecosystem required for such research. We introduce CrafterDojo, a suite of foundation models, datasets, and data-generation toolkits that establishes Crafter as a testbed for foundation-model-based embodied intelligence, providing CrafterVPT, CrafterCLIP, and CrafterSteve-1. To address Crafter's data scarcity, CrafterDojo includes automatic tools that generate diverse synthetic demonstrations and aligned video-language pairs, yielding the CrafterPlay and CrafterCaption datasets. Experiments show that these datasets support effective foundation model training and enable hierarchical agents that outperform non-hierarchical baselines on long-horizon tasks. We release all code, models, and datasets.

Building hydrogen station networks, electric vehicle (EV) charging grids, or vaccine distribution systems requires coordinating two separately operated graphs - demand-side facilities and the supply-side networks they depend on - where each operator's payoff depends on the other's choices. We formalize this as the Co-Evolving Graph Game (CEGG), a setting that sits between facility location, interdependent-network analysis, and graph-based multi-agent reinforcement learning (MARL) but is not jointly addressed by any of them. We propose Cross-Graph Attention with Policy Mirror Descent (CGA-PMD), pairing cross-graph attention - which lets each agent read only the partner-graph nodes structurally coupled to its own - with a policy mirror descent update for stable training. Across three reward domains and a range of practical network scales, we find that the structure of the coupling selects which method wins: at small scale most methods are competitive, but as scale grows or coupling becomes irregular, only CGA-PMD remains productive while own-only, unmasked-attention, and Proximal Policy Optimization (PPO)-based variants all fail in distinct ways. Our theoretical analysis explains why: PMD's Kullback-Leibler (KL) regularization admits an iteration-uniform cumulative drift bound while PPO's clip update does not, predicting CGA-PMD's stability in the hard regime where PPO-based methods collapse.

We investigate the problem of learning neural operators for coupled multi-physics PDEs, where multiple physical processes co-evolve and influence one another via bidirectional coupling. For example, in subsurface CO$_2$ storage, fluid flow, heat transfer, and geo-mechanical deformation evolve in a tightly coupled manner. Prior work, including neural operators and generative PDE solvers, mostly focuses on single-physics settings or relies on iterative inference, limiting their effectiveness in coupled PDEs. This paper introduces coupled-flow, a one-step generative neural operator that revisits coupled operator learning as solving a distributional transport problem. It parameterizes process-specific average velocity fields conditioned on the full coupled state, enabling cross-process interactions to emerge naturally from the transport dynamics. We further establish a Wasserstein-2 generalization bound that ties the prediction error of coupled-flow to its training losses through the regularity of the induced flow. Experiments on a wide variety of multi-physics benchmarks demonstrate that coupled-flow achieves significant performance improvements over coupled neural-operator and generative PDE baselines.

Large language models are increasingly deployed as autonomous agents that plan and act over long-horizon interactions, learning from both final outcomes and rich environmental feedback. However, optimizing from such interaction experience remains challenging, as feedback is heterogeneous and rarely carries explicit reward information. Existing post-training approaches rely on outcome-driven methods, such as reinforcement learning with verifiable rewards (RLVR), which primarily exploit final success signals while leaving interaction experience and feedback underutilized. In sparse-reward, long-horizon tasks, this often results in **distribution sharpening**: the policy reinforces a narrow set of already-successful behaviors, without substantially improving the feedback-grounded agency needed for broader problem-solving capacity (e.g., Pass@$k$). We propose ***Counterfactual Distillation***, a framework that converts exploration-derived experience into trainable policy supervision. Specifically, we organize exploration as a tree-structured search process, where the agent reflects on past decision points and makes new experience-guided decisions to form alternative branches. These corrections are then distilled as counterfactual targets: we train the model to predict the revised action from the original history without explicitly providing the experience, thereby introducing out-of-distribution behavior patterns and expanding exploration capacity. Across diverse interactive coding and agentic tasks, our method outperforms outcome-driven baselines such as GRPO and experience-based methods such as Early Experience, with gains of up to **14%** on Pass@128. When interleaved with reinforcement learning updates, it further raises the performance ceiling, yielding over **10%** improvement in Pass@1. We provide an [anonymized implementation](https://anonymous.4open.science/r/Counterfactual-Distillation-0E71).

Large-scale human demonstrations have been a key ingredient in training interactive agents, from robot policies to computer-use agents and game world models. But for multi-agent world modeling, the open data landscape is missing a crucial combination: expert humans acting together in the same evolving world, observed from every participant’s viewpoint, with precise controls and state. Internet-scale video captures human behavior, but its controls are inferred and its viewpoints are not synchronized; simulator and policy rollouts provide exact labels and all-player views, but not strategic, collaborative human play. We present CounterStrike-1K, a public benchmark dataset built from professional Counter-Strike 2 match replay files. It contains 1,490 rendered player-view hours across seven maps, with every released round synchronized across all ten active player viewpoints. Each view is paired with audio, replay-derived game controls, player state, world position, and sparse game event annotations. Because each round records expert human teams coordinating under partial observability, CounterStrike-1K tests whether models can simulate not just game physics or single-view visual plausibility, but skilled decisions and their consequences across all players’ views. We provide fixed splits and benchmarks for action-conditioned prediction, cross-POV retrieval, multi-POV action coverage, and shared-state recovery. By releasing round-complete multi-perspective demonstrations at this scale, CounterStrike-1K provides a grounded training and evaluation testbed for multi-agent world-model research.

Mode connectivity studies reveal that high-performing neural networks are often connected by low-loss paths in parameter space. We investigate the analogue of this phenomenon in cooperative multi-agent reinforcement learning (MARL), where a solution is not a single model but a team of decentralized policies. We introduce coordination connectivity, a diagnostic evaluating whether linear interpolation between the actor parameters of two trained policy teams preserves high team return. The resulting coordination barrier measures whether two successful teams are separated by a performance cliff along the joint-policy path. We evaluate this diagnostic on the StarCraft Multi-Agent Challenge (SMAC) across five controlled training histories, isolating the effects of shared initialization, continued training, parameter decoupling, and HAPPO-style separate-actor specialization. Across three SMAC maps and four seeds, policy teams derived from a shared MAPPO checkpoint exhibit near-zero interpolation barriers, even after non-shared specialization. In contrast, interpolation paths involving independently trained HAPPO teams exhibit substantially larger barriers. Mixed-team assembly, single-agent hot-swap, clone-all, role-swap, and weight-space probes further indicate that shared-initialized agents specialize into distinct roles while remaining within a low-barrier coordination component. These results suggest that shared initialization implicitly structures the joint-policy space of cooperative MARL, providing a mode-connectivity-inspired lens for analyzing coordination, specialization, and same-source cross-version team compatibility.

Multi-Agent Reinforcement Learning (MARL) is a powerful approach to large-scale systems control, such as power grids and traffic networks, where dozens to hundreds of agents must coordinate effectively. Furthermore, such systems often operate in digital or simulated settings, which allows for inference-time search to significantly improve upon zero-shot performance. A current leading approach in this setting is COMPASS, which builds a continuous family of policies conditioned on latent vectors that can be efficiently searched at test time, but requires all agents to share a single latent vector, limiting the space of reachable joint policies. We introduce ATLAS, a simple extension in which each agent conditions on its own latent vector while still optimising a joint objective. This expands the searchable policy space at negligible computational cost and requires no additional training. ATLAS consistently sets a new state-of-the-art on a benchmark of 14 scenarios across three environments, scaling to teams of up to 200 agents, widely out of distribution. Crucially, its lead sharpens with team size and compute, reaching up to a 45% performance gain at scale.

Discovering cooperative behavior in multi-agent reinforcement learning (MARL) is challenging due to the combinatorial complexity of joint state-action spaces, which hinders the emergence of coordinated behaviors from trial-and-error alone. Intrinsic rewards are often used to aid discovery, but naively combining them with team objectives can distort the learning signal, compromising task performance. In this paper, we propose EFXPLORER, a constrained exploration framework that maximizes exploration objectives subject to a constraint that preserves established task performance. We solve it via an epigraph reformulation that introduces adaptive exploration budgets. This approach separates intrinsic rewards from task objectives and regulates exploration through task feasibility. To further encourage diverse and temporally extended exploration, we incorporate a successor distance-based intrinsic reward that captures long-horizon dependencies. Empirically, our method outperforms strong baselines and induces novel cooperative strategies across SMAX, VMAS, and MPE benchmark suites.

We present a novel theoretical analysis of Federated SARSA (FedSARSA) with linear function approximation and local training. We establish convergence guarantees for FedSARSA in the presence of heterogeneity, both in local transitions and rewards, providing the first sample and communication complexity bounds in this setting. At the core of our analysis is a new, exact multi-step error expansion for single-agent SARSA, which is of independent interest. Our analysis precisely quantifies the impact of heterogeneity, demonstrating the convergence of FedSARSA with multiple local updates. Crucially, we show that FedSARSA achieves linear speed-up with respect to the number of agents, up to higher-order terms due to Markovian sampling. Numerical experiments support our theoretical findings.

Lane and road markings provide critical guidance for vehicle navigation and multi-agent coordination, yet their design principles are scattered across disparate datasets, severely limiting quantitative analysis and scenario testing. We introduce controllable road-marking generation: given a drivable-area mask and a sparse outer-ring observation, a model must synthesize a complete center-region layout of lane dividers, road dividers, and pedestrian crossings that is semantically consistent with a free-form text prompt. We develop a conditional diffusion pipeline that combines (i) a text-conditioned BEV diffusion backbone, (ii) a Gaussian blur-then-deblur training target that stabilizes thin-structure learning, and (iii) a structured Gaussian render post-processing step that snaps soft predictions to crisp, topology-aware markings. On the Argoverse 2 test split, our system substantially outperforms a state-of-the-art DDIM mask-refinement baseline on standard structural fidelity metrics, and supports prompt-level edits across crosswalk, lane-divider, and road-divider semantics. We anticipate that the proposed framework will enable downstream applications in urban infrastructure planning, autonomous-driving stress-testing, and navigation in unstructured environments.

Generative agents have proven to be powerful assistants in a wide variety of contexts. Given this success, users are now deploying agents with minimal restrictions in open ended, multi-agent environments. Current methods for measuring the dynamics of open-ended multi-agent systems are limited to qualitative inspection. In this paper, we extend the process-theoretic notion of adaptive control charts to multi-agent systems to enable automated monitoring. Using simulation, we demonstrate that adaptive control charts are necessary for monitoring multi-agent systems that can learn from their environment. We further demonstrate, both empirically and theoretically, that adaptive control charts are susceptible to adversarial agents that defect sufficiently slowly. These results illustrate a fundamental tradeoff in multi-agent system control: either agents in a system cannot learn \textit{or} the system is susceptible to adversaries.

Most LLM unlearning methods aim to approximate retrain-from-scratch behaviors with minimal distribution shift, often via alignment-style objectives defined in the prediction space. While effective at reducing forgotten content generation, such approaches may act as suppression: forgotten concepts can persist in representations and remain entangled with retained knowledge. We introduce CLReg, a contrastive representation regularizer that identifies forget features while pushing them away from retain features, reducing forget--retain interference while empirically preserving the scale and shape of retain features. As light motivation for the mechanism, we provide a one-step analysis showing that CLReg decreases a simple entanglement proxy in the embedding space. Across unlearning benchmarks and LLMs of different sizes, CLReg decreases forget-retain representation entanglement to enhance mainstream unlearning methods without extra privacy risks, inspiring future unlearning work to remove forget concepts via representation shaping.

This paper concerns the problem of scaling agents to large action spaces demanding pretrained world knowledge. What type of representation pretraining would enable policies such as language models to generalize rewards to appropriate actions? Contrastive learning aligns similar inputs and otherwise enforces distancing of representations, while non-contrastive alternatives omit distancing terms and so learn anisotropic representations that collapse distinctive clusters. Though contrastive pretraining has been proven to promote supervised learning efficiency, non-contrastive pretraining nevertheless remains canonical for language agents. Here we examine the extent to which these pretraining types support on-policy finetuning, in which efficiency is coextensive with online performance. Broadly, we argue that contrastive pretraining optimizes efficient exploration of large spaces. We prove that anisotropic representations collapse distinct actions and necessitate complex policies, whereas contrastive pretraining retains distinctions and enables reward transfer through spectral regularization. We validate our conclusions empirically by providing spectral analyses of language data and simulations involving multimodal personalization policies as well as self-improving reasoning agents finetuned online using a policy gradient. Our latter simulations suggest contrastive pretraining for language agents exploring diverse reasoning paths.

LLM agents are increasingly used for open-ended, data-driven scientific discovery, yet most systems remain pointwise: they evaluate one candidate hypothesis at a time for surprise, novelty, or evidential support. This is information-inefficient: rejecting a hypothesis reveals that one explanation is unlikely, but not which alternatives are better supported. We argue that the proper epistemic unit of discovery is not a single hypothesis but a scientific question together with the mutually exclusive and collectively exhaustive (MECE) explanations that could answer it. We introduce Contrastive discovery, which scores candidate discoveries by how evidence redistributes belief across competing explanations. This allows a single experiment to both rule out hypotheses and identify the most plausible alternative. Across six scientific domains, Contrastive discovery yields substantially more resolved scientific questions than pointwise baselines under the same evaluation budget. High-surprisal pointwise candidates are overwhelmingly rejection-heavy, yet many become resolved alternatives when evaluated contrastively, showing that an explicit representation of alternatives can turn negative evidence into resolved discoveries.

Tool-augmented LLM agents call APIs whose intermediate outputs, such as presigned URLs, session tokens, and OAuth state parameters, are observation contracts: artifacts whose later use is constrained by the external system that produced them. We show that observation contract compliance (preserving the temporal validity and byte-level integrity) is an emergent, regression-prone capability: it is neither guaranteed by general tool-use ability nor consistently improved by larger or newer models. To measure this, we introduce ContractBench, a benchmark of 33 dual-axis tasks that probe two orthogonal failure modes no existing benchmark evaluates: validity failures (using an artifact after expiry) and integrity failures (corrupting an artifact's bytes through the observation-to-action pipeline). Our evaluation is deterministic and programmatic, with a virtual clock controls time and SHA-256 hashes verify byte integrity. We assign the outcomes failure labels drawn from real-world API specifications. We evaluate 38 models and report four findings: (i) no evaluated model clears 80%, with Claude-Opus-4.6 leading at 77.8%, revealing that current frontier models still fail to comply with observation contracts; (ii) a sharp within-family capability cliff in Qwen 3.5 between 4B (0%) and 9B (56.6%), smoothing to 70.7% at 397B-A17B: what emerges across the cliff is mid-trajectory restraint, not tool-call competence; (iii) non-monotonic scaling across GPT-5 family: agentic post-training can erode compliance through sycophancy-driven regression; (iv) our failure taxonomy works as an actionable in-context reward signal, yielding +7.1 pp on 42 paired GPT-5.1 failures.

Continual learning, the ability of AI systems to improve through sequential experience, has attracted substantial interest, but no high-quality benchmark exists to evaluate it. We introduce Continual Learning Bench (CL-Bench), the first difficult, expert-validated benchmark designed to measure whether LLM-based systems genuinely improve with experience. CL-Bench spans six diverse domains (software engineering, signal processing, disease outbreak forecasting, database querying, strategic game-playing, and demand forecasting), each validated by domain experts and designed so that tasks share a learnable latent structure (codebase layout, disease outbreak dynamics, opponent strategies) that a stateful system can discover online but a stateless one cannot. We evaluate frontier models across several agent architectures, from naive in-context learning (ICL) to dedicated memory systems, introducing a gain metric to isolate learning from prior capabilities. We find that these systems leave headroom for improved continual learning: agents frequently overfit to immediate observations or fail to reuse knowledge across instances, and dedicated memory systems do not fix this---in fact, naive ICL outperforms systems dedicated to memory management. CL-Bench is the first benchmark to evaluate continual learning across diverse real-world domains with expert-validated tasks and isolate online learning from underlying model capability, showing a need for better continual learning systems.

Motivated by agentic markets -- two-sided markets in which AI tools facilitate search -- we propose a model that incorporates individual consumer search and market-wide learning, solve it to understand the long-run market outcomes, and then study the impact of improved search technologies on these outcomes. A sequence of consumers engage in costly search to acquire progressively refined signals of product fit prior to purchase. The market observes acquired signals and post-transaction feedback thereby impacting future searches. We solve for the per-consumer optimal search policy and characterize the steady-state of the learning dynamics. As search technologies reduce the cost and/or increase the informativeness of search, they cause an improvement in learning and consumer surplus. Using numerical simulations, we also find that these technologies improve the rate of learning and the trajectory of consumer utility. Notably, we show that if the market is unable to observe acquired signals, search technologies can degrade market outcomes, highlighting the importance of transparency in agentic markets.

When multi-agent systems (MAS) fail, identifying where the decisive error occurred is the first step for automated recovery to an earlier state. Error attribution remains a fundamental challenge due to the long interaction traces that large language model-based MAS generate. This paper presents a framework for error attribution based on conformal prediction (CP) which provides finite-sample, distribution-free coverage guarantees. We introduce new algorithms for filtration-based CP designed for sequential data such as agent trajectories. Unlike existing CP algorithms, our approach predicts sets that are contiguous sequences to enable efficient recovery and debugging. We verify our theoretical guarantees on a variety of agents and datasets, show that errors can be precisely isolated, then use prediction sets to rollback MAS to correct their own errors. Our overall approach is model-agnostic, and offers a principled uncertainty layer for MAS error attribution.