NeurIPS 2026 论文列表

The ability to learn a new task from a few examples has been crucial to the success of foundation models such as large language models (LLMs). However, current foundation models for structured data like tables and relational databases still require tens of thousands of labeled examples to perform well on new tasks. Here we show that when pretrained at scale with the right recipe, Relational Transformers (RTs) can make state-of-the-art (SoTA) predictions with only hundreds of labels. This capability arises from the following ingredients: First, we assemble THE JOIN, the largest pretraining corpus of relational data to date, comprising 6k forecasting tasks across 650 real-world databases from diverse domains. Second, we develop a pretraining recipe that combines (1) mixed context sizes for multi-scale learning, (2) multi-cell masking for dense supervision, and (3) a novel random- walk-based retriever to efficiently gather relevant context. Third, we identify context ensembling and context tuning as complementary axes for scaling test-time compute to further improve few-shot performance. Pretrained with our recipe on THE JOIN, RT achieves parity with LLM Agent + TabICLv2 and RDBLearn + TabICLv2 pipelines with 32–23×fewer labels respectively, and even surpasses the prior SoTA of full task-specific training, on average nMAE for RelBench regression tasks. Context ensembling and tuning improves this further by up to 3%, and 4% respectively. Our ablations highlight the importance of schema semantics, multi-cell masking, and random-walk retrieval. Overall, our work paves the way for developing relational foundation models with strong few-shot capabilities.

Transformer-based large language models are increasingly used for long-horizon tasks; however, their attention mechanism scales poorly with context length. To handle this, we study a sleep-like consolidation mechanism in which a model periodically converts recent context into persistent fast model weights before clearing its key-value cache. During the sleep phase, the model performs $N$ offline recurrent passes over the accumulated context and updates the fast weights in its state-space model (SSM) blocks through a learned local rule. This shifts extra computation to the sleep phase while preserving the latency of wake-time prediction. We test our method on controlled synthetic tasks, including cellular automata and multi-hop graph retrieval, as well as a more realistic math reasoning task, on which a regular transformer as well as SSM-attention hybrid models fail. We then show that increasing sleep duration $N$ for our models improves performance, with the largest gains on examples that require deeper reasoning.

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.

Guardrails are a critical safety layer for modern AI systems, but their operating regime is changing. As LLMs are deployed as customized assistants, safety policies are increasingly specified at inference time by users, organizations, or regulatory contexts. This makes safety enforcement fundamentally dynamic: the guardrail should adapt to changing safety policies without retraining. Yet this requirement creates a fundamental tension: faithfully judging complex policy contexts demands reasoning capability, while practical deployment requires low-latency responses. We introduce Latent Policy Guardrail (LPG), a guardrail framework that learns **semantic latent deliberation** over dynamic policies. LPG compresses the internal deliberation needed for intent interpretation and policy grounding into continuous states supervised by decision-relevant semantics. At inference time, it generates only a compact verdict anchored to the violated policy clauses, preserving auditability while avoiding the latency of explicit reasoning. Across policy guardrail benchmarks, LPG-4B reaches 84.5% average safety accuracy and 77.9% F1 by compressing deliberation into just 10 latent tokens, outperforming the strongest dynamic baseline while running roughly 11 times faster than Qwen3-4B-Thinking under the single-sample evaluation setup.

Confidence-weighted routing, selective abstention, and ensemble weighting all assume that a model's stated confidence is informative about its capability on the question being asked. They presume functional metacognition, the capacity to assess one's own capabilities, without exercising them, relative to other frontier language models. Aggregate calibration is well studied, with mixed results, but the underlying structure of elicited confidence is less well understood. We decompose binary confidence judgements from 20 frontier Large Language Models (LLMs) across six benchmarks using tetrachoric factor analysis paired with pairwise calibration, asking whether two models that differ in confidence also differ in performance. On factual recall and information retrieval benchmarks the cross-model confidence matrix is approximately rank-one and a single dominant factor captures most of the latent variance. Models retrieving facts share an item-level difficulty axis and differ mainly in their decision thresholds along it. Across all benchmarks the relationship between confidence and performance collapses once items that all models agree on are removed. Inter-model pairwise calibration is small even where statistically significant, and what remains shrinks to nothing once base-rate differences along the shared factor are controlled for. Mathematical reasoning is the apparent exception, but this turns out to be a confound where reasoning models answer questions about their confidence by trying to solve them in their chain of thought, bypassing the sub-symbolic self-knowledge we seek to measure. We find no evidence for significant verbalised individuated metacognition in any tested domain. A trivial logistic regression on surface features of the confidence-judgement reasoning text beats the median confidence model's repeatedly sampled judgements on half the benchmarks, evidence that extracting uncertainty from the reasoning trace recovers signal the binary judgement does not.

Scaling large recommendation systems requires advancing three major frontiers: processing longer user histories, expanding candidate sets, and increasing model capacity. While promising, transformers' computational cost scales quadratically with the user sequence length and linearly with the number of candidates. This trade-off makes it prohibitively expensive to expand candidate sets or increase sequence length at inference, despite the significant performance improvements. We introduce **LIME**, a novel architecture that resolves this trade-off. Through two key innovations, LIME fundamentally reduces computational complexity. First, low-rank ``link embeddings" enable pre-computation of attention weights by decoupling user and candidate interactions, making the inference cost nearly independent of candidate set size. Second, a linear attention mechanism, **LIME-XOR**, reduces the complexity with respect to user sequence length from quadratic ($O(N^2)$) to linear ($O(N)$). Experiments on public and industrial datasets show LIME achieves near-parity with state-of-the-art transformers but with a 10$\times$ inference speedup on large candidate sets or long sequence lengths. When tested on a major recommendation platform, LIME improved user engagement while maintaining minimal inference costs with respect to candidate set size and user history length, establishing a new paradigm for efficient and expressive recommendation systems.

Large reasoning models, such as OpenAI o1 and DeepSeek-R1, tend to become increasingly verbose as their reasoning capabilities improve. These inflated Chain-of-Thought (CoT) trajectories often exceed what the underlying problems require, wasting compute, latency, and context budgets. While introducing length-based efficiency rewards during reinforcement learning offers a natural remedy, existing methods struggle with two fundamental challenges: the optimal balance between correctness and efficiency is non-stationary throughout training, and intrinsic reasoning budgets vary drastically across problems. Relying on static reward weights and global length constraints inevitably forces a compromise between degraded accuracy and unrealized compression. To overcome these limitations, we propose LEAD (Length-Efficient Adaptive and Dynamic reasoning), a method that replaces static heuristics with online, self-adaptive mechanisms. LEAD dynamically calibrates the correctness-efficiency trade-off at each step using a Potential-Scaled Instability, directing optimization capacity to the most informative learning signal. Furthermore, it estimates an adaptive per-problem target length online based on the model’s own correct rollouts, applying a symmetric efficiency reward that penalizes both overthinking and over-compression. Evaluated on five mathematical reasoning benchmarks, LEAD achieves the highest accuracy and Accuracy-Efficiency Score among RL-trained efficient-reasoning methods while producing substantially shorter outputs than the base model.

This paper studies test-time aggregation, an approach that generates multiple reasoning traces and aggregates them into a final answer. Most existing methods rely on evaluation signals collected from candidate traces in isolation or answer frequencies, while ignoring comparative interactions among candidates. We propose *Joint Consistency* (JC), formulated as a constrained Ising-type energy minimization problem, where independent evaluation signals act as external fields and pairwise comparisons act as interactions. JC provides a unified framework for test-time aggregation that subsumes existing voting and weighted aggregation methods as special cases. Our construction of the interaction matrix leverages LLM-as-a-judge comparisons, and admits a theoretical interpretation under answer-level homogeneity assumptions. Moreover, we develop an efficient approximation strategy that makes interaction modeling practical for large-scale test-time aggregation. Experiments on math and code reasoning benchmarks show that JC consistently outperforms existing baselines across tasks, judge models, trace budgets, and trace-generation settings.

Universal Transformers (UTs) show strong performance on abstract reasoning tasks such as ARC-AGI, yet the specific sources of their performance gains remain underexplored. We conduct systematic ablations over UT variants and find that performance improvements on abstract reasoning are driven primarily by (i) the recurrent inductive bias induced by parameter sharing across depth and (ii) strong nonlinearity inside the Transformer block, rather than by elaborate architectural heuristics. Motivated by this finding, we propose the Universal Reasoning Model (URM). First, we propose ConvSwiGLU, which augments the feed-forward block with a channel-wise short convolution to effectively enhance nonlinearity of UT. Second, we adopt Truncated Backpropagation Through Loops (TBPTL) to eliminate noise from early loop when training with many loops, by freezing the gradients of early loops. Our approach substantially improves abstract reasoning performance, achieving 57.5% pass@1 on ARC-AGI 1 and 16.0% pass@1 on ARC-AGI 2. More importantly, we pretrain a 8-billion-parameter large language model based on the URM architecture from scratch, and the results show that URM still substantially outperforms the baseline across all reasoning tasks.

Emerging latent reasoning paradigms allow Large Language Models to “think” in latent embedding spaces, offering a more efficient alternative to explicit Chain-of-Thought. However, they generally face two critical limitations: First, relying exclusively on latent tokens amplifies uncertainty due to their inherently high entropy, thereby degrading solution correctness. Second, the overthinking problem remains prevalent, and existing approaches typically rely on rigid heuristics for stopping, often resulting in suboptimal termination. To address these challenges, we view efficient reasoning as a learnable control problem over both how to think and when to stop. We formulate this within a Reinforcement Learning framework, training lightweight per-step gates on a frozen LLM. This modulation enables the model to (i) adaptively interleave latent soft-token reasoning with explicit token generation, and (ii) predict cumulative exit probabilities for robust early termination. Experiments on diverse reasoning benchmarks across varying model scales and families demonstrate that our method surpasses both standard CoT and existing latent baselines, achieving accuracy gains of ∼3% while reducing token consumption by up to 40%.

Obtaining meaningfully diverse high-quality samples from Large Language Models (LLMs) for a fixed prompt remains an open challenge. Current methods often only operate at the token-level, paraphrasing the same response. This is problematic as it leads to poor exploration on reasoning problems and to unengaging, repetitive conversational agents. To address this, we propose Intent Factored Generation (IFG), factorising the sampling process into two stages. First, we sample a semantically dense intent that anchors the sample, e.g., a summary or keywords. Second, we sample the final response conditioning on both the original prompt and the intent from the first stage. This factorisation allows the use a higher temperature during the intent step to promote conceptual diversity, and a lower temperature during the final generation to ensure the outputs are coherent. We find that prompting the model to explicitly state its intent for each step of the chain-of-thought before generating the step is beneficial for reasoning tasks. We show that our simple method is highly effective across a diverse set of tasks. We find that this method improves both exploration and final performance on math and code tasks, and combines well with Reinforcement Learning from Verifier Feedback (RLVF). We also show that our method can be used to increase the diversity of instruction-tuned models. Finally, we demonstrate that our method leads to higher quality diverse samples on a language modelling task, on a new dataset that we open-source. IFG is easy to implement, and can be used with off-the-shelf LLMs to improve the performance-diversity tradeoff. Full Implementation at https://anonymous.4open.science/r/IFG-Anon/

Supervised fine-tuning (SFT) provides the standard approach for teaching LLMs new behaviors from offline expert demonstrations. However, standard SFT uniformly fits all samples—including those with low likelihood under the base model—which can disproportionately drive training updates toward overfitting specific samples rather than learning the target behavior. Moreover, adapting to these unlikely samples induces substantial policy shifts that degrade prior capabilities. Existing methods mitigate this by filtering, regenerating, or down-weighting low-likelihood data. In doing so, they often suppress precisely the novel behaviors the base model has yet to learn. We propose InfoSFT, a principled weighting scheme for the SFT objective that concentrates learning signals on maximally informative, medium-confidence tokens—those neither overly familiar to the base model nor too unlikely to cause instability. Requiring only a one-line modification to the standard token-wise loss, InfoSFT demonstrably improves generalization over vanilla SFT and likelihood-weighted baselines across math, code, and chain-of-thought tasks with diverse model families, while better preserving pre-existing capabilities.

Building general-purpose role-playing agents that faithfully portray any character from a natural-language profile remains challenging. The dominant paradigm---supervised fine-tuning---encourages behavioral mimicry without deep, human-like internal thought processes, resulting in poor out-of-distribution generalization. Therefore, we propose Psy-CoT, a psychology-grounded chain-of-thought framework that decomposes pre-response reasoning into three role-specific steps---Interaction Perception, Psychological Empathy, and Logical Construction---so that the model thinks dynamically from the profile rather than merely mimicking surface patterns. While structured reasoning provides a foundation, it alone is insufficient; reinforcement learning is essential to further align the model with character fidelity. However, we observe that under LLM-based reward models, both generic phrases that hack the reward model and genuinely role-specific phrases receive identical gradient signals---this hacking accumulates over training, misleading the model into treating both as equally optimal choices. To address this, we propose Role-Aware Policy Optimization (RAPO), which uses profile--token mutual information to weight gradients asymmetrically---amplifying role-specific tokens under positive advantage while attenuating them under negative advantage. Experiments on CoSER, CharacterBench, and CharacterEval demonstrate that Psy-CoT outperforms existing role-playing CoT methods, and RAPO consistently surpasses GRPO across multiple model scales.

A core obstacle to alignment evaluation is evaluation awareness: capable models can tell when they are being tested rather than deployed, weakening the conclusions a safety audit can support. We present two techniques that make audits in Petri, where an auditor model red-teams a target over multiple turns to elicit concerning behavior, harder to distinguish from real deployments. Our first technique, \emph{critique refinement}, spends additional inference-time compute on each auditor action: an instance of the target model scores candidate actions for realism, provides feedback, and the auditor iterates before the highest-scoring candidate is selected. On Sonnet-4.6, realism win rate (the fraction of pairings in which the audit transcript is judged more realistic than a real deployment transcript) rises monotonically from 12\% to 39\% as refinement depth increases, and verbalized evaluation awareness drops to near zero; Haiku-4.5 and Opus-4.7 show a similar pattern. Our second technique, DISH (Deployment-Imitating SWE-Agent Harness), wraps the target in a Claude Code agent harness, reducing the gap between auditor-simulated and real deployment environments; in coding settings DISH raises realism win rate from 7\% to 18\% on Sonnet-4.6. The gains are additive on Sonnet-4.6 (12pp over either alone) but not on Opus-4.7.

Self-evolution offers a scalable path to stronger reasoning: a pretrained language model improves itself with only minimal external supervision. Yet existing methods either depend on extensively curated or teacher-generated training data, or, when the generator runs unsupervised, reward it by a difficulty heuristic that need not improve the solver. We introduce **INFUSER**, an iterative co-training framework with two co-evolving roles: a *Generator* that drafts questions and reference golden answers from a pool of unstructured, automatically collected documents, and a *Solver* that improves by training on them. The solver is trained with standard correctness rewards against the generator-provided answers, while the generator is rewarded by an *optimizer-aware influence score* that measures whether each proposed question would actually improve the solver on the target distribution. Because this continuous, noisy influence score is poorly served by standard GRPO, we propose **DuGRPO**, a dual-normalized variant of GRPO, for generator training. Together, these turn the document pool into an *adaptive curriculum* that favors questions useful to the current solver, not just hard ones. On Qwen3-8B-Base, INFUSER outperforms strong self-evolution baselines with over 20% relative improvement on Olympiad and SuperGPQA benchmarks, and an 8B INFUSER co-evolving generator outperforms a frozen 32B thinking generator on math and coding. Ablations confirm each design choice is necessary. Code is available at

Compute-optimal scaling laws have become the central organizing tool of language-model training, yet they have been measured exclusively under Euclidean readouts. We give the first compute-optimal scaling-law characterization of hyperbolic language modeling. Trained from scratch on OpenWebText across a paired 5×4 grid spanning 86M to 1.32B parameters and 66M to 8.4B tokens, our hyperbolic LM HypTip follows a markedly different scaling law than its Euclidean counterpart. Fitting the Hoffmann form L = E + A·N^(−α) + B·D^(−β) to each track yields a compute-optimal data-to-parameter ratio of D*/N* = 8.4 for the hyperbolic readout versus 20.8 for the Euclidean baseline. The hyperbolic readout therefore reallocates compute by 2.5× toward parameter-heavy training, winning 17 of 20 paired cells across the grid. This shift is mechanistically grounded: across all 20 trained checkpoints, the per-row Lorentzian radius of the LM head correlates with token log-frequency at Spearman ρ down to −0.83, and the strength of this Zipfian token hierarchy scales with training-token count rather than parameter count, indicating that the geometry-as-prior emerges causally during training rather than being read off a frozen Euclidean checkpoint as in prior post-hoc analyses. Held-out evaluation on WikiText-103 perplexity, LAMBADA, and PIQA preserves the same advantage, and a cross-architecture replication on GPT-2 and DeepSeek-V3 confirms that LM-head geometry, not body details, drives the effect. Output-layer geometry alone therefore reshapes the compute-optimal scaling law of language modeling and brings hyperbolic LMs into the same predictive framework as the Euclidean models that have shaped the field.

Vision-language models (VLMs) show strong multimodal capabilities but still struggle with fine-grained vision-language reasoning. We find that long chain-of-thought (CoT) reasoning exposes diverse failure modes, including perception, reasoning, knowledge, and hallucination errors, which can compound across intermediate steps. However, most existing data for reinforcement learning with verifiable rewards (RLVR) does not involve complex reasoning chains grounded in visual evidence throughout, leaving these failure modes understressed during training. We therefore propose HopChain, a scalable framework that constructs an out-of-distribution (OOD) proxy task: OOD only in query format, while strengthening foundational abilities shared with downstream benchmarks to support generalizable gains. Concretely, HopChain synthesizes multi-hop vision-language reasoning chains, each a logically dependent chain of instance-grounded hops where earlier hops establish the instances, sets, or conditions needed for later hops, ending in an unambiguous numerical answer for verifiable rewards. In experiments, we train Qwen3.5-35B-A3B and Qwen3.5-397B-A17B under two RLVR settings (with and without HopChain's multi-hop data) and compare them across 24 benchmarks spanning STEM and Puzzle, General VQA, Text Recognition and Document Understanding, and Video Understanding. Although the multi-hop data is not designed for any specific benchmark, it improves 20 of 24 benchmarks on both models, indicating broad and generalizable gains. Consistently, replacing full chained queries with half-multi-hop or single-hop variants reduces the average score across five representative benchmarks from 70.4 to 66.7 and 64.3, respectively. Improvements are particularly substantial in long-CoT and ultra-long-CoT regimes, peaking at more than 50 accuracy points in the ultra-long-CoT regime. These results show that OOD proxy tasks for long-chain visual reasoning are a scalable RLVR supervision source for generalizable VLM reasoning.

Deployed approaches for AI text detection often rely on training-time access to labeled datasets of both human-written and AI-generated text. This approach is vulnerable to three types of distribution shifts that occur continually post-deployment, and for which labeled data is often unavailable: adversarial humanization, new LLMs being released, and temporal drift in human writing. Simultaneously, existing approaches do not leverage a key signal of LLM usage: inference-time homogeneity. We propose a test-time adaptation (TTA) approach, using semi-supervised learning, that adapts to distribution shifts by leveraging homogeneity among unlabeled samples observed at inference time. Empirically, we find that state-of-the-art supervised detectors systematically fail when they encounter distribution shifts in AI-generated and human writing, both adversarial and natural, while test-time adaptation with semi-supervised learning is largely robust; e.g., the commercial model Pangram detects just 24.1\% of our adversarial AI-generated text, compared to 90.5\% for our test-time approach. We establish that test-time adaptation is a promising framework for AI text detection in the wild.

Large reasoning models achieve high accuracy through extended chain-of-thought but generate 5--8 more tokens than necessary, applying verbose reasoning uniformly regardless of problem difficulty. We propose Hint Tuning, a data-efficient approach that teaches models to calibrate reasoning depth. Our key insight: the corresponding instruct model serves as an ideal difficulty probe. By testing what the instruct model can solve with varying guidance, we automatically construct training data across three states: No-Hint (direct answer), Sparse-Hint (minimal prefix), and Full-Hint (complete reasoning). This converts the abstract challenge of difficulty labeling into a measurable consistency check between the instruct and reasoning models. With only 1K self-annotated samples, Hint Tuning achieves 24--66% token reduction (31.5% average) across mainstream reasoning models (Qwen3-Thinking, DeepSeek-R1-Distill) at multiple scales (4B--32B) while maintaining competitive accuracy on five benchmarks. Unlike methods requiring massive distillation datasets or expensive RL, we achieve superior efficiency through simple alignment with the instruct model's capabilities.