NeurIPS 2026 论文列表

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.

When auditing machine learning models, human experts sequentially inspect model predictions and explanations, gradually building a mental model of the model’s behavior to uncover errors. This manual process is critical for safe deployment in high-stakes domains but fundamentally limited in scalability. In this work, we propose a novel method to simulate this human auditing behavior for automating the auditing process. We formalize auditing as a sequential decision-making problem, modeled as a Markov decision process. Our state representation summarizes the auditor’s accumulated knowledge of model predictions and feature usage. Building on this formalization, we solve the MDP using reinforcement learning, yielding RLAuditor, which learns to select informative test samples to efficiently uncover model errors. We further connect our formalization to Theory of Mind, showing that our state mirrors how human auditors build mental models: its update rule is equivalent to Bayesian belief revision and it converges to the model’s behavioral pattern at an interpretable rate. We validate our approach on multiple ML models and datasets across different data modalities. In a human subject study, we demonstrate that RLAuditor helps human auditors produce more accurate audit reports in less time.

Recurrent neural networks are widely used in neuroscience and machine learning, but their weight matrices often appear dense and unstructured, defying clean interpretation. The Schur decomposition expresses any recurrent weight matrix in quasi-triangular form, allowing network dynamics to be interpreted as a latent circuit of "functionally feedforward" interactions between orthogonal modes. However, its application for interpreting recurrent weight matrices has been limited due to its well-known combinatorial non-uniqueness. We introduce SchurMO (Schur Manifold Optimization), a method for discovering structured Schur forms via Riemannian gradient descent on the manifold of orthogonal similarity transformations, enabling recovery of target motifs. Compared to a greedy baseline, SchurMO more reliably and more efficiently recovers ground-truth latent motifs (chain-like, banded, and modular structure) from noisy weight matrices. Applying SchurMO to nonlinear recurrent networks enables recovery of latent chain-like structure, providing a link between connectivity and the underlying dynamics of sequence production. Our results highlight SchurMO as a scalable, computationally efficient approach to discovering latent structure in the Schur forms of recurrent weights.

Naturalistic language encoding models based on large language model features can predict cortical responses well, but their learned representations remain difficult to interpret. Three challenges are central: semantic selectivity is fundamentally non-identifiable in high-dimensional correlated feature spaces, cortical organization remains implicit in voxelwise weight maps, and cross-subject structure is usually addressed only after model fitting. We introduce a multisubject coupled sparse autoencoder that jointly reconstructs language-model features and predicts subject-specific brain responses, learning a sparse basis of semantic-cortical atoms that makes semantic structure, cortical organization, and cross-subject correspondence explicit within a single model. In naturalistic story-listening fMRI, the model retains 88\% of matched dense-feature ridge brain prediction performance using 50 atoms and $k{=}5$ active atoms per TR. Relative to matched two-stage baselines, joint training yields the strongest overall tradeoff between prediction and interpretability. The learned atoms are semantically coherent and interpretable, recover more reproducible cross-subject brain maps than standard post hoc encoding-model analyses, align with large-scale cortical structure from Yeo~7 and Neurosynth, and transfer better than ridge to held-out subjects in the low-data regime. This framework provides a more interpretable basis for cognitive neuroscience analyses of semantic-to-brain encoding by linking semantic features to corresponding cortical response patterns across subjects.

Learning in neural systems arises from synaptic changes that reshape the representations underlying behavior. While low-rank recurrent neural networks (RNNs) have emerged as a powerful framework for linking connectivity to function, a theoretical understanding of their learning process remains elusive. Here, we extend the low-rank framework from activity to learning by deriving gradient-descent dynamics directly in a reduced overlap space. We formulate a closed-form, low-dimensional system of ODEs that governs learning in this space, exact for linear RNNs and asymptotically exact for nonlinear RNNs in the large-$N$ Gaussian limit. Central to our analysis is a distinction between two classes of overlaps: _loss-visible_ overlaps, which fully determine network activity, output, and loss, and _loss-invisible_ overlaps, which do not affect function but are required to describe learning. We illustrate the consequences of this decomposition through two phenomena. First, we show that learning can serve as a perturbation that exposes differences in connectivity between functionally equivalent networks. Second, we show that loss-invisible overlaps can act as memory variables that encode training history, and characterize the conditions under which this occurs. Finally, we present several testable predictions for biological learning experiments derived from our theory.

Arena-style rankings of language models are a widely used evaluation framework in leaderboards like LMArena, research papers, and model evaluation in production. These rankings aim to realistically assess model performance using pairwise comparisons of anonymous models on user prompts, aggregating votes via the Bradley-Terry model to produce the final ranking. However, we find that these leaderboards are highly vulnerable to several manipulation strategies that unscrupulous providers could use to boost a model's rank: strategic voting (individual votes that boost a model's rank, including votes on irrelevant models), strategic prompting (choosing prompts that favor a model), and strategic nomination (boosting the rank of a model by inserting irrelevant models). We demonstrate these effects on public leaderboard data, analyze the circumstances that cause these vulnerabilities, and describe approaches to safeguard against each attack. Notably, these issues hold for any pairwise preference-based evaluation that uses the Bradley-Terry model. These manipulations boost the rank of nearly every tested model; each attack can boost a model by up to 15-21 places in a ranking and, combined, they can boost a model by up to 45 places.

Unified models (UMs) hold promise for their ability to understand and generate content across heterogeneous modalities. Compared to merely generating visual content, the use of UMs for interleaved cross-modal reasoning is more promising and valuable, e.g., for solving understanding problems that require dense visual thinking, improving visual generation through self-reflection, or modeling visual dynamics of the physical world guided by stepwise action interventions. However, existing UMs necessitate pixel decoding as a bridge due to their disjoint visual representations for understanding and generation, which is both ineffective and computationally cumbersome. In this paper, we introduce LatentUM, a novel unified model that represents all modalities within a shared semantic latent space, eliminating pixel-space mediation for model-internal reasoning over generated visual states. By making generated visual tokens directly interpretable to the model itself, LatentUM enables flexible interleaved cross-modal reasoning and generation. Empirically, LatentUM achieves state-of-the-art performance on the Visual Spatial Planning benchmark, pushes the limits of visual generation through self-reflection and demonstrates world modeling capability by predicting future visual states in the shared semantic latent space.

Goal-conditioned visual navigation has been a long-standing testbed for embodied AI. We study a natural language-conditioned variant, language-conditioned visual navigation (LCVN), in which an embodied agent must follow a natural language instruction given only an initial egocentric observation. Without access to goal images, the agent must rely on language to shape its perception and continuous control. We introduce the LCVN Dataset, a benchmark of 39,016 trajectories and 117,048 human-verified instructions spanning diverse environments and instruction styles. Building on this benchmark, we study two complementary paradigms: (i) latent-imagination policy learning, in which a diffusion-based world model (LCVN-WM) imagines future observations and an actor-critic agent (LCVN-AC) learns its policy entirely within the imagined latent space; and (ii) unified autoregressive prediction, in which a single multimodal backbone (LCVN-Uni) jointly predicts actions and observations in one forward pass over a shared token sequence. Experiments show that two paradigms offer complementary strengths: latent imagination produces more temporally coherent rollouts, whereas unified prediction generalizes better to unseen environments. Targeted ablations further isolate the contributions of language guidance, conditioning signals, and instruction style, clarifying when language grounding versus dynamics modeling is the performance bottleneck. Together, these findings position LCVN as a concrete testbed for studying how language, imagination, and decision-making interact in embodied agents.

Tabular data is widely applied in critical domains, where tree-based models remain dominant, yet their performance is often constrained by manual heuristic splitting criteria and limited ensemble diversity. While hyperbolic geometry naturally fits the hierarchical topology of tabular data, existing hyperbolic algorithms are hindered by complex Riemannian optimization and the difficulty of defining subtree priors in non-Euclidean spaces. To address these challenges, we propose HOT, a hyperbolic random forest framework based on data decoupling. First, we leverage intrinsic geometric structures and LLM reasoning to guide subtree construction, replacing traditional manual heuristics with geometric-semantic priors. Second, we introduce an orthogonal hyperbolic subspace projection mechanism via tangent space isometry, effectively decoupling feature dependencies and maximizing ensemble diversity. Finally, we establish an end-to-end collaborative training paradigm where gradient residuals from a graph network directly guide the growth of new hyperbolic trees, bypassing expensive iterative optimization. Experiments on 13 benchmark datasets demonstrate that HOT significantly outperforms state-of-the-art baselines in both computational efficiency and predictive performance.

AI agents increasingly excel at generating, testing, and refining code. However, they fall short on tasks requiring formal guarantees of full coverage that testing alone cannot provide. Distributed systems are a prime example: properties such as consistency between reads and writes must hold under every possible interleaving of events. Mechanized formal verification can guarantee such correctness, but typically demands months to years of expert effort. As evidence, even SOTA coding agents (Claude Code with Opus 4.6, Codex with GPT-5.4) succeed on only 2/7 distributed key-value-store specifications. In this paper, we present the first effective approach to addressing this gap, Inductive Deductive Synthesis (IDS), which jointly and incrementally synthesizes implementation and proof, and learns from failed attempts to systematically try promising strategies. Built as an agentic LLM system, IDS achieves 7/7 in about 6.8 hours and $106 per spec on average, roughly 200× faster than expert effort and 17% cheaper than SOTA agents. IDS further incorporates performance feedback into the same loop, yielding implementations up to 3× faster than published verified systems.

Building capable medical vision-language models (VLMs) is challenging due to the scarcity of domain data and the difficulty of eliciting pre-trained medical visual knowledge. We present HuatuoGPT-5, a medical VLM built on a simple yet effective recipe: large-scale multimodal pre-training followed by reinforcement learning (RL). We first construct PubMedVision-Plus, the largest medical image-text dataset with 60M pairs from public PubMed literature, along with a clean 12M high-quality subset. After pre-training on this data, we find that traditional supervised fine-tuning (SFT) struggles to elicit the pre-trained knowledge. In contrast, by applying RL with verifiable and rubric-based instances synthesized directly from the pre-training data, we successfully incentivize the model to use this pre-trained medical visual knowledge. On PubMed-derived data, HuatuoGPT-5 outperforms similarly sized VLMs on multiple medical benchmarks and expert evaluations, significantly narrowing the gap between medical VLMs and leading proprietary VLMs. We hope these data, models, and findings can help accelerate open research on capable medical VLMs.

Uncertainty quantification for individual treatment effects (ITEs) is a daunting challenge in causal inference. Motivated by recent advances in conformal prediction, several works aim to construct distribution-free prediction sets for ITEs with desired coverage under standard assumptions such as strong ignorability and overlap. In this paper, we show that such goals are fundamentally unattainable in the presence of continuous covariates. Specifically, we establish finite-sample and asymptotic impossibility results demonstrating that any distribution-free prediction set achieving desired coverage for ITEs must be trivial, in the sense that it has infinite expected length. Our analysis relies on a connection between ITE inference and the hardness of conditional independence testing, and highlights the intrinsic limitations imposed by the missing data nature of causal inference. These results provide a new perspective on existing methods, clarifying that their apparent success necessarily relies on additional structural assumptions beyond standard causal assumptions.

Concept-based explanation methods have emerged as a prominent framework for interpreting deep neural networks using human-understandable concepts. However, existing approaches are limited to assigning a scalar importance score to a concept, failing to capture the interaction between the concept and its visual context. We argue that overlooking this interaction leaves a predictor's reasoning only partially understood. To address this gap, we introduce IDEA, a novel global post-hoc explanation framework that explicitly models the interaction between a concept and its visual context. Through this modeling, IDEA decomposes concept-level predictive information into three atomic components: uniqueness (independent concept contribution), redundancy (information shared with the context), and synergy (information arising strictly from concept-context interaction). Consequently, IDEA extends interpretability beyond merely identifying influential concepts to diagnosing how a predictor reasons -- for instance, flagging potential context dependencies that scalar scores cannot reveal. IDEA requires no task-specific concept annotations and no access to classifier internals. Evaluations on controlled and real-world settings demonstrate that IDEA can surface reasoning patterns invisible to traditional scalar concept scores.

"Hearing is never like being.'' —Persian Proverb Large language models have become the default advisors for life-critical human problems. While they democratize access to personal counseling, they suffer from a silent foundation bias: the internet is a record of people with the freedom to act. Current benchmarks assume users have this same agency, ignoring the reality of those in war zones or navigating statelessness. These long-tail experiences are not only missing from training data, but authentic evaluation data to measure them is equally scarce. We introduce HearSayBench, a human-verified dataset of 400 scenarios from respected archives like the United Nations, covering 80 regions across three specific barriers: social, personal, and environmental. Our work uses Capabilities Approach to test if a model can distinguish between what a person is legally promised and what they are actually free to do in their specific environment. While these models may have "heard'' about global inequality during training, we find that this knowledge is merely hearsay. Across 11 frontier and open-weight models, we identify a systemic %37 performance drop between situational comprehension and structural reasoning. When faced with the most vulnerable users, models consistently offer a "Checklist of Impossible Things'': polite, fluent advice that is physically impossible or legally suicidal to follow. Ultimately, we show that the true digital divide is no longer about access to technology, but about whether an AI can recognize the reality of your life.

Embodied agents make decisions by combining information from heterogeneous memory modules, including declarative memory for explicit knowledge, procedural memory for learned routines, and working memory for current observations. During interaction, the memory contents required for decision making change with the agent's state, goal, and task progress, while many contents remain shared across consecutive decisions. However, existing memory modules are typically accessed through separate interfaces, causing each updated memory context to be processed as a new input to the foundation model. This prevents selective key-value (KV) cache reuse for unchanged memory contents and forces full KV cache recomputation even when much of the memory context remains valid. We introduce HeMeR, an inference framework that performs KV reconciliation between dynamically updated memory contexts and reusable KV states. HeMeR tracks declarative, procedural, and working memory contents across action decisions, preserves KV states for unchanged contents, and reconciles only the states affected by memory-content updates. This enables efficient and stable inference as memory contents evolve during interaction. We evaluate HeMeR on ALFWorld, Habitat, and real-world robotic deployments requiring coordinated use of heterogeneous memory contents. On Real-world, HeMeR achieves $83.20\%$ task success and reduces per-step inference time from $2.94$ to $1.87$ seconds compared with MemRL by preserving compatible KV states as task progress updates the memory context.

We introduce **Graph Cascades**, a mesoscopic rewiring strategy for Graph Neural Networks (GNNs) and Graph Transformers (GTs) that captures intermediate-scale graph structure beyond purely local edges or fully global attention. Using contagion-based diffusion processes, Graph Cascades constructs, in $\mathcal{O}(|V|+|E|)$ time, an auxiliary graph where node pairs supported by repeated multi-hop reinforcement are promoted to direct neighbors. We theoretically characterize when reinforcement-based rewiring helps: sufficient conditions under which a reinforcement-based edge-selection rule has higher label agreement than direct adjacency, an SBM witness in which two-hop reinforcement is perfectly homophilic, and formalizing mesoscopic connectivity via graph effective resistance. Empirically, across node-classification benchmarks, Graph Cascades improves multiple GNN and sparse-GT backbones, with the most reliable gains observed on heterophilic and moderate- to high-degree homophilic graphs. The theoretical conditions also identify regimes where mesoscopic rewiring is unlikely to be beneficial -- low-degree regular graphs and graphs with structural bottlenecks -- and these predictions match the observed failures. We additionally observe tight correlations between performance and structural properties in the rewired graphs.

Similarity measures such as canonical correlation analysis (CCA), representational similarity analysis (RSA) and center kernel alignment (CKA), are widely used to compare the representational geometries used by different neural networks to solve the same task. Yet, because existing methods compare the *extrinsic geometry* of neural representations in state space, rather than their *intrinsic geometry*, they may fail to capture subtle yet crucial distinctions between fundamentally different computations performed by biological and artificial neural networks. Here, we introduce metric similarity analysis (MSA), a novel method which leverages tools from Riemannian geometry to compare the intrinsic geometry of neural manifolds. Within our mathematical framework, we derive several properties of MSA, including coordinate, scale and rotation invariances. We show that MSA can be used to i) disentangle neural computations of deep networks in rich vs. lazy learning regimes, ii) compare the computation-through-dynamics of recurrent neural networks and state-space models during neuroscience working memory tasks, and iii) investigate the statistical manifolds of diffusion models. Overall, we introduce a mathematically grounded and broadly applicable framework to understand the mechanisms behind neural computations by contrasting their Riemannian geometries, with broad applicability to both neuroscience and machine learning.

Large Multimodal Models (LMMs) often struggle with geometric reasoning due to visual hallucinations and a lack of mathematically precise Chain-of-Thought (CoT) data. To address this, we propose the GeoSym Engine, an automated and scalable neuro-symbolic framework. By leveraging a type-conditional grammar and an analytic SymGT Solver, it derives exact symbolic ground truths and seamlessly integrates with a robust rendering pipeline to produce high-precision geometric diagrams. Using this engine, we construct GeoSym127K, a difficulty-stratified dataset featuring 51K high-resolution images, 127K questions with symbolic ground truths, and 55K answer-verified CoT QA pairs. We also introduce GeoSym-Bench, an expert-curated suite of 511 complex samples for rigorous evaluation. Through extensive supervised fine-tuning (SFT), we demonstrate that GeoSym drives concentrated improvements specifically on diagram-dependent and multi-step geometry tasks. Our Qwen3-VL-8B model gains an absolute +22.21% on the MathVerse Vision-Only subset and reaches 61.52% (+6.19% improvement) on WeMath, mitigating long-horizon logic fragmentation and outperforming advanced closed-source models like Doubao-1.8. Furthermore, applying Reinforcement Learning with Verifiable Rewards (RLVR) via GRPO reveals that initializing from structural SFT checkpoints substantially elevates the performance ceiling over zero-shot RL. Driven by deterministic exact-match signals, this showcases the robust scaling potential of our verifiable reasoning synthesis. Datasets and code are available at https://huggingface.co/datasets/Tomie0506/GeoSym127K and https://github.com/Tomie56/GeoSym127K

Generative modeling can be formulated as learning a mapping $f$ such that its pushforward distribution matches the data distribution. The pushforward behavior can be carried out iteratively at inference time, for example, in diffusion/flow-based models. In this paper, we propose *Drifting Models*, a generative modeling framework that evolves the pushforward distribution during training and naturally admits one-step inference. We introduce a drifting field that governs the sample movement and achieves equilibrium when the distributions match. This leads to a training objective that allows the neural network optimizer to evolve the distribution. In experiments, our one-step generator achieves state-of-the-art results on ImageNet 256$\times$256, with FID 1.54 in latent space and 1.61 in pixel space. We hope that our work opens up new opportunities for high-quality one-step generation.