NeurIPS 2025 论文列表

Efficient optimization is essential for training large language models. Although intra-layer selective updates have been explored, a general mechanism that enables fine-grained control while ensuring convergence guarantees is still lacking. To bridge this gap, we propose \textbf{MGUP}, a novel mechanism for selective updates. \textbf{MGUP} augments standard momentum-based optimizers by applying larger step-sizes to a selected fixed proportion of parameters in each iteration, while applying smaller, non-zero step-sizes to the rest. As a nearly {plug-and-play} module, \textbf{MGUP} seamlessly integrates with optimizers such as AdamW, Lion, and Muon. This yields powerful variants such as \textbf{MGUP-AdamW}, \textbf{MGUP-Lion}, and \textbf{MGUP-Muon}. Under standard assumptions, we provide theoretical convergence guarantees for \textbf{MGUP-AdamW} (without weight decay) in stochastic optimization. Extensive experiments across diverse tasks, including MAE pretraining, LLM pretraining, and downstream fine-tuning, demonstrate that our \textbf{MGUP}-enhanced optimizers achieve superior or more stable performance compared to their original base optimizers. We offer a principled, versatile, and theoretically grounded strategy for efficient intra-layer selective updates, accelerating and stabilizing the training of large-scale models. The code is publicly available at https://github.com/MaeChd/MGUP.

We study gradient flows for loss landscapes of fully connected feedforward neural networks with commonly used continuously differentiable activation functions such as the logistic, hyperbolic tangent, softplus or GELU function. We prove that the gradient flow either converges to a critical point or diverges to infinity while the loss converges to an asymptotic critical value. Moreover, we prove the existence of a threshold $\varepsilon>0$ such that the loss value of any gradient flow initialized at most $\varepsilon$ above the optimal level converges to it. For polynomial target functions and sufficiently big architecture and data set, we prove that the optimal loss value is zero and can only be realized asymptotically. From this setting, we deduce our main result that any gradient flow with sufficiently good initialization diverges to infinity. Our proof heavily relies on the geometry of o-minimal structures. We confirm these theoretical findings with numerical experiments and extend our investigation to more realistic scenarios, where we observe an analogous behavior.

Reward model (RM) plays a pivotal role in reinforcement learning with human feedback (RLHF) for aligning large language models (LLMs). However, classical RMs trained on human preferences are vulnerable to reward hacking and generalize poorly to out-of-distribution (OOD) inputs. By contrast, strong LLM judges equipped with reasoning capabilities demonstrate superior generalization, even without additional training, but incur significantly higher inference costs, limiting their applicability in online RLHF. In this work, we propose an uncertainty-based routing framework that efficiently complements a fast RM with a strong but costly LLM judge. Our approach formulates advantage estimation in policy gradient (PG) methods as pairwise preference classification, enabling principled uncertainty quantification to guide routing. Uncertain pairs are forwarded to the LLM judge, while confident ones are evaluated by the RM. Experiments on RM benchmarks demonstrate that our uncertainty-based routing strategy significantly outperforms random judge calling at the same cost, and downstream alignment results showcase its effectiveness in improving online RLHF.

As Video Large Multimodal Models (VLMMs) rapidly advance, their inherent complexity introduces significant safety challenges, particularly the issue of mismatched generalization where static safety alignments fail to transfer to dynamic video contexts. We introduce SafeVid, a framework designed to instill video-specific safety principles in VLMMs. SafeVid uniquely transfers robust textual safety alignment capabilities to the video domain by employing detailed textual video descriptions as an interpretive bridge, facilitating LLM-based rule-driven safety reasoning. This is achieved through a closed-loop system comprising: 1) generation of SafeVid-350K, a novel 350,000-pair video-specific safety preference dataset; 2) targeted alignment of VLMMs using Direct Preference Optimization (DPO); and 3) comprehensive evaluation via our new SafeVidBench benchmark. Alignment with SafeVid-350K significantly enhances VLMM safety, with models like LLaVA-NeXT-Video demonstrating substantial improvements (e.g., up to 42.39%) on SafeVidBench. SafeVid provides critical resources and a structured approach, demonstrating that leveraging textual descriptions as a conduit for safety reasoning markedly improves the safety alignment of VLMMs in complex multimodal scenarios.

This position paper argues that there is an urgent need to restructure markets for the information that goes into AI systems. Specifically, small-to-medium sized producers of information (such as journalists, news organizations, researchers, and creative professionals) need to be able to appoint representatives who can carry out "collective bargaining" with AI product builders in order to receive a reasonable terms and a fair return on the informational value they contribute. Obstacles to this market structure can be removed through technical work that facilitates collective bargaining in the information economy (e.g., explainable data value estimation and federated data management tools) and regulatory/policy interventions (e.g., support for trusted data intermediary organizations that represent guilds or syndicates of information producers). We argue that without collective bargaining in the information economy, AI will exacerbate a large-scale "information market failure" that will lead not only to undesirable concentration of capital, but also to a potential "ecological collapse" in the informational commons. On the other hand, collective bargaining in the information economy can create market conditions necessary for a pro-social AI future. We provide concrete actions that can be taken to support a coalition-based approach to achieve this.

Searching through chemical space is an exceptionally challenging problem because the number of possible molecules grows combinatorially with the number of atoms. Large, autoregressive models trained on databases of chemical compounds have yielded powerful generators, but we still lack robust strategies for generating molecules with desired properties. This molecular search problem closely resembles the "alignment" problem for large language models, though for many chemical tasks we have a specific and easily evaluable reward function. Here, we introduce an algorithm called energy rank alignment (ERA) that leverages an explicit reward function to produce a gradient-based objective that we use to optimize autoregressive policies. We show theoretically that this algorithm is closely related to proximal policy optimization (PPO) and direct preference optimization (DPO), but has a minimizer that converges to an ideal Gibbs-Boltzmann distribution with the reward playing the role of an energy function. Furthermore, this algorithm is highly scalable, does not require reinforcement learning, and performs well relative to DPO when the number of preference observations per pairing is small. We deploy this approach to align molecular transformers and protein language models to generate molecules and protein sequences, respectively, with externally specified properties and find that it does so robustly, searching through diverse parts of chemical space.

We study the problem of infrequent exploration in linear bandits, addressing a significant yet overlooked gap between fully adaptive exploratory methods (e.g., UCB and Thompson Sampling), which explore potentially at every time step, and purely greedy approaches, which require stringent diversity assumptions to succeed. Continuous exploration can be impractical or unethical in safety-critical or costly domains, while purely greedy strategies typically fail without adequate contextual diversity. To bridge these extremes, we introduce a simple and practical framework, INFEX, explicitly designed for infrequent exploration. INFEX executes a base exploratory policy according to a given schedule while predominantly choosing greedy actions in between. Despite its simplicity, our theoretical analysis demonstrates that INFEX achieves instance-dependent regret matching standard provably efficient algorithms, provided the exploration frequency exceeds a logarithmic threshold. Additionally, INFEX is a general, modular framework that allows seamless integration of any fully adaptive exploration method, enabling wide applicability and ease of adoption. By restricting intensive exploratory computations to infrequent intervals, our approach can also enhance computational efficiency. Empirical evaluations confirm our theoretical findings, showing state-of-the-art regret performance and runtime improvements over existing methods.

Achieving balanced alignment of large language models (LLMs) in terms of Helpfulness, Honesty, and Harmlessness (3H optimization) constitutes a cornerstone of responsible AI. Existing methods like data mixture strategies face limitations, including heavy reliance on expert knowledge and conflicting optimization signals. While model merging offers parameter-level conflict-resolution strategies through integrating specialized models' parameters, its potential for 3H optimization remains underexplored. This paper systematically compares the effectiveness of model merging and data mixture methods in constructing 3H-aligned LLMs for the first time, revealing previously overlooked collaborative and conflict relationships among the 3H dimensions and discussing the advantages and drawbacks of data mixture (\textit{data-level}) and model merging (\textit{parameter-level}) methods in mitigating the conflict for balanced 3H optimization. Specially, we propose a novel \textbf{R}eweighting \textbf{E}nhanced task \textbf{S}ingular \textbf{M}erging method, \textbf{RESM}, through outlier weighting and sparsity-aware rank selection strategies to address the challenges of preference noise accumulation and layer sparsity adaptation inherent in 3H-aligned LLM merging. Extensive evaluations can verify the effectiveness and robustness of RESM compared to previous data mixture (2\%-5\% gain) and model merging (1\%-3\% gain) methods in achieving balanced LLM alignment.

Entity alignment (EA) aims to merge two knowledge graphs (KGs) by identifying equivalent entity pairs. Existing methods can be categorized into symbolic and neural models. Symbolic models, while precise, struggle with substructure heterogeneity and sparsity, whereas neural models, although effective, generally lack interpretability and cannot handle uncertainty. We propose NeuSymEA, a unified neuro-symbolic reasoning framework that combines the strengths of both methods to fully exploit the cross-KG structural pattern for robust entity alignment. NeuSymEA models the joint probability of all possible pairs' truth scores in a Markov random field, regulated by a set of rules, and optimizes it with the variational EM algorithm. In the E-step, a neural model parameterizes the truth score distributions and infers missing alignments. In the M-step, the rule weights are updated based on the observed and inferred alignments, handling uncertainty. We introduce an efficient symbolic inference engine driven by logic deduction, enabling reasoning with extended rule lengths. NeuSymEA achieves a significant 7.6\% hit@1 improvement on $DBP15K_{ZH-EN}$ compared with strong baselines and demonstrates robustness in low-resource settings, achieving 73.7\% hit@1 accuracy on $DBP15K_{FR-EN}$ with only 1\% pairs as seed alignments. Codes are released at https://github.com/chensyCN/NeuSymEA-NeurIPS25.

Training time-series forecast models presents unique challenges in designing effective learning objectives. Existing methods predominantly utilize the temporal mean squared error, which faces two critical challenges: (1) label autocorrelation, which leads to bias from the label sequence likelihood; (2) excessive amount of tasks, which increases with the forecast horizon and complicates optimization. To address these challenges, we propose Time-o1, a transformation-augmented learning objective for training time-series forecasting models. The central idea is to transform the label sequence into decorrelated components with discriminated significance. Models are then trained to align the most significant components, thereby effectively mitigating label autocorrelation and reducing task amount. Extensive experiments demonstrate that Time-o1 achieves state-of-the-art performance and is compatible with various forecast models. Code is available at https://github.com/Master-PLC/Time-o1.

This position paper argues that the theoretical inconsistency often observed among Responsible AI (RAI) metrics, such as differing fairness definitions or trade-offs between accuracy and privacy, should be embraced as a valuable feature rather than a flaw to be eliminated. We contend that navigating these inconsistencies, by treating metrics as divergent objectives, yields three key benefits: (1) Normative Pluralism: maintaining a full suite of potentially contradictory metrics ensures that the diverse moral stances and stakeholder values inherent in RAI are adequately represented; (2) Epistemological Completeness: using multiple, sometimes conflicting, metrics captures multifaceted ethical concepts more fully and preserves greater informational fidelity than any single, simplified definition; (3) Implicit Regularization: jointly optimizing for theoretically conflicting objectives discourages overfitting to any one metric, steering models toward solutions with better generalization and robustness under real-world complexities. In contrast, enforcing theoretical consistency by simplifying or pruning metrics risks narrowing value diversity, losing conceptual depth, and degrading model performance. We therefore advocate a shift in RAI theory and practice: from getting trapped by metric inconsistencies to establishing practice-focused theories, documenting the normative provenance and inconsistency levels of inconsistent metrics, and elucidating the mechanisms that permit robust, approximated consistency in practice.

Language models must be adapted to understand and follow user instructions. Reinforcement learning is widely used to facilitate this —typically using fixed criteria such as "helpfulness" and "harmfulness". In our work, we instead propose using flexible, instruction-specific criteria as a means of broadening the impact that reinforcement learning can have in eliciting instruction following. We propose "Reinforcement Learning from Checklist Feedback" (RLCF). From instructions, we extract checklists and evaluate how well responses satisfy each item—using both AI judges and specialized verifier programs—then combine these scores to compute rewards for RL. We compare RLCF with other alignment methods on top of a strong instruction following model (Qwen2.5-7B-Instruct) on five widely-studied benchmarks — RLCF is the only method to help on every benchmark, including a 4-point boost in hard satisfaction rate on FollowBench, a 6-point increase on InFoBench, and a 3-point rise in win rate on Arena-Hard. We show that RLCF can also be used off-policy to improve Llama 3.1 8B Instruct and OLMo 2 7B Instruct. These results establish rubrics as a key tool for improving language models' support of queries that express a multitude of needs. We release our our dataset of rubrics (WildChecklists), models, and code to the public.

There is growing interest in leveraging mechanistic interpretability and controllability to better understand and influence the internal dynamics of large language models (LLMs). However, current methods face fundamental challenges in reliably localizing and manipulating feature representations. Sparse Autoencoders (SAEs) have recently emerged as a promising direction for feature extraction at scale, yet they, too, are limited by incomplete feature isolation and unreliable monosemanticity. To systematically quantify these limitations, we introduce Feature Monosemanticity Score (FMS), a novel metric to quantify feature monosemanticity in latent representation. Building on these insights, we propose Guided Sparse Autoencoders (G-SAE), a method that conditions latent representations on labeled concepts during training. We demonstrate that reliable localization and disentanglement of target concepts within the latent space improve interpretability, detection of behavior, and control. Specifically, our evaluations on toxicity detection, writing style identification, and privacy attribute recognition show that G-SAE not only enhances monosemanticity but also enables more effective and fine-grained steering with less quality degradation. Our findings provide actionable guidelines for measuring and advancing mechanistic interpretability and control of LLMs.

Large language models can generate rare but catastrophic outputs, such as harmful conversations or insecure code. Existing Reinforcement Learning from Human Feedback (RLHF) typically maximizes average reward, leaving high-risk tail events insufficiently controlled. We introduce Quantile‑Guided Alignment (QA), a framework that allows users to specify desired improvements at any quantile—individually or across multiple reward dimensions—thus shifting the distribution of outputs with finer control toward safer, more desirable outcomes. The method extends standard RLHF via an augmented reward formulation that enforces quantile constraints. Experiments on conversation and code‐generation tasks show that quantile alignment significantly enhances quality at targeted tails while maintaining overall performance. The results position QA as a principled route to risk‑calibrated language models with tail‑focused alignment.

Direct Preference Optimization (DPO) has emerged as a promising approach for aligning large language models with human preferences. While prior work mainly extends DPO from the aspect of the objective function, we instead improve DPO from the largely overlooked but critical aspect of data selection. Specifically, we address the issue of parameter shrinkage caused by noisy data by proposing a novel margin-maximization principle for dataset curation in DPO training. To further mitigate the noise in different reward models, we propose a Bayesian Aggregation approach that unifies multiple margin sources (external and implicit) into a single preference probability. Extensive experiments in diverse settings demonstrate the consistently high data efficiency of our approach. Remarkably, by using just 10\% of the Ultrafeedback dataset, our approach achieves 3\% to 8\% improvements across various Llama, Mistral, and Qwen models on the AlpacaEval2 benchmark. Furthermore, our approach seamlessly extends to iterative DPO, yielding a roughly 3\% improvement with 25\% online data, revealing the high redundancy in this presumed high-quality data construction manner. These results highlight the potential of data selection strategies for advancing preference optimization.

Many jailbreak attacks on large language models (LLMs) rely on a common objective: making the model respond with the prefix ``Sure, here is (harmful request)''. While straightforward, this objective has two limitations: limited control over model behaviors, yielding incomplete or unrealistic jailbroken responses, and a rigid format that hinders optimization. We introduce AdvPrefix, a plug-and-play prefix-forcing objective that selects one or more model-dependent prefixes by combining two criteria: high prefilling attack success rates and low negative log-likelihood. AdvPrefix integrates seamlessly into existing jailbreak attacks to mitigate the previous limitations for free. For example, replacing GCG's default prefixes on Llama-3 improves nuanced attack success rates from 14\% to 80\%, revealing that current safety alignment fails to generalize to new prefixes. Code and selected prefixes are released.

Reinforcement learning (RL) often relies on manually designed reward functions, which are difficult to specify and can lead to issues such as reward hacking and suboptimal behavior. Alternatives like inverse RL and preference-based RL attempt to infer surrogate rewards from demonstrations or preferences but suffer from ambiguity and distribution mismatch. A more direct approach, inspired by imitation learning, avoids reward modeling by leveraging expert demonstrations. However, most existing methods align actions only at individual states, failing to capture the coherence of long-horizon trajectories. In this work, we study the problem of directly aligning policies with expert-labeled trajectories to preserve long-horizon behavior without relying on reward signals. Specifically, we aim to learn a policy that maximizes the probability of generating the expert trajectories. Nevertheless, we prove that, in its general form, this trajectory alignment problem is NP-complete. To address this, we propose Trajectory Graph Learning (TGL), a framework that leverages structural assumptions commonly satisfied in practice—such as bounded realizability of expert trajectories or a tree-structured MDP. These enable a graph-based policy planning algorithm that computes optimal policies in polynomial time under known dynamics. For settings with unknown dynamics, we develop a sample-efficient algorithm based on UCB-style exploration and establish sub-linear regret. Experiments on grid-world tasks demonstrate that TGL substantially outperforms standard imitation learning methods for long-trajectory planning.

Human feedback plays a pivotal role in aligning large language models (LLMs) with human preferences. However, such feedback is often noisy or inconsistent, which can degrade the quality of reward models and hinder alignment. While various automated data cleaning methods have been proposed to mitigate this issue, a systematic evaluation of their effectiveness and generalizability remains lacking. To bridge this gap, we introduce the first comprehensive benchmark for evaluating 13 preference data cleaning methods in the context of LLM alignment. Our framework offers a standardized protocol to assess cleaning strategies in terms of alignment performance and generalizability across diverse datasets, model architectures, and optimization algorithms. By unifying disparate methods and rigorously comparing them, we uncover key factors that determine the success of data cleaning in alignment tasks. This benchmark lays the groundwork for principled and reproducible approaches to improving LLM alignment through better data quality—highlighting the crucial but underexplored role of data preprocessing in responsible AI development. We release modular implementations of all methods to catalyze further research: https://github.com/deeplearning-wisc/PrefCleanBench.

Steering has emerged as a practical approach to enable post-hoc guidance of LLMs towards enforcing a specific behavior. However, it remains largely underexplored for multimodal LLMs (MLLMs); furthermore, existing steering techniques, such as \textit{mean} steering, rely on a single steering vector, applied independently of the input query. This paradigm faces limitations when the desired behavior is dependent on the example at hand. For example, a safe answer may consist in abstaining from answering when asked for an illegal activity, or may point to external resources or consultation with an expert when asked about medical advice. In this paper, we investigate a fine-grained steering that uses an input-specific linear shift. This shift is computed using contrastive input-specific prompting. However, the input-specific prompts required for this approach are not known at test time. Therefore, we propose to train a small auxiliary module to predict the input-specific steering vector. Our approach, dubbed as L2S (Learn-to-Steer), demonstrates that it reduces hallucinations and enforces safety in MLLMs, outperforming other static baselines. We will open-source our code.