NeurIPS 2026 论文列表

Flexible switching between reasoning trajectories (i.e., thoughts switching) has significantly enhanced the reasoning capabilities of Large Reasoning Models (LRMs). However, existing models often switch excessively yet fail to sustain promising reasoning thoughts---a phenomenon termed ''under-thinking''. While recent efforts suppress switching to mitigate this, such over-correction may discard valuable trajectories. To address this challenge, we propose Steady Thought (ST), a novel thought-level preference optimization framework. ST formalizes under-thinking as a preference issue at switching points, forcing single-path continuations from identified points to yield high-quality reasoning. Then, ST performs thought-level preference optimization by treating the newly generated response as preferred and the original one as dis-preferred. Experiments across multiple models and datasets show that ST effectively reduces token consumption while maintaining or even improving accuracy. It reduces output length by up to 63.8% while improving accuracy by up to 11.4%. Further analysis suggests that ST helps models acquire a generalizable ability to reduce redundant switching across domains and languages.

The performance of *Graph Domain Adaptation* (*GDA*) is fundamentally limited by *covariate-driven* discrepancies between source and target graphs, captured by *Covariate Shift* (*CS*) in the joint feature-structure space beyond label-related formulations. We show that CS induces bias at two non-interchangeable stages of *Graph Neural Network* (*GNN*) computation: *Feature Shift* (*FS*) distorts representations before message passing, while *Feature-Conditional Structure Shift* (*FCSS*) biases propagation, rendering single-stage alignment intrinsically insufficient. Based on this, we propose *Dual Alignment for Covariate Shift* (*DACS*), a stage-aware GDA framework that aligns discrepancies at their source. DACS mitigates FS via adversarial feature alignment and addresses FCSS through layer-wise reweighting to correct propagation bias, followed by final adversarial alignment for residual mismatch. This design follows directly from the stage-wise structure of CS rather than heuristic combinations of alignment modules. We further show that representation and propagation biases correspond to distinct components of target-domain error that cannot be eliminated in isolation. Experiments on synthetic and real-world benchmarks demonstrate that DACS consistently outperforms prior methods, especially under complex and coupled distribution shifts.

The widespread deployment of generative AI has made it increasingly difficult to distinguish synthetic content from real data. Consequently, synthetic data is inevitably incorporated into the training pipelines of future model generations, forming a *self-consuming training loop*. Prior work has studied the effects of such recursive self-consuming training, but analyses have largely been limited to *isolated* models, where a model consumes only its own synthetic data, or to simplified interactions between two models. This paper takes a first step toward understanding *networked* self-consuming generative models, in which multiple models consume synthetic data generated by one another through complex interaction pathways. We introduce a theoretical framework representing models as nodes in a directed, weighted graph, with edge weights governing the flow of synthetic data among models. Using this framework, we analyze the long-term behavior of networked models under retraining dynamics, establishing conditions for convergence and characterizing the resulting fixed points. We further investigate how the system's long-term stability and diversity are shaped by each model’s access to real data, cross-model data consumption, and the structure of the interaction graph.

Latent action models (LAMs) aim to learn compact representations of state transitions directly from visual observations, but they suffer from a fundamental ambiguity: the latent code can encode target-state information rather than true dynamics. Existing approaches address this issue through restrictive bottlenecks, which reduce leakage at the cost of limiting expressivity. We propose StaDy, a regularization framework that instead enforces conditional informativeness: the latent code should aid prediction only when paired with the source state, while remaining uninformative about the target on its own. Concretely, we introduce a likelihood-matching objective that aligns the decoder’s predictions conditioned solely on the latent variable with its unconditional predictions. This discourages target memorization without constraining latent capacity. Experiments in video modeling show that, unlike bottleneck-based methods, StaDy scales effectively with increased latent dimensionality, reducing static content leakage while improving the representation of more complex dynamics.

Online platforms connect users with relevant products and services using ads. A key challenge is that a user's search query often leaves their true intent ambiguous. Typically, platforms passively predict relevance based on available signals and in some cases offer query refinements. The shift from traditional search to conversational AI provides a new approach. When a user's query is ambiguous, a Large Language Model (LLM) can proactively offer several clarifying follow-up prompts. In this paper we consider the following: what if some of these follow-up prompts can be ``sponsored,'' i.e., selected for their advertising potential. How should these ``suggestion slots'' be allocated? And, how does this new mechanism interact with the traditional ad auction that might follow? This paper introduces a formal model for designing and analyzing these interactive platforms. We use this model to investigate a critical engineering choice: whether it is better to jointly optimize the user interaction and the final ad auction, or to decouple them into separate mechanisms for the suggestion slots and another for the subsequent ad slot. We show that the VCG mechanism can be adopted to jointly optimize the sponsored suggestion and the ads that follow; while this mechanism is more complex, it achieves outcomes that are efficient and truthful. On the other hand, we prove that the simple-to-implement modular approach suffers from strategic inefficiency: its Price of Anarchy is unbounded, for various natural choices for the two mechanisms. While we show that this inefficiency can be mitigated through revenue redistribution, we argue that the resulting mechanism sacrifices the practical simplicity that motivated the modular design in the first place.

E(3)-equivariant networks are promising for 3D atomistic system modeling, yet their scalability is limited by the $O(L^6)$ complexity of the Clebsch-Gordan Tensor Product (CGTP). The recently proposed Gaunt Tensor Product (GTP) reduces the complexity but is unable to capture the antisymmetric paths, resulting in incomplete expressivity. In this work, we present SpinGT, an approach to overcome the GTP incompleteness by generalizing from scalar functions to Spin-Weighted Spherical Harmonics (SWSH). By relying on the algebraic properties of SWSH, SpinGTP recovers the missing antisymmetric interactions while maintaining the asymptotic efficiency of GTP. It also allows for a more expressive equivariant basis that naturally accounts for the parity-odd components of tensor products. We evaluate SpinGTP across diverse benchmarks, including Tetris, 3BPA, SPICE-MACE-OFF, and OC20. Our results show that SpinGTP achieves accuracies comparable to full CGTP. Notably, by explicitly capturing antisymmetric paths, SpinGTP exhibits superior performance in tasks involving chiral materials and non-centrosymmetric geometries. This work provides a scalable, complete, and mathematically rigorous path toward high-order equivariance in large-scale 3D atomistic system simulations.

We prove that the speedup of exact speculative decoding is governed by a per-token rate, not by the path-level prefix-overlap capacity that folklore suggests, and that the gap between the two can be unbounded. The two candidates are the per-token Leviathan rate σ_L(P, Q) and the path-level coupling capacity Ω_L(P, Q) = Σ_{t=1}^L ov(P_{≤t}, Q_{≤t}). We show σ_L ≤ Ω_L always, the inequality is strict already on i.i.d. products, and there exists an explicit fixed-precision binary trajectory-exact rejection-style protocol on L = 2 with E[τ] = 19/16 > 9/8 = σ_L, ruling out σ_L as a universal upper bound across all trajectory-exact rounds. Inside the deployable admissible Leviathan-style clipped-ratio class, σ_L is the tight rate, and the speculation decoding complexity satisfies SDC_{B,L}^Q(M; u, T) = Θ(T / (σ_{L,B}(P_M^{u,T}, Q) + 1)), achieved by iterated Leviathan and matched up to constants by Wald. On the path side, Ω_L exhibits a sharp two-sided dichotomy in the prefix-KL profile η_t: Σ_t √(η_t / 2) = o(L) forces Ω_L = L − o(L), while geometrically bounded overlap o_t ≤ ρ^t forces Ω_L ≤ 1/(1−ρ) uniformly in L, with no unconditional KL converse. Two explicit fixed-precision Transformer families of size O(log L) realize the two extremes under the same three-element proposal family Q = {Q_flat, Q_const_0, Q_const_1}: a constant-readout family F⁺ has SDC ≤ T / ((1−ε)L + 1), while an alternating-parity family F⁻ has SDC ≥ T/5, an unconditional Θ(L) architecture separation. Speculative decoding is not path coupling; it is per-token clipped-ratio acceptance, and the architecture of the target controls the rate.

Forward-backward (FB) representations provide a powerful framework for learning the successor representation (SR) in continuous spaces by enforcing a low-rank factorization. However, a fundamental spectral mismatch often exists between the high-rank transition dynamics of continuous environments and the low-rank bottleneck of the FB architecture, making accurate low-rank representation learning difficult. In this work, we analyze temporal abstraction as a mechanism to mitigate this mismatch. By characterizing the spectral properties of the transition operator, we show that temporal abstraction acts as a low-pass filter that suppresses high-frequency spectral components. This suppression reduces the effective rank of the induced SR while preserving a formal bound on the resulting value function error. Empirically, we show that this alignment is a key factor for stable FB learning, particularly at high discount factors where bootstrapping becomes error-prone. Our results identify temporal abstraction as a principled mechanism for shaping the spectral structure of the underlying MDP and enabling effective long-horizon representations in continuous control.

The standard metric LP for correlation clustering has $\Theta(n^2)$ pairwise marginals and $\Theta(n^3)$ triangle inequalities, while dense pairwise supervision is often the primary bottleneck in large relational pipelines. We separate three sparsification questions that are often treated together: preserving the objective of every integral clustering, rounding from a sparse set of LP marginals, and clustering when the signed weighted input itself is only sparsely observed. First, we prove that the VC dimension of the signed-edge disagreement class induced by all clusterings of $n$ vertices is exactly $n-1$, so weighted edge sampling yields additive $\varepsilon$-coresets of size $\tilde O(n/\varepsilon^2)$ with optimal $n$-dependence. Second, we introduce *Sparse-LP-Pivot*, which imputes missing LP marginals from triangle witnesses, and analyze it at two levels: a universal but coarse perturbation bound for every Lipschitz LP-PIVOT rule, and a sharper sparse-rounding theorem under an explicit weighted influence-stability certificate. For pseudometric-weighted CC, this gives a $10/3$ exact-marginal baseline and an unconditional coarse sparse-marginal bound; a sharp same-edge bound governed by $\overline{\Gamma}_w$ follows conditionally if the rounding rule is additionally proved to satisfy edge-weight-dominated influence stability. Third, we show that uniform marginal sampling has a triangle-witness threshold at $m=\Theta(n^{3/2})$ for typical pairs, while $m=O(n^{3/2}\sqrt{\log n})$ is sufficient for all pairs with high probability. Finally, in the stricter sparse edge-observation model, we prove an $\Omega(n^{3/2})$ lower bound on the expected approximation ratio from $o(n)$ uniformly sampled edges for general weighted instances. Experiments illustrate the witness transition and the proposed diagnostics; the formal guarantees are stated independently of those empirical observations.

Multiplicative motor and observation noise, along with internal noise corrupting computation, are central features of the sensorimotor system in biological and robotic agents. Yet, analytical solutions for stochastic optimal control are largely restricted to additive noise models and neglect internal noise. Here, we consider the problem of finding optimal control and filter laws for partially observable stochastic linear systems under quadratic costs with multiplicative control and observation noise, as well as internal noise. We provide an efficient, analytically derived coordinate-descent algorithm that computes mutually optimal linear control and filter laws by reformulating the problem as a constrained optimization. Our method provably guarantees monotonic decrease of the expected cost and convergence to a critical point, and overcomes the suboptimality and incompleteness of prior analytical approaches. Compared to state-of-the-art numerical methods, it achieves orders-of-magnitude computational speedups. Our solution also proves instrumental in revealing novel internal-noise-dependent, task-structured control strategies in a redundant arm control task.

Multilingual large language models often produce inconsistent reasoning and answers across semantically equivalent prompts in different languages. Prior work suggests that intermediate representations can be relatively language-agnostic, but generation becomes increasingly language-specific as models commit to discrete output tokens. This is problematic because language-specific lexical choices can cause semantically equivalent reasoning paths to diverge across languages. This motivates searching for a cross-lingual alignment signal that is less tied to any single vocabulary item or script. We propose **SOLAR**, an auxiliary objective for supervised fine-tuning that aligns _soft-token_ representations across languages, using English as a pivot. Soft tokens are probability-weighted mixtures over the vocabulary embeddings, yielding continuous representations that can aggregate information from semantically related tokens across languages. We then align each non-English soft-token summary to its English counterpart in the shared embedding space. Across four multilingual reasoning benchmarks, SOLAR improves accuracy by up to +17.7 points over the base model and +3.8 over standard supervised fine-tuning, with the largest gains on low-resource languages. SOLAR also strengthens final-layer cross-lingual similarity and substantially reduces language-cluster separability, suggesting that aligning soft-token representations helps preserve shared semantic structure during multilingual reasoning.

Existing soccer commentary models are often designed for pre-segmented clips or localized events. When deployed on untrimmed full-match videos with sliding windows, they can produce delayed, repeated, or poorly synchronized commentary. Recent streaming video-language models make low-latency full-match narration feasible, but they often favor general real-time descriptions rather than professional, event-grounded soccer commentary. They may describe nearby actions while failing to mention key events such as goals, cards, substitutions, and offsides correctly and on time. In this paper, we introduce **SoccerNarrate**, a data, model, and evaluation framework that bridges low-latency streaming narration with event-grounded professional soccer commentary. First, we construct **SoccerNarrate-Data**, a large-scale time-anchored dataset with word-level ASR alignment, quality filtering, and entity calibration. Second, we perform **macro-window preference alignment(MWPA)**, which aligns the model with complete event-level commentary semantics by comparing causal multi-step rollouts from the same streaming prefix while keeping second-level inference unchanged. We further use event-mismatched human commentaries as counterfactual negatives to emphasize event semantics over commentary style. Third, we introduce **SoccerNarrate-Eval**, an event-centric benchmark based on temporally constrained event entailment. Full-match experiments show that SoccerNarrate improves event coverage and event-aligned precision over strong offline and streaming baselines. Data and models will be released.

This paper studies the role of sinks and diagonal patterns as attention switch and anti-oversmoothing mechanisms. We analyze geometric conditions under which sinks can be represented, showing a necessary alignment between the embedding of the sink and all other embeddings. Next, we refine the current understanding of the role of sinks in oversmoothing prevention: we specify the conditions under which dense attention provably smooths more than sparse attention, and empirically verify that such conditions are often satisfied in practice. We further prove an equivalence between sinks and hard attention switch, in which the output of the attention is identically 0. Finally, we relax the hard attention switch by allowing token self-communication: we provide a quantitative comparison of the costs of representing sinks vs.\ diagonal patterns, showing why sinks are favored in pretrained transformers. The introduction and analysis of diagonal patterns and the generalization of the attention switch close the gap between what oversmoothing prevention requires and what sinks provide, while also establishing when and why attention layers act like MLPs if token communication is not necessary.

Large language models (LLMs) can generate student-like responses fluently enough to serve as simulated students, i.e., virtual learners used to train and evaluate AI tutors and human educators. Yet such simulators are typically evaluated by output similarity to real students, not by whether they behave like students with coherent misconceptions during interaction. We introduce a controlled framework for evaluating misconception faithfulness, whether a simulator maintains a misconception-driven belief state and updates selectively when feedback addresses the underlying misconception. Central to our framework is a misconception-contrastive feedback protocol that compare targeted feedback against two controls: misaligned feedback, targeting a different but plausible misconception, and generic feedback, which only signals that the answer is incorrect. We propose Selective Flip Score (SFS), which quantifies how much more often a simulator flips its answer under targeted feedback than under the contrastive controls. Across seven LLMs (4B–120B), multiple datasets, and prompting strategies, simulators exhibit near-zero SFS, correcting their answers at similarly high rates regardless of feedback relevance to the true misconception. Further analysis reveals a sycophantic failure mode: models behave less like students with stable misconceptions and more like problem solvers who treat any corrective signal as a cue to abandon the simulated misconception and recompute from internal knowledge. To improve faithfulness, we develop a post-training pipeline spanning supervised fine-tuning (SFT), preference optimization, and reinforcement learning (RL) with an SFS-aligned reward. SFT yields notable gains up to +0.56; SFS-aligned RL provides more consistent further improvements than preference optimization. Our results establish misconception faithfulness as a challenging but trainable property of student simulators, motivating a shift from static output matching toward interaction- and belief-aware student modeling.

Multi-objective Bayesian optimization (MOBO) is commonly approached through specialized acquisition functions or scalarization schemes designed to explicitly account for trade-offs among non-preferential objectives. In this work, we show that such complexity might be unnecessary. We propose a framework that extends standard single-objective acquisition functions directly to the multi-objective setting through a hypervolume-based transformation. We further extend hedge strategies for acquisition functions, which are typically used only in single-objective optimization, to the multi-objective regime. Our approach requires minimal modification to existing Bayesian optimization pipelines and avoids the need for bespoke multi-objective formulations. We demonstrate how a broad class of commonly used single-objective acquisition functions and hedge strategies can be adapted in a principled manner to handle multiple objectives, while preserving their intuitive interpretation and computational efficiency. Empirically, we evaluate the proposed methods across a range of synthetic and real-world multi-objective benchmarks. Despite their simplicity, our extensions consistently match or outperform more complex state-of-the-art MOBO methods in terms of optimization performance and sample efficiency. These results suggest that effective multi-objective Bayesian optimization can be achieved by reusing and carefully extending well-established single-objective acquisition strategies, offering a simpler and more flexible alternative to existing approaches.

We present the first algorithms for generalized linear contextual bandits under shuffle differential privacy and joint differential privacy. While prior work on private contextual bandits has been restricted to linear reward models---which admit closed-form estimators---generalized linear models (GLMs) pose fundamental new challenges: no closed-form estimator exists, requiring private convex optimization; privacy must be tracked across multiple evolving design matrices; and optimization error must be explicitly incorporated into regret analysis. We address these challenges under two privacy models and context settings. For stochastic contexts, we design a shuffle-DP algorithm with regret $\tilde{O}\!\big(d^{3/2}\sqrt{T \log T} + d^{5/4} \sqrt{T/\varepsilon} (\log T)^{3/4}\big)$ in the dominant term, matching the non-private rate $\tilde{O}(d \sqrt{T \log T})$ in the leading-in-$T$ term up to a multiplicative factor of $\sqrt d$ and a privacy correction term of $\tilde{O}(d^{5/4}\sqrt{T \log T/\varepsilon})$. For adversarial contexts, we provide a joint-DP algorithm with regret $\tilde{O}\!\big(d\sqrt{T}\log T + d^{3/4}\sqrt{T/\varepsilon}\,(\log T)\,(d+\log T)^{1/4}\big)$ matching the non-private rate $\tilde{O}(d\sqrt{T}\log T)$ in the leading term. Unlike prior work on locally private GLM bandits, our methods require no spectral assumptions on the context distribution beyond $\ell_2$ boundedness.

Emergent misalignment (EM) has been linked to the activation of misaligned persona vectors and evil character traits, suggesting that EM operates through disruption of the model's aligned character rather than direct learning of harmful content. Motivated by this connection, we study self-generated text recognition (SGTR) finetuning as a character-targeted intervention that is distinct from existing in-training defenses. We conduct two-stage finetuning experiments across three models (GPT-4.1, Qwen2.5-32B-Instruct, Seed-OSS-36B-Instruct) and multiple EM datasets to compare SGTR finetuning against benign finetuning baselines (correct domain-specific data, general knowledge, and word counting) to find it an effective defense in both reversal and prevention settings. We find that all interventions produce comparable EM reversal, but only when restoring capabilities that EM had degraded. For prevention, only SGTR finetuning consistently reduces misalignment without exacerbating any individual metric, suggesting that character fortification specifically drives prevention. We provide further evidence for EM's relation to the LLM's default character by showing that EM finetuning induces diversity into the LLM's identity self-reports, artificially corrupting self-recognition exacerbates misalignment caused by EM finetuning, and that removing the model's identity-bearing system prompt substantially reduces the effect of EM finetuning. Together, these findings reframe EM not as the adoption of a coherent misaligned persona but as the destabilization of aligned character.

While large language models (LLMs) are trained to align with human values, their generations may still violate safety constraints. A growing line of work addresses this problem by modifying the model’s sampling policy at decoding time using a safety reward. However, existing decoding-time steering methods often intervene unnecessarily, modifying generations that would have been safe under the base model. Such unnecessary interventions are undesirable, as they can distort key properties of the base model such as helpfulness, fluency, style, and coherence. We propose a new test-time steering method designed to reduce such unnecessary interventions while improving the safety of unsafe responses. Our approach filters tokens using a value-based safety criterion and provides an explicit bound on the probability of false interventions. A single threshold hyperparameter controls this bound, allowing practitioners to trade off higher rates of unnecessary intervention for better output safety. Across multiple datasets and experiments, we show that our value-filtered decoding method outperforms existing baselines, achieving better trade-offs between safety, helpfulness, and similarity to the base model.

Rotation invariance is a fundamental requirement across many computer vision tasks. Historically, this inductive bias has been encoded through hand-crafted rotation-invariant representations. These are compact, interpretable, and fast to compute, but they come at the cost of descriptive power. More recently, architectures achieve inductive bias through learned representations. These are highly descriptive and achieve strong empirical performance, at the cost of efficiency and interpretability. In this work, we propose an alternative at the intersection of both paradigms. We introduce the selective disk bispectrum (SDB), a complex-valued rotation-invariant vector that preserves all information about the image except its orientation. Our key theoretical contributions are the selective disk bispectrum, its inversion, its (reduced) spatial and computational complexities (compared to the full disk bispectrum), and its expectation and variance under noise. Furthermore, we propose a numerical SDB approximation and provide theoretical guarantees for its accuracy and rotation invariance. Empirically, we validate SDB's invariance and robustness to noise classification tasks. We test our reconstruction algorithm on multi-reference alignment of rotated images.