NeurIPS 2026 论文列表

Long-horizon planning often requires composing familiar local transitions while remaining grounded in a specified goal. We study this setting from single-stage data: on a controlled no-skip graph benchmark, training provides only adjacent-stage demonstrations while evaluation requires composing them on unseen long-horizon start–goal pairs; an event-chain diagnostic decomposes success into legality, stage-bridging, and goal-conditioned termination. Across a minimal Transformer, Qwen2.5-3B, and an external validation on MuSiQue multi-hop QA, supervised fine-tuning (SFT) learns local transitions but degrades with horizon, with failures localized to stage-bridging. Reinforcement learning (RL) narrows this composition gap, especially at longer horizons, but unanchored RL drifts—producing valid stage progress while ignoring the instructed target. Stable gains require two complementary forms of anchoring: instruction-style prompting strengthens prompt-side goal-conditioning, especially at shorter horizons, whereas KL regularization to the SFT reference policy prevents RL-induced drift across horizons; reward shaping mainly provides denser credit. Finally, we identify a Decomposition Paradox: models trained only on long compositions can reach intermediate targets yet fail the corresponding prefix subtasks because stopping at instructed intermediates is never supervised.

LLM agents are flexible in language interaction, but using them as long-horizon planners is costly and prone to hallucinated or invalid actions. This paper targets a different role for LLMs: constrained bridges between language and symbolic planning. We propose MemPlan, a memory-conditioned PDDL planning framework for partially observable interactive text environments. MemPlan keeps the PDDL domain fixed and reconstructs the PDDL problem at each step from parsed observations, candidate hypotheses, and interaction memory. Positive graph memory assigns relevance-based action costs to candidate hypotheses for missing objects, locations, or conditions, guiding the planner toward more plausible checks. Negative memory records execution-refuted assumptions and compiles them into exclusion constraints, preventing repeated attempts at candidates already contradicted by feedback. A unified schema-conditioned language--symbol interface, trained offline with supervised fine-tuning, connects text environments to the planning loop through observation-to-fact parsing and symbolic-action-to-command grounding. Experiments on TextWorld, ALFWorld, and Robotouille show that MemPlan improves task completion and step efficiency over language-only and PDDL-aware baselines, while substantially reducing online token cost compared with LLM-planning baselines such as ReAct. The same fine-tuned interface is also reused across TextWorld task types without task-specific fine-tuning. Our code and data are available at: \url{https://anonymous.4open.science/r/MemPlan-1EE5/}.

We study the problem of multiclass classification within the framework of *marginal-nonuniform* learning, where the constants in the learning rate may depend on the unknown marginal distribution generating the unlabeled data, but not on the target concept. This notion naturally lies between classical PAC learning, where the constants are independent of both the marginal distribution and the target concept, and more recent universal learning, where the constants may depend on both. We give a complete characterization of the optimal marginal-nonuniform learning rates, establishing a fundamental trichotomy: exponential, linear, and arbitrarily slow rates. This characterization is witnessed by a new combinatorial dimension, termed the *uncentered DS-eluder dimension*. Technically, a key ingredient in the upper-bound proofs is a set of new *distributional exhaustion* lemmas that exploit the sequential nature of this dimension and may be of independent interest.

Finding the optimal sample complexity of realizable PAC learning was a major open problem in learning theory, which was settled by the breakthrough result of Hanneke (2016), building on prior work by Simon (2015). Hanneke's algorithm is quite involved and requires training $\text{poly}(n)$ Empirical Risk Minimizers (ERMs) on carefully crafted subsets of the data. Since using a single ERM call is provably suboptimal, recent work by Aden-Ali (2024) proposed the majority-of-three ERMs as a candidate for the *simplest* optimal PAC learner, and established its optimality in the *in-expectation* regime. Their main open question was whether this algorithm is optimal in the classical *high-probability* regime. In this work, building on their ideas, we answer this question affirmatively.

Vision-Language Models (VLMs) are increasingly adopted as end-to-end driving planners, yet they suffer from cross-city domain shift. Existing methods treat this shift as a single quantity addressed by uniform alignment or single-site editing. We refute this view on two counts: ***(i) domain shift in VLM trajectory planners is inherently structured across modality and layer, and (ii) a locate-then-repair recipe matched to this structure outperforms every uniform-objective baseline***. We analyze domain shift via activation patching and uncover two regularities. **(F1) Modality axis.** Image and text tokens exhibit distinct shift patterns in each VLM. **(F2) Layer axis.** The layer-wise shift partition is architecture-determined and dataset-stable. These findings recast domain shift as a **two-dimensional tensor** indexed by *modality* (image vs. text tokens) and *layer*, motivating a simple principle: ***locate where shift occurs, and repair only there***. We instantiate this principle in **MoLaRx** (**Mo**dality–**La**yer **R**epair), a locate-then-repair framework. The Causal Importance Score (CIS) estimates the domain shift as a modality–layer tensor and partitions the layers into functional zones. CIS-Guided Selective Repair (CISR) then applies a matched operator (Maximum Mean Discrepancy (MMD) / low-rank adaptation (LoRA) / freezing) to repair each zone, with the composition adapting automatically to each architecture. We evaluate **MoLaRx** on three architecturally distinct VLMs (LFM2-VL-1.6B, Qwen2-VL-2B, and PaliGemma2-3B) across two cross-city benchmarks (nuScenes and Argoverse 2). MoLaRx attains the lowest cross-city trajectory-error gap on every (VLM, dataset) cell, while using $50$–$60\%$ fewer trainable parameters and $8.4$–$10.9$ ms lower inference latency. In contrast, the strongest scalar baseline (MMD-only) collapses on every cell—direct evidence that no scalar objective can repair structured shift. Code: [anonymous.4open.science/r/MoLaRx-E3B4](https://anonymous.4open.science/r/MoLaRx-E3B4/).

Value iteration (VI) is the unit-step explicit-Euler discretization of a continuous-time *Bellman residual flow* $\dot V=\operatorname{BR}(V)$, and its standard linear dependence on the effective horizon $1/(1-\gamma)$ can be viewed as reflecting the step-size restriction that explicit integration imposes on stiff dynamics. We propose *Implicit-Euler Value Iteration* (IE-VI): a planning routine equivalent to the implicit-Euler discretization of the same flow, in which each step is unconditionally stable and contracts at rate $1/(1+h(1-\gamma))$ for every step size $h>0$. The implicit step requires solving an equation in the true Bellman operator, which we realize at finite cost by *iterative defect correction*: a short sequence of cheap planning rounds in an approximate model $\hat{\mathcal{P}} \approx \mathcal{P}$ (a simulator or a learned dynamics model), with the true model queried once per round to refresh a residual correction term. For Policy Evaluation and Control, IE-VI converges to the value function of $\mathcal{P}$ for any $\hat{\mathcal{P}}$, however inaccurate---in contrast to model-based planners, which generally converge to a biased fixed point. When the model error shrinks at a sublinear-in-horizon rate, IE-VI attains polylogarithmic iteration complexity in $1/(1-\gamma)$, an exponential improvement over VI; outside this regime its complexity stays within a logarithmic factor of VI. We also give an adaptive variant A-IE-VI, which removes the need to know the model-error scale, and a Dyna-style sample-based variant IE-Dyna for tabular RL.

World models enable model-based planning through learned latent dynamics, but imagined rollouts often become unstable as the planning horizon grows or the dynamics distribution shifts. We argue that this instability arises from two missing structures in planner-facing latents: history-conditioned memory for approximate Markov completeness, and geometric organization that separates configuration, momentum, and task semantics. We propose HaM-World, a structured world model that decomposes the latent state into a canonical $(q, p)$ subspace and a context subspace $c$, while incorporating Mamba selective state-space memory as a history-conditioned input to the same latent dynamics. Within this unified interface, $(q, p)$ evolves under a Soft-Hamiltonian dynamics composed of an energy-derived Hamiltonian vector field and learnable residual and control dynamics, while $c$ captures semantic, dissipative, and non-conservative factors. This design provides the planner with a single latent representation shared across dynamics prediction, reward and value estimation, imagined rollouts, and cross-entropy method (CEM) planning. On four DeepMind Control Suite tasks, HaM-World achieves the highest average AUC (+9.5% over strong baselines), reduces long-horizon rollout error to 45% of a competitive model, and wins 11 out of 12 $k \in \{3,5,7\}$ rollout MSE metrics. Under 12 out-of-distribution perturbations, HaM-World consistently attains the highest returns, with average gains of 10.2% on Finger Spin and 13.6% on Reacher Easy. Mechanism diagnostics further demonstrate bounded energy drift under action-free rollouts, structured energy variation under policy control, and coherent control-induced energy transfer, supporting the effectiveness of the proposed Soft-Hamiltonian latent dynamics. Code: https://anonymous.4open.science/r/HaM_World-47CD

Temporal point processes (TPPs) are the canonical framework for event sequence data, with applications spanning seismology, social media, and electronic health records. Recent neural extensions have been reported to substantially outperform classical baselines such as Hawkes processes, but a systematic comparison across model families and intensity parameterizations is lacking, and large-scale medical benchmarks — arguably one of the most promising application domains — have not been considered. Here, we revisit this comparison through a systematic benchmark of neural and non-neural TPPs on existing datasets and on two large-scale medical benchmarks derived from UK Biobank disease histories and MIMIC-IV ICU records. First, our results show that the previously reported gap between Hawkes processes and neural TPPs largely disappears once Hawkes processes are equipped with flexible and learnable kernels. Second, we find that the flexibility of the intensity parameterization also limits prior neural TPPs: a transformer with a spline-based intensity head outperforms prior neural TPP implementations. Third, our benchmark underscores the value of complex datasets when assessing the performance of TPPs. In fact, exclusively on the MIMIC-IV ICU records did we find clear evidence for structures — time-varying or higher-order interactions — beyond what pairwise flexible Hawkes process kernels can capture. Our results guide both future method development and practitioners’ model choice. Our TPP framework is released as an open-source package, providing consistent and feature rich implementations of classical and neural TPPs.

Latent world models based on Joint-Embedding Predictive Architecture (JEPA) are deterministic by design. While successful in fully observable scenarios, this paradigm breaks down when past observations and actions lead to multiple plausible future possibilities, e.g., due to occlusion. We introduce EpicWorldModel, a framework to train stochastic JEPAs for environments and tasks with inherent uncertainty under partially observability. We jointly train the EpicWorldModel predictor with its latent representation space to directly predict multiple potential future states using a flow-matching objective, when the goal-relevant scene content is absent from the conditioning history. We show that flow predictive variance approximates an upper bound on Expected Information Gain and serves as a useful exploration guidance for planning. By incorporating this uncertainty signal into Cross-Entropy Method (CEM)-based planning, our approach balances goal-reaching with exploration of uncertain regions where occluded or unseen goals are most likely to be located. We demonstrate the effectiveness of EpicWorldModel through a series of latent planning experiments with the best performance among all baseline methods, showing up to $22\%$ empirical improvement in success rate over LeWorldModel.

Discrete diffusion models are often said to outperform autoregressive models on planning tasks, but the specific mechanisms enabling this are not yet well understood. To explain why, we abstract the complexity of studying ``planning'' into graph traversal problems, which offer a minimal setup for evaluating language models' planning and lookahead capabilities. We find that non-autoregressive language models are able to leverage an inherent directional asymmetry in lookahead planning tasks to outperform the training efficiency of their autoregressive counterparts. Through a mechanistic study of the training and inference dynamics of autoregressive and non-autoregressive models, we discover that while forward traversal through junctions in the graph requires complex sequential planning, the reverse path remains deterministic. Consequently, non-autoregressive models naturally develop a fundamentally different solution strategy. Effectively, they learn a reverse-decoding pattern which reduces an $\ell^{th}$-order lookahead problem to a sequence of simple $1^{st}$-order transitions. Latent-space analysis further confirms these divergent strategies by demonstrating the distinct internal representations formed during generation. Ultimately, these findings highlight the advantages of non-autoregressive modeling while clarifying the inherent planning limitations induced by autoregressive architectures.

Offline goal-conditioned reinforcement learning (GCRL) learns goal-directed policies from reward-free data, but in long-horizon tasks, goal-conditioned value functions often provide unstable guidance due to sparse rewards and discounting. Hierarchical methods partially mitigate this issue via subgoal decomposition; however, high-level decision-making still relies on noise-sensitive value estimates, leading to unstable behavior in complex environments. We address this limitation by proposing \textbf{D}iffusion \textbf{S}ubgoal \textbf{P}lanning (\textbf{DSP}), a diffusion-based framework for high-level subgoal generation. DSP casts high-level planning as guided generative inference over goal-conditioned subgoals and learns both conditional and unconditional flows, enabling classifier-free guidance to introduce a goal-directed bias at inference time. By removing explicit value-based guidance from high-level planning, DSP generates reachable and goal-directed subgoals through a generative model while retaining hierarchical execution. Experiments on offline GCRL benchmarks demonstrate that DSP outperforms prior methods on a range of navigation and manipulation tasks, with particularly strong performance in maze environments that require multi-step subgoal planning.

Classical planners can effectively solve very large deterministic MDPs represented in STRIPS or PDDL where states are sets of atoms over objects and relations, and lifted action schemas add or delete these atoms. This compact representation yields strong search heuristics and provides an ideal setting for structural generalization, since lifted relations and action schemas give rise to infinitely many domain instances. A central challenge is to learn these relations and action schemas from data, and recent approaches have addressed this problem using different types of observations. In this work, we develop a novel neural network architecture for learning action schemas from traces where states are fully observed but action arguments are unobserved. The problem is a simplification but an important step towards learning planning domains from sequences of images and action labels, and we aim to solve this simplification in a nearly perfect manner. The challenge lies in learning the action schemas while simultaneously identifying the action arguments from observed state changes. Our approach yields a robust differentiable component that can then be integrated into larger neuro-symbolic models. We evaluate the architecture on various planning domains, where the learned lifted action schemas must recover the ground-truth structure. Additionally, we report experiments on robustness to observation noise and on a variation related to slot-based dynamics models.

Deep research tasks require agents to navigate large knowledge spaces, synthesize evidence across many sources, and adapt their plans as intermediate findings emerge. Directed acyclic graph (DAG)-based multi-agent systems are well suited to this setting because they support parallel execution and isolate each sub-task within a focused dependency context. Yet existing DAG-based agents typically instantiate a task-level plan before execution and then repair the graph only after failures or missing evidence are observed. This Plan-then-Patch strategy is brittle for deep research: the system commits most strongly when its evidence is weakest, and later revisions often waste computation on branches that should not have been planned in the first place. To bridge this gap, we propose DAGent, a DAG-based multi-agent framework that introduces Evaluate-then-Grow incremental planning. Instead of committing to a full DAG upfront, an Orchestrator grows the task graph one batch at a time, conditioning each new expansion on confidence and uncertainty signals from completed nodes. To support long-horizon evidence use without overloading each sub-task, DAGent maintains a hierarchical context layer that propagates compact QueryDocs by default while preserving full execution traces for on-demand recall. The cleanly recorded DAG topology in turn admits structural RL signals that an outcome-only recipe cannot define; we instantiate this with DAGRPO, a GRPO adaptation that injects topology-conditioned credit on Executor rollouts and a structural compliance regularization on Orchestrator plans. Across BrowseComp-Plus, GAIA, and xbench-DeepSearch, DAGent surpasses the strongest open-source baseline by 5.3 / 5.8 / 2.0 points at the Qwen3-235B-A22B scale, with the lead replicating across four open-source backbones from four vendors and extending to GPT-5 at 327K context. At the Qwen3-8B scale, DAGRPO further improves over a same-budget outcome-only GRPO baseline by 3.0 average Pass@1 points. A same-architecture comparison further shows that incremental, evidence-conditioned planning recovers accuracy at lower per-task token, tool-call, and step footprints than its Plan-then-Patch counterpart, supporting the view that DAGent's gains come from more targeted evidence expansion rather than additional computation. Anonymized code is available at \url{https://anonymous.4open.science/r/DAGent-F2D6}.

Chain-of-thought (CoT) monitoring is increasingly relied upon to detect misbehavior in LLMs, under the implicit assumption that misbehaving actors must reason about their misbehavior, leaving detectable traces in the CoT. We show this assumption fails for a class of attacks we call planning-execution decoupled attacks, in which misbehavior is injected via a corrupted plan before the actor begins reasoning. Using the investigator agent framework of Li et al. [2025], we discover that exposing an actor to a corrupted reasoning plan upstream causes it to reproduce the flawed reasoning in its own CoT, embedding misbehavior into natural-looking reasoning with few suspicious traces. The attack scales to stronger reasoning models and harder tasks, steering them toward misbehavior while evading detection. Evaluating a suite of thinking and non-thinking monitors, we find that thinking monitors substantially outperform non-thinking ones, though even the best miss a meaningful fraction of attacks. Critically, the relationship between monitor thinking budget and detection is not monotonic: extra reasoning sometimes improves detection but can also hurt it when monitors talk themselves into accepting corrupted CoTs as benign. This refines Guan et al. [2025], who show thinking budget generally helps monitoring; we find detection also depends on whether the additional reasoning is directed toward critical evaluation rather than rationalization.

We study the degree-weighted work required to compute $\ell_1$-regularized PageRank using the standard one-gradient-per-iteration accelerated proximal-gradient method (FISTA). For non-accelerated local methods (ISTA), the best known worst-case work scales as $\widetilde{O} ((\alpha\rho)^{-1})$, where $\alpha$ is the teleportation parameter and $\rho$ is the $\ell_1$-regularization parameter. A natural question is whether FISTA can improve the dependence on $\alpha$ from $1/\alpha$ to $1/\sqrt{\alpha}$ while preserving the $1/\rho$ locality scaling. The challenge is that acceleration can break locality by transiently activating nodes that are zero at optimality, thereby increasing the cost of gradient evaluations. We analyze FISTA with respect to a less regularized objective and show that, under a checkable confinement condition, all spurious activations remain inside a boundary set $\mathcal{B}$. This yields a bound consisting of an accelerated $(\rho\sqrt{\alpha})^{-1}\log(\alpha/\varepsilon)$ term plus a boundary overhead $\sqrt{vol(\mathcal{B})}/(\rho\alpha^{3/2})$. We provide graph-structural conditions that imply such confinement. Experiments on synthetic and real graphs show the resulting speedup and slowdown regimes under the degree-weighted work model.

Diffusion models provide strong priors for generating structured data, but many tasks require outputs beyond the scale on which these models are typically trained. Compositional generation addresses this by composing overlapping local plans from a pretrained short-horizon prior into a long-horizon output. However, standard composition primarily enforces agreement between neighboring local plans, yielding local consistency without directly specifying the global structure of the full composition. As a result, locally compatible plans may still form an implausible route, task sequence, or temporal evolution. Existing methods improve global coherence by repeatedly propagating local consistency signals or by adding inference-time optimization, but these procedures become expensive as the number or dimensionality of local plans increases. We propose Coarse-to-Fine Compositional Diffusion (CoFi), an inference-time sampler that separates global structure formation from local detail refinement. CoFi first aligns local denoised estimates around a shared coarse structure, producing a global scaffold that captures the long-range task-level arrangement. It then diffuses this scaffold to an intermediate noise level and denoises it with the same pretrained local prior, restoring local fine structure while preserving the scaffold-induced global coherence. Across long-horizon robotic planning, panoramic image generation, and long video generation, CoFi not only improves both global coherence and local sample quality over prior compositional baselines, but also requires $2$--$8\times$ fewer denoiser evaluations.

Monte-Carlo Tree Search (MCTS) typically stores only scalar mean estimates, which makes it poorly suited for tail-risk objectives such as conditional value-at-risk (CVaR). We introduce **Categorical Thompson Sampling with Optimism (CATSO)** and **Particle Thompson Sampling with Optimism (PATSO)**, two distributional MCTS algorithms that maintain finite-support empirical backup laws at Q-edges. CATSO represents Q-edge laws with categorical atoms and Dirichlet Thompson sampling, while PATSO represents them with adaptive particles. Both algorithms select actions using the CVaR of a Thompson-sampled Q-edge law plus a polynomial optimism bonus, and propagate scalar continuation values through visit-weighted averages of child Q-edge CVaR scores. For the finite-depth nested CVaR planning target, we prove an $O(n^{-1/2})$ bound on the root value error after $n$ rollouts, which upper-bounds the risk-sensitive simple regret. CATSO adds a fixed-grid discretization bias, while capped PATSO adds a tunable Wasserstein compression bias. Experiments on stochastic planning benchmarks show that distributional Q-edges improve lower-tail return and reduce catastrophic outcomes when mean-optimal and risk-sensitive behavior differ.

We trace factual sycophancy across three post-training pipelines (OLMo 3 7B Think and Instruct, Llama 3.1 8B Instruct, Tulu 3) on AMPS and MedQuAD with a dual-track evaluation: a GPT-4o-judged generative track and a log-probability track reporting $\Delta$LogOdds separately under \emph{in-context} rebuttals and \emph{preemptive} wrong-answer assertions. Decomposing by challenge type $\times$ context reveals three patterns. (1) \textbf{Challenge type:} the sycophantic shift is alternative-conditioned, so simple pushback with no asserted answer produces no shift, while ethos/justification/citation challenges produce substantial positive shifts. (2) \textbf{Context:} every base model's sycophantic shift concentrates preemptively, with weak or defensive in-context response. On computational, IC diverges into four pipeline-specific endpoints. On medical, the same four pipelines converge with no defensive IC developing on any recipe. (3) \textbf{Behavioral vs log-probability:} post-training produces a \emph{surface-predictive dissociation}: matched-subset behavioral flip rate drops significantly on identical items while preemptive $\Delta$LogOdds on those same items grows.

Material synthesis planning (MSP) remains a fundamental and underexplored bottleneck in AI-driven materials discovery, as it requires not only identifying suitable precursor materials but also designing coherent sequences of synthesis operations to realize a target material. Although several AI-based approaches have been proposed to address isolated subtasks of MSP, a methodology for solving the entire MSP task has yet to be established. We propose MSP-LLM, a structured LLM-based framework that formulates MSP as a two-stage process composed of two constituent subproblems: precursor prediction (PP) and synthesis operation prediction (SOP). Our approach introduces a discrete material class as an intermediate decision variable that organizes both tasks into a chemically consistent decision chain. For SOP, we further incorporate hierarchical precursor types as synthesis-relevant inductive biases and employ precursor constraint factorization that preserves precursor-related information in the autoregressive decoding state. Extensive experiments show that MSP-LLM consistently outperforms existing methods on both PP and SOP, as well as on the MSP task, demonstrating an effective and scalable framework with practical potential for autonomous MSP.