NeurIPS 2026 论文列表

In autonomous driving, multi-agent collaboration enhances individual perception capabilities through information sharing. However, in real-world applications, differences in models and training data among heterogeneous agents inevitably lead to domain gaps between shared features, bringing challenges to collaboration. To address this, existing methods typically rely on a one-to-one adaptation paradigm, necessitating specific training or fine-tuning for every new agent. This results in a large accumulation of adapters on the vehicle, which is unsustainable for resource-constrained edge devices. To overcome these limitations, we propose UniMTO, a Unified adaptation framework that enables Many-To-One adaptation across heterogeneous agents using a single, fixed adapter. Specifically, it treats domain adaptation as a pattern transformation task, disentangling domain-invariant content and domain-specific pattern from features, and achieving source-to-target pattern transformation through recombination based on a Mixture-of-Experts (MoE) mechanism. Extensive experiments on V2V4Real and V2X-Real demonstrate that UniMTO achieves superior generalization, and exhibits strong zero-shot adaptation capabilities, outperforming all state-of-the-art methods, enabling many-to-one domain adaptation with a single, fixed adapter.

LLM-based generation of SystemVerilog Assertions (SVA) is often reported as nearing saturation, with the strongest specialized model reaching ${\sim}76\%$ accuracy on NL2SVA-Human. We show that this aggregate hides a temporal gap: models that appear strong overall still collapse to a few implication templates on bounded-delay and liveness specifications. The core issue is that the dominant recipe, supervised fine-tuning on NL/SVA pairs, optimizes token-level mimicry rather than the property equivalence that defines SVA correctness. We introduce Reward-Weighted On-Policy Distillation (RWOPD), an on-policy distillation method that samples student rollouts, scores them with an open SymbiYosys+Z3 Property-Equivalence Checker (PEC), and applies a verifier-reward-weighted forward-KL gradient from a frozen 14B teacher on verifier-passable rollouts. This keeps the supervision dense at every response token while grounding both selection and loss weight in property-equivalent behavior. RWOPD distills CodeV-SVA-14B into a Qwen2.5-Coder-7B-Instruct student that sets a new state of the art on NL2SVA-Human and NL2SVA-Machine across pass@1, pass@5, and pass@10, surpassing both specialized prior SOTA models and 671B general-purpose baselines.

On-policy training plays an important role in modern machine learning, especially in reinforcement learning (RL) for language models. However, quantifying how training data influences the policy's behavior is difficult in on-policy settings: one must account for the cascading effect by which a training event affects its immediate policy update, which in turn affects future action samples, which further affects future policy updates, and so on. Attribution methods developed for off-policy and supervised learning (e.g. influence functions) do not account for such \emph{policy-data interaction}, and can therefore fail in on-policy settings. In this work, we formalize the problem of attributing behaviors to individual training events in the on-policy setting. We derive an influence measure that accounts for policy-data interaction while also being tractable to compute in practical training setups (e.g. LM finetuning) and for realistic optimizers (e.g. AdamW) at a cost that scales linearly in the total number of training events. Empirically, our influence measure accurately predicts the true effect of ablating specific events during RL, while traditional methods fail to do so. Further, we show that this influence accurately identifies helpful and harmful RL training events, allowing us to improve performance via influence-weighted retraining. Overall, we show that accounting for policy-data interaction is both necessary and tractable, allowing for principled training intervention in modern RL pipelines.

Aligned models can misbehave in several ways: they are often sycophantic, fall victim to jailbreaks, or fail to include appropriate safety warnings. Consistency training is a promising new alignment paradigm to mitigate such failures by training invariants into the model using contrastive input pairs. Existing consistency training procedures generate the supervision signal once, offline, and use supervised fine-tuning (SFT) to update the model. Unfortunately, the resulting models tend to merely memorize the surface forms of the training distribution and thus generalize poorly and regress in their capabilities. We introduce On-Policy Consistency Training (OPCT), a new consistency training approach where the objective is computed over the model's own responses to prompts, supervised by itself conditioned on corresponding contrastive prompts. We evaluate OPCT on three safety axes: sycophancy, jailbreaking, and safety awareness. Across three model families, OPCT outperforms its SFT counterpart on all safety desiderata. It nearly halves the sycophancy rate relative to baseline (8.1\% vs. 15.4\%, compared to 11.2\% for SFT). Under an adaptive per-target attacker, OPCT holds jailbreak defense success near 99\% on held-out jailbreak behaviors, whereas SFT achieves 87\% on average. On safety awareness, OPCT outperforms SFT in two out of three models, and matches it on the other. OPCT also largely avoids the capability regressions that SFT induces, such as a 28-point drop on MATH-500. Our results suggest that consistency training is best implemented as OPCT rather than as SFT, especially when generalization beyond the training distribution is desired.

Chain-of-Thought reasoning has enabled large language models to achieve substantial performance gains on complex tasks. However, these gains come at the cost of dramatically increased token consumption. This raises a fundamental question: is every token in the reasoning trace equally valuable? We present a diagnostic and optimization framework grounded in a key empirical finding: the value of tokens within a CoT reasoning sequence is highly non-uniform, and this non-uniformity can be effectively characterized by token-level log probability signals. We show that normalized log probability helps distinguish core tokens, which carry structural and decisive reasoning content, from redundant tokens, which are exploratory, low-confidence filler that contributes less directly to the final answer. Building on these findings, we formulate the TokenProbe framework around two empirical findings and one claim: findings identify token value inequality first and then establish TokenProbe as a core-token proxy, and the claim introduces an efficient GRPO objective positing that selectively compressing redundant tokens can yield Pareto improvements in the accuracy–token efficiency space. Empirically, our method preserves reasoning quality while reducing the token usage by 76% of the baseline. Under matched reasoning-length budgets, we show that it can even outperform strong flagship baselines like Gemini-3.1-Pro.

Dictionary learning has long been studied from both optimization and probabilistic perspectives. While formulations with element-wise sparsity regularization (e.g., L1-based sparse coding) admit well-established probabilistic interpretations, many structured variants that impose global constraints lack a clear and tractable generative view. In this paper, we revisit a class of practically effective yet theoretically under-explored dictionary learning methods that impose a simple global regularization on the number of activated dictionary atoms, which we term parsimoniously activated dictionary learning (PADL). We show that PADL admits an equivalent formulation as maximum a posteriori estimation under a structured generative model, with auxiliary latent variables that govern global activation patterns. This formulation allows us to derive generalization guarantees that are difficult to obtain under the original formulation. More importantly, it yields an analytical characterization of the tradeoff between sparsity, storage cost, and reconstruction accuracy, enabling data-driven estimation of optimal hyperparameters. Based on this connection, we develop an efficient and interpretable PADL algorithm that eliminates manual hyperparameter tuning, achieving improved reconstruction performance under comparable sparsity levels on visual benchmarks. We further demonstrate its practical utility in accelerating inference for vision-language models.

Direct Preference Optimization (DPO) has been widely adopted for aligning language models with human preferences. Recent empirical work has revealed that DPO induces remarkably sparse parameter updates, yet no theoretical explanation for this phenomenon has been offered. We analyze DPO's optimization dynamics under a neural tangent kernel linearization and show that the logistic structure of the DPO loss gives rise to an implicit maximum-margin bias. Building on this observation, we introduce the concept of dual sparsity: the DPO update direction is a sparse linear combination (data-level sparsity, driven by support vectors) of sparse gradient feature vectors (parameter-level sparsity, arising from gradient cancellation between preferred and dispreferred responses). We formalize this structure in our main theorem and present a framework relating it to weight disentanglement---a condition favorable for model merging. Experiments on GPT-2 Small and Medium with the Anthropic HH-RLHF and UltraFeedback datasets confirm both sources of sparsity and show that DPO reduces cross-task interference by up to $35 \times$ compared to supervised fine-tuning.

Designing small-molecule drugs that selectively bind to a target protein while avoiding unintended interactions with other proteins is essential for reducing side effects that contribute substantially to clinical attrition. Although structure-based drug design (SBDD) has advanced rapidly with deep learning, most generative models still optimize binding to a single target and disregard selectivity due to limited training data and unknown off-target interactions. Motivated by this gap, we introduce SelectBench, a comprehensive benchmark assessing the selectivity of generative SBDD models. SelectBench provides three curated evaluation tasks reflecting different real-world drug discovery scenarios: literature-based case studies, safety screening panels used in pharmaceutical development, and large-scale structure-based evaluation sets. We further introduce a suite of metrics to quantify selectivity and conduct a systematic evaluation of nine representative SBDD models, including methods explicitly trained for selectivity and widely-used inference-time guidance techniques. Our results show that even selectivity-trained models achieve only modest off-target discrimination and inference-time guidance provides limited improvements, underscoring the need for method improvement. SelectBench provides both a characterization of where current generative SBDD stands on selectivity and an open, extensible platform to accelerate progress.

In decoder-only (causal) transformers, the computation graph created by causal masking routes information through both direct-path attention and indirect paths formed by intermediate tokens. We denote these indirect paths between token pairs as their \textit{runways}. We argue that certain failure modes of causal transformers as observed by a growing body of recent works are likely exacerbated by a misalignment between these two information propagation modes. We formalize \textit{runway cascade} as a phenomenon whereby this misalignment results in redundancies and irrelevant information cascading to token representations despite adequately learned attention patterns. As a solution, we propose \textit{runway-aware rewiring} as a more explicit way of incorporating runway context directly into each token's direct-path attention. This mechanism re-wires the attention pattern for each token based on a summary of its runway landscape, enabling awareness of accumulating representational influences and allowing for more balanced information propagation. Our proposed methodology introduces no additional parameters and can seamlessly be integrated into standard attention mechanism. Empirically, our rewired transformer results in steady improvements in general language modeling as well as noticeably stronger information retrieval and extrapolation abilities compared to standard transformers.

Associative Recall (AR) is the cognitive ability to learn and retrieve links between items in memory. In NLP, AR is used as a benchmark for evaluating the in-context memory capacity of architectures such as Mamba, and has been found to strongly correlate with language modeling performance. This paper explores AR from the perspective of mechanistic interpretability, aiming to reverse-engineer the exact internal algorithm used by Mamba to perform recall. Our key insight is that Mamba performs recall by implicitly learning linear hash functions, and we identify the low-level circuit that enables this behavior. Building on these findings and inspired by theoretical tools in similarity-preserving hashing, such as the Johnson–Lindenstrauss (JL) lemma, we develop a theoretical framework for analyzing AR, which we term recall scaling laws. For example, given a model’s state, context length, and embedding dimensions, our theory predicts a lower bound on the largest vocabulary size that can be perfectly recalled in the AR task. Empirical results show that this bound is tight and predictive, offering insights into how AR capacity scales with vocabulary size, state size, embedding size, and model architecture.

Contrastive vision-language models such as CLIP are trained to align images and text, not to represent taxonomies. Yet empirical studies show that CLIP embeddings exhibit a hierarchical geometry aligned with the semantic hierarchy of concepts. A theoretical characterization of why such hierarchy emerges from a non-hierarchical objective remains open. In this paper, we formalize this phenomenon as Hierarchical Angular Separation (HAS), a geometric property whereby each non-leaf concept embedding is more similar to its descendants than to off-branch concepts. Empirically, we show that HAS already holds weakly in randomly initialized CLIP image embeddings, suggesting that hierarchical structure may be induced by the raw image geometry, while in text embeddings it emerges only after contrastive training. To explain this asymmetry between text and image embeddings, we analyze an unconstrained features model in which the image embeddings are fixed and only the text embeddings are trained. We show that when the image embeddings satisfy HAS with a positive margin, the global minimizer of the contrastive objective induces HAS in the text embeddings, thereby transferring hierarchical structure across modalities. The key mechanism is the hierarchical co-occurrence structure of image-text pairs, which biases text embeddings toward root-centered image prototypes. For a linear contrastive loss, we derive a closed-form solution that makes this mechanism explicit. We extend the analysis to the InfoNCE loss and establish analogous guarantees under sufficient regularization. Synthetic experiments validate the theory and show that HAS emerges even beyond the regimes covered by our analysis.

Verifiable Reward (RLVR) has emerged as a promising approach for enhancing the reasoning capabilities of large language models. However, it still remains unclear whether RLVR can extend reasoning capabilities beyond what is already encoded in the base model. In this paper, we study this question through the lens of *primacy bias* — the tendency of models to over-rely on already-learned representations, disproportionately improving performance on problems they already handle well while underexploring harder ones that require departing from those representations. Through our controlled experiments and large-scale analysis, we show that RLVR training on mixed data tends to further sharpen performance on easier problems while yielding only marginal gains on harder ones. We find that although hard problems lead to modest gains overall, they are actually the ones that help most in expanding the model’s reasoning capabilities and pushing the representations. However, prolonged training on these hard problems may also hurt and lead to forgetting on easy problems. These findings help to better understand some of the prior conflicting observations about RLVR, specifically, why on (OOD) domains where the base model performs poorly, improvements on hard problems can expand reasoning; while on well-trained domains, forgetting can reduce performance and mask genuine gains obtained from hard problems. Data and code for reproducing our experiments are available at: https://anonymous.4open.science/r/neurips26_sub-0DE8.

As large language models (LLMs) are increasingly trained on sensitive user data, understanding the fundamental cost of privacy in language learning becomes essential. We initiate the study of differentially private (DP) language identification and generation in the agnostic statistical setting, establishing algorithms and matching lower bounds that precisely quantify the cost of privacy. For both tasks, approximate $(\varepsilon, \delta)$-DP with constant $\varepsilon > 0$ recovers the non-private error rates: $\exp(-r(n))$ for identification (for any $r(n) = o(n)$) and $\exp(-\Omega(n))$ for generation. Under pure $\varepsilon$-DP, the exponents degrade by a multiplicative factor of $\min\\{1, \varepsilon\\}$, which we show is tight up to constants. Notably, for generation under pure DP with mild assumptions, the upper bound $\exp(-\min\\{1,\varepsilon\\} \cdot \Omega(n))$ matches the lower bound up to some constants, establishing an optimal rate. Our results show that the cost of privacy in language learning is surprisingly mild: absent entirely under approximate DP, and exactly a $\min\\{1,\varepsilon\\}$ factor in the exponent under pure DP.

A classic result in learning and optimization theory is that exponentiated gradient descent (EG) or the multiplicative weight update method computes an $\\epsilon$-approximate minimizer of an $\\ell_1$-Lipschitz function over the $d$-dimensional probability simplex $\\Delta_d :=\\{x \in \\mathbb{R}^d_{\\geq0}:\\sum_{i\in[d]}x_i=1\\}$ with $\\tilde{O}(\epsilon^{-2})$ queries to a first-order oracle. We investigate whether there are improved parallel algorithms for this fundamental problem. We show that, up to logarithmic factors, the answer is negative for $\\epsilon=\\tilde \\Omega(d^{-1/6})$ - any randomized algorithm that makes $O(\\mathrm{poly}(d))$ subgradient queries per round requires $\\tilde{\\Omega}(1/\\epsilon^2)$ rounds to output an $\epsilon$-optimal point with constant success probability. Moreover, to obtain this result we provide an analogous characterization of the parallel complexity of minimizing a $\ell_1$-Lipschitz convex function over the unit $\\ell_1$ ball. Previous lower bounds for non-constant $\\epsilon$ for this $\\ell_1$-Lipschitz convex optimization problem either made additional assumptions on the domain, or instead attained a bound of $\\Omega(\\epsilon^{-2/3})$ [DG19].

Algorithms have been estimated to increase AI pretraining FLOP efficiency by a factor of $22,000$ between 2012 and 2023 (Ho et al., 2024) . Running ablation experiments on key innovations from this time period, we are able to account for less than $10\times$ of these gains at small compute scales. Surveying the broader literature, we estimate that additional innovations not included in our ablations also account for less than $10\times$, yielding less than $100\times$ efficiency gains total in low-compute regimes. This leads us to conduct scaling experiments, which reveal that much of this efficiency gap can be explained by algorithms with scale-dependent efficiency improvements. In particular, we conduct scaling experiments between LSTMs and Transformers, finding exponent differences in their compute-optimal scaling law while finding little scaling difference for many other innovations. These experiments demonstrate that -- contrary to standard assumptions -- an algorithm's efficiency gains are tied to compute scale. Using experimental extrapolation and literature estimates, we account for $3,360\times$ efficiency gains over the same time period, with scale-dependent innovations accounting for close to 90\% of log-efficiency gains by 2023. Our results indicate that algorithmic progress for small models has been far slower than previously assumed, and that measures of algorithmic efficiency are strongly reference-dependent.

Kullback-Leibler (KL) regularization is widely used in offline decision-making and offers several benefits, motivating recent work on the sample complexity of offline learning with respect to \emph{KL-regularized performance metrics}. Nevertheless, the exact sample complexity of KL-regularized offline learning remains largely from fully characterized. In this paper, we study this question in the setting of multi-armed bandits (MABs). We provide a sharp analysis of KL-PCB~(Zhao et al., 2026), showing that it achieves a sample complexity of $\tilde{O}(\eta SAC^{\pi^\*}/\epsilon)$ under large regularization $\eta = \tilde{O}(\epsilon^{-1})$, and a sample complexity of $\tilde{\Omega}(SAC^{\pi^\*}/\epsilon^2)$ under small regularization $\eta = \tilde{\Omega}(\epsilon^{-1})$, where $\eta$ is the regularization parameter, $S$ is the number of contexts, $A$ is the number of arms, $C^{\pi^\*}$ is the policy coverage coefficient at the optimal policy $\pi^*$, $\epsilon$ is the desired sub-optimality, and $\tilde{O}$ and $\tilde{\Omega}$ hide all poly-logarithmic factors. We further provide a pair of sharper sample complexity lower bounds, which matches the upper bounds over the entire range of regularization strengths. Overall, our results provide a nearly complete characterization of offline multi-armed bandits with KL regularization.

The goal of this paper is to strengthen the reasoning of Omnimodal Large Language Models (Omni-LLMs) at inference time, without additional training. These models jointly process video, audio, and text, and given the large number of tokens they consume, how attention is routed across them is central to their behaviour. We focus specifically on attention sinks, tokens that absorb a disproportionate share of attention mass regardless of their semantic content, to understand how this routing unfolds. To this end, we conduct a systematic analysis of sink behaviour in Omni-LLMs. Our analysis yields two key findings: (i) high sink attention does not solely indicate head redundancy, suggesting that sink value representations play additional functional roles; (ii) the sink value vector acts as a shared bias added to every token's output, serving as a global signal that organises the representation as a whole. Building on this, we propose OutRo, which correspondingly aligns non-sink token representations with the sink in feature space, and relaxes the causal mask for sink tokens at an early layer to sharpen this bias before the rest of decoding proceeds. This design enhances the reasoning process without requiring additional forward passes or access to attention maps. Based on extensive experiments, OutRo consistently improves performance on seven video QA benchmarks and demonstrates strong generalisation, while incurring only a 1.1\x decoding overhead.

There is an extensive literature that studies how to find optimal policies in resource allocation problems, taking the underlying design parameters that define the allocation, such as what data is collected, how many people can be served, and quality of service as fixed constraints. Yet, from a planner's perspective, these design parameters are themselves optimization variables that are just as important in determining overall welfare as selecting the optimal targeting rule for a given set of constraints. This realization motivates a rich set of meta-design questions exploring how planners should make principled decisions about investments in prediction, capacity constraints, and treatment quality, all of which lie upstream of classical policy optimization. Building on initial theoretical work in this space, our paper has three main contributions. First, we formally define the broad meta-design space of resource allocation problems. Second, we develop empirical tools that enable practitioners to reliably navigate it. Third, we demonstrate the framework in two real-world case studies on German employment services and targeted cash transfer programs in Ethiopia.

Diffusion models are central to modern generative modeling, and understanding how they balance memorization and generalization is critical for reliable deployment. Recent work has shown that memorization in diffusion models is shaped by training dynamics, with generalization and memorization emerging at different stages of training. However, deployed diffusion models are often further distilled, introducing an additional training phase whose impact on memorization remains unclear. In this work, we analyze how distillation reshapes memorization behavior in diffusion models, taking consistency distillation as a representative framework. Empirically, we show that when applied to a teacher model that has memorized data, consistency distillation significantly reduces transferred memorization in the student while preserving, and sometimes improving, sample quality. To explain this behavior, we provide a theoretical analysis using a random feature neural network model [Bonnaire et al., 2025], showing that consistency distillation suppresses unstable feature directions associated with memorization while preserving stable, generalizable modes. Our findings suggest that distillation can serve not only as an acceleration tool, but also as a mechanism for improving the memorization-generalization trade-off.