NeurIPS 2026 Papers

As the era of large language models (LLMs) unfolds, Preference Optimization (PO) methods have become a central approach to aligning LLMs with human preferences and improving performance. We propose Maximum a Posteriori Preference Optimization (MaPPO), a methodology for learning from preferences that explicitly incorporates prior reward knowledge into the optimization objective. Building on the paradigm employed by Direct Preference Optimization (DPO) and its variants of treating preference learning as a Maximum Likelihood Estimation (MLE) problem, MaPPO integrates prior reward estimates into a principled Maximum a Posteriori (MaP) objective. This not only generalizes DPO and its variants, but also enhances alignment by mitigating the oversimplified binary classification of responses. Additionally, MaPPO introduces no additional hyperparameters, and supports preference optimization in both offline and online settings. In addition, MaPPO can be used as a plugin for DPO variants, including widely used SimPO, IPO and CPO, and produce consistent improvements. Extensive empirical evaluations of different model sizes and model series on three standard benchmarks (MT-Bench, AlpacaEval 2.0, and Arena-Hard) demonstrate consistent improvements in alignment performance without sacrificing computational efficiency.

Driven by increasingly complex real-world applications, Retrieval-Augmented Generation (RAG) has evolved from textual pipelines to multimodal settings, comprising two complementary dimensions: knowledge organization and retrieval enhancement. However, existing methods still face challenges in high construction cost, context window limitations, and the lack of unified training across offline and online knowledge sources. To address these challenges, we propose MMGraph-Agent, a unified multimodal RAG framework based on cache-inspired multimodal knowledge hypergraphs. MMGraph-Agent enables extremely lightweight hypergraph construction with zero API overhead and dynamic memory-based retrieval that alleviates long-context issues. Experiments across offline, online, and joint retrieval settings demonstrate consistent improvements in performance, efficiency, and architectural unification. Our software and data are publicly available.

Multimodal dynamic graphs are pervasive in real-world applications, where nodes and edges are accompanied by multimodal signals, including text and images. Despite their practical importance, benchmarks and methods that explicitly study multimodal learning in dynamic graphs remain scarce, hindering the full understanding and effective utilization of multimodal information for downstream tasks. To address this gap, we introduce MM-DyGraph, to the best of our knowledge, the first benchmark for multimodal dynamic graph learning, consisting of diverse real-world datasets with fine-grained temporal graph evolution and rich multimodal attributes. MM-DyGraph provides three evaluation tasks to comprehensively assess model performance in dynamic settings. We benchmark five state-of-the-art dynamic graph methods and reveal that naive feature concatenation yields target-level inconsistencies: multimodal inputs benefit certain prediction targets while harming others. This observation indicates that the modality informativeness depends on the specific prediction target over time. Building on these observations, we propose a Target-aware Spatial-Temporal Multimodal Fusion (TST-MF) framework, which jointly conditions multimodal fusion on the prediction target and spatial-temporal relations, enabling context-adaptive cross-modal interplay for target-specific representations. Extensive experiments over diverse datasets and tasks show that TST-MF delivers robust and competitive performance, achieving clear gains in multimodal settings.

We study the $K$-armed logistic bandit problem, where at each round, the agent observes $K$ feature vectors associated with $K$ actions. Existing approaches that achieve a rate-optimal $\widetilde{\cal O}(\sqrt{dT})$ regret bound rely heavily on context diversity assumptions, such as strict positivity of the minimum eigenvalue of a context covariance matrix. These assumptions, however, impose strong restrictions on the context process, as they rule out the situation where the context vectors are concentrated in a low-dimensional subspace. In this paper, we propose SupSplitLog, which, to the best of our knowledge, is the first algorithm for logistic bandits that achieves $\widetilde{\cal O}(\sqrt{dT})$ regret without any context diversity assumption. The key idea is to split the collected samples into two disjoint subsets when constructing estimators; one is used to compute an initial-point estimator, while the other is used to apply a Newton-type one-step correction procedure. The splitting rule is carefully designed to balance the accuracy requirements of the initial-point estimator and the one-step correction procedure. Moreover, SupSplitLog strictly improves on the existing algorithms in terms of the dependence on dimension $d$ in the regret upper bound. Furthermore, SupSplitLog can be adapted simply to deduce a regret bound that grows with a data-dependent complexity measure, avoiding a direct dependence on $d$, which is favorable when the context vectors are concentrated in a low-dimensional subspace. We also provide experimental results that demonstrate numerically the superiority of our algorithm, validating the theoretical results.

Sparse Mixture of Experts (MoE) models offer a scalable and efficient architecture for training large neural networks by activating only a subset of parameters (“experts”) for each input. A learned router computes a distribution over these experts, and assigns input tokens to a small subset. However, without auxiliary balancing mechanisms, routers often converge to using only a few experts, severely limiting model capacity and degrading performance. Most current load balancing mechanisms encourage a uniform routing probability across experts. Early in pretraining, this can result in inconsistent routing behavior, resulting in the model spending its capacity learning redundant knowledge. We address this by introducing a novel load balancing loss that helps preserve relational structure, encouraging consistent expert choices for similar inputs during training. Our experimental results show that applying our loss to the router results in 36% faster convergence and lower redundancy compared to a popular load balancing loss.

Modern language models are trained on heterogeneous web-scale text corpora. Consequently, studying knowledge and skill acquisition is difficult, as prior exposure to related content is hard to characterize. To address this challenge, we introduce **LittleCurriculum**, a meticulously curated 88B-token pretraining corpus tailored to U.S. elementary school material, explicitly excluding concepts, facts, and vocabulary taught above Grade 5. Training a 2B-parameter LLM from scratch on **LittleCurriculum** yields **LittleLearner**, a model with sufficient language competence for open-ended evaluation, yet with clear knowledge and capability boundaries mapped to interpretable curriculum guidelines. We release **LittleCurriculum** and **LittleLearner** as a developmentally restricted sandbox to study how models acquire, represent, and use data under a well-defined training scope. We illustrate the sandbox's utility in a first suite of experiments on injecting new knowledge through post-training and in-context learning. These methods let **LittleLearner** better utilize existing knowledge, but do not raise out-of-scope capabilities. Our findings underscore the value of this controlled environment for future investigations.

Large software projects often depend on older versions of libraries, even as APIs continue to evolve across releases. This creates a challenge for LLMs: they must maintain knowledge of multiple API versions, not merely the latest or most common one. However, current LLMs, trained on temporally mixed corpora and lack explicit mechanisms for such version-specific reasoning, leading to anachronistic errors -- calling APIs as they exist in a different library version. To systematically evaluate this phenomenon, we introduce LibEvoBench, a multi-task benchmark spanning multiple versions of widely used Python libraries, along with a new metric, the Software Evolution Understanding Score (SEUS), to measure models' consistency when working with evolving APIs. Our results show that state-of-the-art models are largely version-oblivious: performance degrades for evolving APIs, while for stable APIs it remains the same across versions. Moreover, simply specifying the target version provides no benefit, while relevant documentation significantly boosts models' accuracy. These findings highlight a systematic limitation of current training paradigms and motivate new approaches for temporally grounded knowledge in code generation.

Understanding human behavior while interacting with the surrounding world is crucial for many applications of embodied AI. First-person videos are particularly informative for this problem, as they well capture how activities reshape the scene over time. However, existing approaches often rely on implicit visual or language-aligned representations, disregarding structured reasoning over the scene dynamic. In this paper we argue that explicit, compositional and editable representations of human-environment interactions can play a crucial role for rich grounded activity understanding. To this end, we introduce SG-Ego, a large scale annotation set extending Ego4D with spatio-temporal scene graphs, where relations triplets are consolidated over time into explicit time-evolving descriptions of the scene state. To effectively reason over this representation, we propose GLEN, a graph-based model that operates over scene graph sequences to both align them with textual actions and model their temporal evolution. In addition, we formulate the activity-driven graph-edit forecasting (A-GEF) problem, a novel task that casts scene dynamics as a sequence of structured transformations conditioned on ongoing actions, enabling explicit reasoning about how scenes change over time. We demonstrate the effectiveness of our approach across multiple downstream tasks, spanning retrieval benchmarks as EgoMCQ and the more complex compositional EgoCVR, as well as long-horizon reasoning benchmarks as EXPLORE-Bench and the newly introduced A-GEF. GLEN achieves competitive or state-of-the-art performance compared to raw video baselines and it excels in reasoning settings, typically addressed only with MLLMs, while enabling controllable and structured predictions of scene dynamics driven by human activities. We believe our results establish spatio-temporal scene graphs, together with models that reason over them, as strong compositional and interpretable representations for egocentric video understanding and potentially beyond. Code will be released upon acceptance.

Graph Neural Networks (GNNs) operate under the implicit assumption that the underlying data manifold is Riemannian, where the metric tensor is strictly positive-definite. While effective for homophilic graphs, this inductive bias creates a fundamental geometric mismatch for heterophilic graphs, where edges often signify dissimilarity or structural repulsion. In this work, we propose a paradigm shift from learning on fixed manifolds to learning the manifold itself. We postulate that heterophilic graphs are naturally embedded in pseudo Riemannian manifolds endowed with an indefinite metric, allowing for negative squared distances to model repulsive interactions. To formalize this, we introduce the Graph Signature Index, a spectral invariant that diagnoses the geometric nature of a graph. This index enables our proposed Pseudo-Riemannian Attention Network (PRAT) to dynamically learn an indefinite metric tensor, effectively capturing both attractive and repulsive interactions. By unifying these opposing forces within a single framework, PRAT demonstrates how a subtle shift in the metric signature yields substantial performance gains on heterophilic benchmarks. Our analysis reveals that PRAT’s learned geometry emergently aligns with the spectral signature of the graph, validating our geometric hypothesis.

GraphRAG enhances Large Language Model reasoning by organizing retrieved knowledge into structured evidence graphs, enabling inference over connected evidence rather than isolated text fragments. Yet, existing GraphRAG methods often either miss query-critical links or introduce noisy or conflicting evidence that distracts reasoning. We propose CONSIST, a Conflict-aware Nash-optimized Minimal Sufficient graph learning algorithm that regulates evidence graphs to resolve conflicts during reasoning. Specifically, we cast graph learning as an exploration–editing process, where a candidate graph is expanded via structural and semantic connections and then refined by pruning unreliable edges. We further introduce Gumbel-based edge editing, framing it as a Nash-guided bargaining process, where current and future players negotiate over a customized utility to determine reliable edits that improve both immediate answer support and long-term graph quality. Experiments on multi-hop question answering and conflict-aware benchmarks demonstrate that CONSIST consistently outperforms recent state-of-the-art baselines, achieving average performance gains up to 7.0% EM and 6.3% F1. Moreover, CONSIST also produces substantially more compact evidence graphs, reducing retained nodes by 42.9% and edges by 54.7% on average.

Modern reasoning models depend on reasoning data, today sourced from human annotations or distilled from stronger LLMs. However, a rich and largely untapped source of supervision lies in expert systems (e.g., game engines, classical planners, theorem provers), which routinely produce near-optimal actions across diverse domains. But these experts are silent: they commit to an action without writing down the chain of thought (CoT) behind it. Recovering that CoT as natural-language reasoning would distill expert knowledge into a student that generalizes beyond the demonstrated actions. We treat it as a latent variable and study how to recover it from the action alone. Our approach, LeAct (**Le**arning to reason from **Act**ions), optimizes this latent variable: the student samples candidate CoTs for each expert action, and we retain those that measurably improve its own probability of recovering the action. Across imperfect-information games at multiple scales and a simulated robotics benchmark, LeAct reaches the solver's numerical floor on small enumerable games. At larger scale, it is $5\times$ closer to the solver than the strongest expert-iteration baseline. At Flop Hold'em ($\sim 10^9$ infosets), LeAct wins head-to-head by $+60$ mbb/g, and on the robotics probe it is the only training recipe that improves on direct imitation. We present a principled framework and the result: expert systems become a categorically new source of reasoning teachers for foundation models.

Current analyses of LLMs' parametric knowledge are largely output-centric, drawing conclusions about what a model knows without verifying what it was actually trained on. This leaves fundamental questions, such as whether a correct response reflects genuine generalization or rote memorization, grounded in speculation rather than evidence. To resolve these ambiguities, we introduce LUMOS, a diagnostic framework that traces knowledge along the causal chain from training-data exposure to behavioral output, leveraging OLMo 2 with its fully transparent training corpus. By grounding analysis in verified exposure, we reveal that models internally encode rare facts with high separability (84%) yet fail to express them behaviorally (54%), though this retrieval gap narrows with scale. Furthermore, when models are asked to self-reflect on their own answers, they perform reliably on trained content (83%) but drop to random-baseline levels (49%) on unseen content. This collapse persists even under chain-of-thought prompting, which inflates confidence signals rather than improving calibration. Collectively, these findings demonstrate that incorporating the training-data axis into LLM evaluation transforms speculative diagnoses into verifiable claims, and we advocate that this axis should be a standard component of knowledge assessment in LLMs.

Modern LLMs continue to exhibit significant variance in behavior across languages, such as being able to recall factual information in some languages but not others. While typically studied as a problem to be mitigated, in this work, we propose leveraging this cross-lingual inconsistency as a tool for interpretability in mixture-of-experts (MoE) LLMs. Our knowledge localization framework contrasts routing for sets of languages where the model correctly recalls information from languages where it fails. This allows us to isolate model components that play a functional role in answering about a piece of knowledge. Our method proceeds in two stages: (1) querying the model with difficult factual questions across a diverse set of languages to generate "success" and "failure" activation buckets and then (2) applying a statistical contrastive analysis to the MoE router logits to identify experts important for knowledge. To validate the necessity of this small number of experts for answering a knowledge question, we deactivate them and re-ask the question. We find that despite only deactivating about 20 out of 6000 experts, the model no longer answers correctly in over 40% of cases. Generally, this method provides a realistic and scalable localization approach to increasingly complex LLMs, but also suggests redundant and highly dispersed knowledge parameterization in MoEs.

Large language models (LLMs) have shown promise in scientific discovery tasks such as protein optimization, but remain limited by the lack of a mechanism for persistent and reusable knowledge accumulation. Existing approaches adapt LLMs through parameter optimization or trajectory refinement, but in both cases, knowledge remains implicit, either encoded in opaque model weights or scattered across transient solution histories. As a result, knowledge cannot be systematically inspected, attributed, or reused. We propose \emph{KnowEvo}, a framework that treats externalized knowledge itself as the optimization target. Our key insight is that scientific knowledge is inherently modular: multiple interacting mechanisms jointly determine the design outcomes. This makes knowledge evolution fundamentally challenging, as improvements cannot be attributed to individual components in isolation. To address this, \emph{KnowEvo} represents knowledge as executable expertise blocks and organizes their refinements into strategy mutation trees, enabling structured exploration over a combinatorial knowledge space. The framework combines a forward exploration phase that adaptively allocates evaluation budget across strategy variants with a backward revision phase in which an agentic sandbox converts rollout evidence into localized program improvements. This design enables persistent accumulation, attribution, and refinement of mechanistic expertise across iterative experimentation. We evaluate \emph{KnowEvo} on protein optimization tasks including thermostability and solubility. Results show that knowledge evolution consistently improves over initial expert heuristics and outperforms monolithic program evolution and prior LLM-guided evolution baselines across ranking metrics such as pass@$k$, MRR, MAP, and NDCG@5. Codes are available at \url{https://anonymous.4open.science/r/KnowEvo-6E3B/}.

Large language models (LLMs) are increasingly used for knowledge base question answering (KBQA), where answering requires selecting entities from a question-specific knowledge-graph subgraph. Yet LLMs are known to hallucinate across tasks, and KBQA is no exception: even when we provide a graph as the knowledge source, the model may rely on parametric knowledge instead of graph evidence or perform invalid reasoning over the given relations. Such hallucinated answer nodes can limit the practical deployment of KBQA systems, especially in high-stakes domains such as healthcare. We formulate hallucination detection in KBQA as an answer-node classification problem and propose a lightweight graph-based framework that treats the answering LLM as a black box. KG-Guard represents each KBQA instance as an augmented graph. It initializes node features with semantic representations of KG entities, marks topic entities and LLM-proposed answer nodes with learned vectors, and connect a virtual question node to the topic entities. A graph encoder then produces verification-oriented node representations, and a small MLP classifies each proposed answer node using its graph representation together with the question embedding. Experiments on WebQSP, ComplexWebQuestions, and PUGG show that our detector achieves the highest F1 on all three benchmarks ($82.0$, $87.4$, and $84.3$), outperforming LLM-judge and sampling-based baselines, while having $\sim305\times$ fewer parameters than the reference approaches. Beyond detection, the node-level feedback is actionable: when flagged answers are fed back to the KBQA system for iterative refinement, downstream KBQA F1 improves by $13.0$-$14.5$ points and Exact Match by $16.9$-$17.6$ points.

Liability adjudication tasks (e.g., legal charge determination), require analyzing event descriptions against domain-specific skills to determine whether a party should bear liability. While Large Language Models (LLMs) have shown promise in this area by leveraging executable skills for interpretable reasoning, existing approaches face three critical limitations: **skill libraries become stale as real-world scenarios dynamically evolve**, **skill descriptions often suffer from a semantic execution gap where LLMs deviate from intended expert logic**, and **growing skill collections exceed context budgets causing critical skill omission**. To address these challenges, we propose **K**nowledge-**D**ata Dual-Driven **A**utomated **S**kill **O**ptimization (KDASO), which continuously evolves a skill library through neighborhood-conditioned induction from recent adjudication data, execution-aligned refinement based on LLM trajectory analysis, and context-budgeted routing for scalable skill selection. Experiments across ride-hailing adjudication, legal liability determination, and logical reasoning judgment demonstrate that KDASO significantly outperforms current general-purpose and domain-specific baselines. KDASO further achieves an 8\% improvement when transferring skills to concept-similar scenarios and maintains consistent performance gains through continuous optimization, underscoring the high transferability and stability of generated skills.

Large language models (LLMs) are increasingly deployed in K–12 education, yet existing benchmarks such as C-Eval, CMMLU, GaokaoBench, and EduEval measure only whether a model can answer an exam question, i.e., factual recall. Effective educational AI further requires curriculum cognition: the structured understanding of how knowledge is organized, including prerequisite chains, concept taxonomies, experiment–concept links, and pedagogical sequencing. Curriculum cognition is neither probed by current benchmarks nor explicitly taught by current instruction-tuning data. To close this gap, we introduce K12-KGraph, a curriculum-aligned knowledge graph extracted from the official People's Education Press textbooks, covering mathematics, physics, chemistry, and biology across primary, middle, and high school, with seven node types (Concept, Skill, Experiment, Exercise, Section, Chapter, Book) and nine relation types spanning taxonomy, prerequisite, association, verification, assessment, location, and order. Building on this single graph, we derive two complementary resources: K12-Bench, a 23,640-question multi-select benchmark across five graph-derived task families (Ground, Prereq, Neighbor, Evidence, and Locate) that jointly probe curriculum cognition; and K12-Train, a KG-guided supervised fine-tuning corpus of ~2,300 QA pairs synthesized from node attributes and edge semantics. Experiments expose a clear gap and a clear remedy: on K12-Bench, even the strongest proprietary model (Gemini-3-Flash) reaches only 57% exact match and the best open-source model (Gemma-4-31B-IT) only 46%, with Prereq and Neighbor being the hardest; yet under a strictly matched 2,300-sample SFT budget on Qwen3-4B-Base and Llama3.1-8B-Base, K12-Train consistently outperforms equally sized subsets of eight mainstream instruction-tuning corpora (OpenHermes, Infinity, UltraChat, WizardLM, DataFlow, LMSYS, SmolTalk, Tulu-3) on both GaokaoBench and EduEval, showing that structural curriculum grounding is remarkably sample-efficient for educational SFT. We release the graph, benchmark, training data, and the full construction pipeline.

Graph neural networks for bot and fraud detection operate on attributed graphs, where adversaries can jointly modify local edges and continuous node features. We study node-level certification under a localized mixed threat model with up to $d$ edge flips in the target message-passing neighborhood and an $\ell_2$ Frobenius feature budget $\epsilon$. We instantiate hybrid randomized smoothing for this graph setting and derive two graph-specific certificates: a sound Sequential NP composition baseline for the joint topology--continuous-feature threat model, and a Hybrid NP certificate whose homogeneous edge-smoothing structure reduces the adversarial adjacency search to an $O(d^2)$ optimization over edge-deletion and edge-addition counts. Across PubMed, Amazon Fraud, YelpChi, and MGTAB, the Hybrid NP certificate yields non-trivial certified regions under joint perturbations. We also construct adaptive joint attacks and show failure cases on PubMed, Amazon Fraud, MGTAB and YelpChi for topology-only certified nodes, demonstrating that topology-only guarantees can overestimate robustness when continuous feature perturbations are allowed. The results support joint discrete--continuous certification as the appropriate robustness notion for attributed graphs with continuous features.

Existing research infrastructure is fundamentally document-centric, providing citation links between papers but lacking explicit representations of methodological evolution. In particular, it does not capture the structured relationships that explain how and why research methods emerge, adapt, and build upon one another. With the rise of AI-driven research agents as a new class of consumers of scientific knowledge, this limitation becomes increasingly consequential, as such agents cannot reliably reconstruct method evolution topologies from unstructured text. We introduce Intern-Atlas, a methodological evolution graph that automatically identifies method-level entities, infers lineage relationships among methodologies, and captures the bottlenecks that drive transitions between successive innovations. Built from $1{,}030{,}314$ papers spanning AI conferences, journals, and arXiv preprints, the resulting graph comprises $9{,}410{,}201$ semantically typed edges, forming a queryable causal network of methodological development. To operationalize this structure, we further propose a self-guided temporal tree search algorithm for constructing evolution chains that trace the progression of methods over time. We evaluate the quality of the resulting graph against expert-curated ground-truth evolution chains and observe strong alignment. In addition, we demonstrate that Intern-Atlas enables downstream applications in idea evaluation and automated idea generation. We position methodological evolution graphs as a foundational data layer for the emerging field of automated scientific discovery.