NeurIPS 2025 论文列表

Controlling specific behaviors in large language models while preserving their general capabilities is a central challenge for safe and reliable artificial intelligence deployment. Current steering methods, such as vector addition and directional ablation, are constrained within a two-dimensional subspace defined by the activation and feature direction, making them sensitive to chosen parameters and potentially affecting unrelated features due to unintended interactions in activation space. We introduce Angular Steering, a novel and flexible method for behavior modulation that operates by rotating activations within a fixed two-dimensional subspace. By formulating steering as a geometric rotation toward or away from a target behavior direction, Angular Steering provides continuous, fine-grained control over behaviors such as refusal and compliance. We demonstrate this method using refusal steering emotion steering as use cases. Additionally, we propose Adaptive Angular Steering, a selective variant that rotates only activations aligned with the target feature, further enhancing stability and coherence. Angular Steering generalizes existing addition and orthogonalization techniques under a unified geometric rotation framework, simplifying parameter selection and maintaining model stability across a broader range of adjustments. Experiments across multiple model families and sizes show that Angular Steering achieves robust behavioral control while maintaining general language modeling performance, underscoring its flexibility, generalization, and robustness compared to prior approaches. Code and artifacts are available at \url{https://github.com/lone17/angular-steering/}.

Prior work (Klochkov \& Zhivotovskiy, 2021) establishes at most $O\left(\log (n)/n\right)$ excess risk bounds via algorithmic stability for strongly-convex learners with high probability. We show that under the similar common assumptions — Polyak-Lojasiewicz condition, smoothness, and Lipschitz continous for losses — rates of $O\left(\log^2(n)/n^2\right)$ are at most achievable. To our knowledge, our analysis also provides the tightest high-probability bounds for gradient-based generalization gaps in nonconvex settings.

Task-based functional magnetic resonance imaging (fMRI) provides invaluable insights into human cognition but faces critical hurdles—low signal-to-noise ratio, high dimensionality, limited sample sizes, and costly data acquisition—that are amplified when integrating datasets across subjects or sites. This paper introduces orthogonal contrastive learning (OCL), a unified multi-representation framework for multi-subject fMRI analysis that aligns neural responses without requiring temporal preprocessing or uniform time-series lengths across subjects or sites. OCL employs two identical encoders: an online network trained with a contrastive loss that pulls together same-stimulus responses and pushes apart different-stimulus responses, and a target network whose weights track the online network via exponential moving average to stabilize learning. Each OCL network layer combines QR decomposition for orthogonal feature extraction, locality-sensitive hashing (LSH) to produce compact subject-specific signatures, positional encoding to embed temporal structure alongside spatial features, and a transformer encoder to generate discriminative, stimulus-aligned embeddings. We further enhance OCL with an unsupervised pretraining stage on fMRI-like synthetic data and demonstrate a transfer-learning workflow for multi-site studies. Across extensive experiments on multi-subject and multi-site fMRI benchmarks, OCL consistently outperforms state-of-the-art alignment and analysis methods in both representation quality and downstream classification accuracy.

In this position paper we argue that conclusions drawn about relative system safety or attack method efficacy via AI red teaming are often not supported by evidence provided by attack success rate (ASR) comparisons. We show, through conceptual, theoretical, and empirical contributions, that many conclusions are founded on apples-to-oranges comparisons or low-validity measurements. Our arguments are grounded in asking a simple question: When can attack success rates be meaningfully compared? To answer this question, we draw on ideas from social science measurement theory and inferential statistics, which, taken together, provide a conceptual grounding for understanding when numerical values obtained through the quantification of system attributes can be meaningfully compared. Through this lens, we articulate conditions under which ASRs can and cannot be meaningfully compared. Using jailbreaking as a running example, we provide examples and extensive discussion of apples-to-oranges ASR comparisons and measurement validity challenges.

Direct Preference Optimization (DPO) is a key framework for aligning text-to-image models with human preferences, extended by Stepwise Preference Optimization (SPO) to leverage intermediate steps for preference learning, generating more aesthetically pleasing images with significantly less computational cost. While effective, SPO's underlying mechanisms remain underexplored. In light of this, we critically re-examine SPO by formalizing its mechanism as gradient guidance. This new lens shows that SPO uses biased temporal weighting, giving too little weight to later generative steps, and unlike likelihood centric views it reveals substantial noise in the gradient estimates. Leveraging these insights, our GradSPO algorithm introduces a simplified loss and a targeted, variance-informed noise reduction strategy, enhancing training stability. Evaluations on SD 1.5 and SDXL show GradSPO substantially outperforms leading baselines in human preference, yielding images with markedly improved aesthetics and semantic faithfulness, leading to more robust alignment. Code and models are available at https://github.com/JoshuaTTJ/GradSPO.

The widespread adoption of large language models (LLMs) across industries has increased the demand for high-quality and customizable outputs. However, traditional alignment methods often require retraining large pretrained models, making it difficult to quickly adapt and optimize LLMs for diverse applications. To address this limitation, we propose a novel \textit{Residual Alignment Model} (\textit{RAM}) that formalizes the alignment process as a type of importance sampling. In this framework, the unaligned upstream model serves as the proposal distribution, while the alignment process is framed as secondary sampling based on an autoregressive alignment module that acts as an estimator of the importance weights. This design enables a natural detachment of the alignment module from the target aligned model, improving flexibility and scalability. Based on this model, we derive an efficient sequence-level training strategy for the alignment module, which operates independently of the proposal module. Additionally, we develop a resampling algorithm with iterative token-level decoding to address the common first-token latency issue in comparable methods. Experimental evaluations on two leading open-source LLMs across diverse tasks, including instruction following, domain adaptation, and preference optimization, demonstrate that our approach consistently outperforms baseline models.

Large Vision-Language Models (LVLMs) pretrained on large-scale multimodal data have shown promising capabilities in Video Anomaly Detection (VAD). However, their ability to reason about abnormal events based on scene semantics remains underexplored. In this paper, we investigate LVLMs’ behavior in VAD from a visual-textual co-occurrence perspective, focusing on whether their decisions are driven by statistical shortcuts between visual instances and textual phrases. By analyzing visual-textual co-occurrence in pretraining data and conducting experiments under different data settings, we reveal a hallucination phenomenon: LVLMs tend to rely on co-occurrence patterns between visual instances and textual phrases associated with either normality or abnormality, leading to incorrect predictions when these high-frequency objects appear in semantically mismatched contexts. To address this issue, we propose VAD-DPO, a direct preference optimization method supervised with counter-example pairs. By constructing visually similar but semantically contrasting video clips, VAD-DPO encourages the model to align its predictions with the semantics of scene rather than relying on co-occurrence patterns. Extensive experiments on six benchmark datasets demonstrate the effectiveness of VAD-DPO in enhancing both anomaly detection and reasoning performance, particularly in scene-dependent scenarios.

We introduce and study the Combinatorial Ski Rental (CSR) problem, which involves multiple items that can be rented or purchased, either individually or in combination. At each time step, a decision-maker must make an irrevocable buy-or-rent decision for items that have not yet been purchased, without knowing the end of the time horizon. We propose a randomized online algorithm, Sorted Optimal Amortized Cost (SOAC), that achieves the optimal competitive ratio. Moreover, SOAC can be extended to address various well-known ski rental variants, including the multi-slope, multi-shop, multi-commodity ski rental and CSR with upgrading problems. Building on the proposed SOAC algorithm, we further develop a learning-augmented algorithm that leverages machine-learned predictions to improve the performance of CSR. This algorithm is capable of recovering or improving upon existing results of learning-augmented algorithms in both the classic ski rental and multi-shop ski rental problems. Experimental results validate our theoretical analysis and demonstrate the advantages of our algorithms over baseline methods for ski rental problems.

The remarkable progress in text-to-video diffusion models enables the generation of photorealistic videos, although the content of these generated videos often includes unnatural movement or deformation, reverse playback, and motionless scenes. Recently, an alignment problem has attracted huge attention, where we steer the output of diffusion models based on some measure of the content's goodness. Because there is a large room for improvement of perceptual quality along the frame direction, we should address which metrics we should optimize and how we can optimize them in the video generation. In this paper, we propose diffusion latent beam search with lookahead estimator, which can select a better diffusion latent to maximize a given alignment reward at inference time. We then point out that improving perceptual video quality with respect to alignment to prompts requires reward calibration by weighting existing metrics. This is because when humans or vision language models evaluate outputs, many previous metrics to quantify the naturalness of video do not always correlate with the evaluation. We demonstrate that our method improves the perceptual quality evaluated on the calibrated reward, VLMs, and human assessment, without model parameter update, and outputs the best generation compared to greedy search and best-of-N sampling under much more efficient computational cost. The experiments highlight that our method is beneficial to many capable generative models, and provide a practical guideline: we should prioritize the inference-time compute allocation into enabling the lookahead estimator and increasing the search budget, rather than expanding the denoising steps.

With the advancement of large language models (LLMs), solving complex tasks (e.g., math problems, code generation, etc.) has garnered increasing attention. Inference-time computation methods (e.g., Best-of-N, MCTS, etc.) are of significant importance, as they have the potential to enhance the reasoning capabilities of LLMs without requiring external training computation. However, due to the inherent challenges of this technique, most existing methods remain proof-of-concept and are not yet sufficiently effective. In this paper, we investigate and benchmark strategies for improving inference-time computation across a wide range of reasoning tasks. Since most current methods rely on a pipeline that first generates candidate solutions (e.g., generating chain-of-thought candidate solutions) and then selects them based on specific reward signals (e.g., RLHF reward, process reward, etc.), our research focuses on strategies for both candidate solution generation (e.g., instructing prompts, hyperparameters: temperature and top-p, etc.) and reward mechanisms (e.g., self-evaluation, reward types, etc.). The experimental results reveal that several previously overlooked strategies can be critical for the success of inference-time computation (e.g., simplifying the temperature can improve general reasoning task performance by up to 5%). Based on extensive experiments (more than 20,000 A100-80G GPU hours with over 1,000 experiments) across a variety of models (e.g., Llama, Qwen, and Mistral families) of various sizes, our proposed strategies outperform the baseline by a substantial margin in most cases, providing a stronger foundation for future research.

We introduce the Deep Value Benchmark (DVB), an evaluation framework that directly tests whether large language models (LLMs) learn fundamental human values or merely surface-level preferences. This distinction is critical for AI alignment: Systems that capture deeper values are likely to generalize human intentions robustly, while those that capture only superficial patterns in preference data risk producing misaligned behavior. The DVB uses a novel experimental design with controlled confounding between deep values (e.g., moral principles) and shallow features (e.g., superficial attributes). In the training phase, we expose LLMs to human preference data with deliberately correlated deep and shallow features---for instance, where a user consistently prefers (non-maleficence, formal language) options over (justice, informal language) alternatives. The testing phase then breaks these correlations, presenting choices between (justice, formal language) and (non-maleficence, informal language) options. This design allows us to precisely measure a model's Deep Value Generalization Rate (DVGR)---the probability of generalizing based on the underlying value rather than the shallow feature. Across 9 different models, the average DVGR is just 0.30. All models generalize deep values less than chance. Larger models have a (slightly) lower DVGR than smaller models. We are releasing our dataset, which was subject to three separate human validation experiments. DVB provides an interpretable measure of a core feature of alignment.

We investigate refinements of the mean-payoff criterion in two-player zero-sum perfect-information stochastic games. A strategy is *Blackwell optimal* if it is optimal in the discounted game for all discount factors sufficiently close to $1$. The notion of *$d$-sensitive optimality* interpolates between mean-payoff optimality (corresponding to the case $d=-1$) and Blackwell optimality ($d=\infty$). The *Blackwell threshold* $\alpha_{\sf bw} \in [0,1[$ is the discount factor above which all optimal strategies in the discounted game are guaranteed to be Blackwell optimal. The *$d$-sensitive threshold* $\alpha_{\sf d} \in [0,1[$ is defined analogously. Bounding $\alpha_{\sf bw}$ and $\alpha_{\sf d}$ are fundamental problems in algorithmic game theory, since these thresholds control the complexity for computing Blackwell and $d$-sensitive optimal strategies, by reduction to discounted games which can be solved in $O\left((1-\alpha)^{-1}\right)$ iterations. We provide the first bounds on the $d$-sensitive threshold $\alpha_{\sf d}$ beyond the case $d=-1$, and we establish improved bounds for the Blackwell threshold $\alpha_{\sf bw}$. This is achieved by leveraging separation bounds on algebraic numbers, relying on Lagrange bounds and more advanced techniques based on Mahler measures and multiplicity theorems.

Aerial-Ground person re-identification (AG-ReID) is an emerging yet challenging task that aims to match pedestrian images captured from drastically different viewpoints, typically from unmanned aerial vehicles (UAVs) and ground-based surveillance cameras. The task poses significant challenges due to extreme viewpoint discrepancies, occlusions, and domain gaps between aerial and ground imagery. While prior works have made progress by learning cross-view representations, they remain limited in handling severe pose variations and spatial misalignment. To address these issues, we propose a Geometric and Semantic Alignment Network (GSAlign) tailored for AG-ReID. GSAlign introduces two key components to jointly tackle geometric distortion and semantic misalignment in aerial-ground matching: a Learnable Thin Plate Spline (LTPS) Transformation Module and a Dynamic Alignment Module (DAM). The LTPS module adaptively warps pedestrian features based on a set of learned keypoints, effectively compensating for geometric variations caused by extreme viewpoint changes. In parallel, the DAM estimates visibility-aware representation masks that highlight visible body regions at the semantic level, thereby alleviating the negative impact of occlusions and partial observations in cross-view correspondence. Extensive experiments on the challenging CARGO benchmark demonstrate the effectiveness of GSAlign, achieving significant improvements of +18.8\% in mAP and +16.8\% in Rank-1 accuracy over previous state-of-the-art methods.

In this paper, we study the fundamental problems of maintaining the diameter and a $k$-center clustering of a dynamic point set $P \subset \mathbb{R}^d$, where points may be inserted or deleted over time and the ambient dimension $d$ is not constant and may be high. Our focus is on designing algorithms that remain effective even in the presence of an \emph{adaptive adversary}—an adversary that, at any time $t$, knows the entire history of the algorithm’s outputs as well as all the random bits used by the algorithm up to that point. We present a fully dynamic algorithm that maintains a $2$-approximate diameter with a \emph{worst-case} update time of $poly(d, \log n)$, where $n$ is the length of the stream. Our result is achieved by identifying a robust representative of the dataset that requires infrequent updates, combined with a careful deamortization. To the best of our knowledge, this is the first efficient fully-dynamic algorithm for diameter in high dimensions that \emph{simultaneously} achieves a $2$-approximation guarantee and robustness against an adaptive adversary. We also give an improved dynamic $(4+\epsilon)$-approximation algorithm for the $k$-center problem, also resilient to an adaptive adversary. Our clustering algorithm achieves an amortized update time of $k^{2.5} d \cdot poly(\epsilon^{-1}, \log n)$, improving upon the amortized update time of $k^6 d \cdot poly( \epsilon^{-1}, \log n)$ by Biabani et al. [NeurIPS'24].

We study a two-institution stable matching model in which candidates from two distinct groups are evaluated using partially correlated signals that are group-biased. This extends prior work (which assumes institutions evaluate candidates in an identical manner) to a more realistic setting in which institutions rely on overlapping, but independently processed, criteria. These evaluations could consist of a variety of informative tools such as standardized tests, shared recommendation systems, or AI-based assessments with local noise. Two key parameters govern evaluations: the bias parameter $\beta \in (0,1]$, which models systematic disadvantage faced by one group, and the correlation parameter $\gamma \in [0,1]$, which captures the alignment between institutional rankings. We study the representation ratio $\mathcal{R}(\beta, \gamma)$, i.e., the ratio of disadvantaged to advantaged candidates selected by the matching process in this setting. Focusing on a regime in which all candidates prefer the same institution, we characterize the large-market equilibrium and derive a closed-form expression for the resulting representation ratio. Prior work shows that when $\gamma = 1$, this ratio scales linearly with $\beta$. In contrast, we show that $\mathcal{R}(\beta, \gamma)$ increases nonlinearly with $\gamma$ and even modest losses in correlation can cause sharp drops in the representation ratio. Our analysis identifies critical $\gamma$-thresholds where institutional selection behavior undergoes discrete transitions, and reveals structural conditions under which evaluator alignment or bias mitigation are most effective. Finally, we show how this framework and results enable interventions for fairness-aware design in decentralized selection systems.

Large Language Model (LLM) inference is typically memory-intensive, especially when processing large batch sizes and long sequences, due to the large size of key-value (KV) cache. Vector Quantization (VQ) is recently adopted to alleviate this issue, but we find that the existing approach is susceptible to distribution shift due to its reliance on calibration datasets. To address this limitation, we introduce $\textbf{NSNQuant}$, a calibration-free Vector Quantization (VQ) technique designed for low-bit compression of the KV cache. By applying a three-step transformation—$\textbf{1)}$ a token-wise normalization ($\textbf{N}$ormalize), $\textbf{2)}$ a channel-wise centering ($\textbf{S}$hift), and $\textbf{3)}$ a second token-wise normalization ($\textbf{N}$ormalize)—with Hadamard transform, NSNQuant effectively aligns the token distribution with the standard normal distribution. This alignment enables robust, calibration-free vector quantization using a single reusable codebook. Extensive experiments show that NSNQuant consistently outperforms prior methods in both 1-bit and 2-bit settings, offering strong generalization and up to 3$\times$ throughput gains over full-precision baselines.

After pre-training, large language models are aligned with human preferences based on pairwise comparisons. State-of-the-art alignment methods (such as PPO-based RLHF and DPO) are built on the assumption of aligning with a single preference model, despite being deployed in settings where users have diverse preferences. As a result, it is not even clear that these alignment methods produce models that satisfy users \emph{on average} --- a minimal requirement for pluralistic alignment. Drawing on social choice theory and modeling users' comparisons through individual Bradley-Terry (BT) models, we introduce an alignment method's \emph{distortion}: the worst-case ratio between the optimal achievable average utility, and the average utility of the learned policy. The notion of distortion helps draw sharp distinctions between alignment methods: \emph{Nash Learning from Human Feedback} achieves the minimax optimal distortion of $(\frac{1}{2} + o(1)) \cdot \beta$ (for the BT temperature $\beta$), robustly across utility distributions, distributions of comparison pairs, and permissible KL divergences from the reference policy. RLHF and DPO, by contrast, suffer $\geq (1 - o(1)) \cdot \beta$ distortion already without a KL constraint, and $e^{\Omega(\beta)}$ or even unbounded distortion in the full setting, depending on how comparison pairs are sampled.

Transformer models have driven breakthroughs across various language tasks by their strong capability to learn rich contextual representations. Scaling them to improve representation, however, often demands substantial memory and compute costs, such as the Key-Value (KV) cache used during auto-regressive decoding. Skip connections offer a promising way to improve representation without bloating resource usage, yet most prior works either improve expressivity while leaving KV costs unchanged, or reduce memory at the cost of weaker representation. In this work, we propose SkipV1Former, a Transformer variant that uses skip connections from the first layer's Value heads to strengthen model representation and reduce KV cache. Specifically, from the second block onward, each layer reuses half of its Value heads from the very first layer, while computing the other half as usual-cutting Value projections and V cache by nearly 50 \%. Theoretically, we show that routing uncompressed first-layer Values into deeper layers restores information lost to compression and accelerates the model’s implicit mesa-optimization-a key pattern of Transformer in auto-regressive tasks. Empirically, across different model scales, SkipV1Former delivers consistent reductions of approximately 25 \% in KV cache while improving perplexity relative to standard Multi-Head Attention (MHA) Transformers and some advanced variants. Moreover, we propose a recipe for uptraining existing MHA Transformer checkpoints to SkipV1Former with only 10-15\% additional compute. Finally, SkipV1Former can seamlessly combine advanced methods like Group-Query Attention and Multi-Latent Attention to achieve further KV cache savings and performance improvement. When combined with YOCO, it cuts KV cache size by nearly 50 \% while still improving performance. The code is available at: https://github.com/Zhoutong-Wu/SkipV1Former.

Reward models are central to aligning LLMs with human preferences, but they are costly to train, requiring large-scale human-labeled preference data and powerful pretrained LLM backbones. Meanwhile, the increasing availability of high-quality synthetic instruction-following datasets raises the question: can simpler, reference-based metrics serve as viable alternatives to reward models during RL-based alignment? In this paper, we show first that BLEU, a basic string-matching metric, surprisingly matches strong reward models in agreement with human preferences on general instruction-following datasets. Based on this insight, we develop BLEUBERI, a method that first identifies challenging instructions and then applies Group Relative Policy Optimization (GRPO) using BLEU directly as the reward function. We demonstrate that BLEUBERI-trained models are competitive with models trained via reward model-guided RL across four challenging instruction-following benchmarks and three different base language models. A human evaluation further supports that the quality of BLEUBERI model outputs is on par with those from reward model-aligned models. Moreover, BLEUBERI models generate outputs that are more factually grounded than competing methods. Overall, we show that given access to high-quality reference outputs (easily obtained via existing instruction-following datasets or synthetic data generation), string matching-based metrics are cheap yet effective proxies for reward models during alignment. We release our code and data at https://github.com/lilakk/BLEUBERI.