NeurIPS 2025 Papers

Finetuning large language models (LLMs) enables user-specific customization but introduces critical safety risks: even a few harmful examples can compromise safety alignment. A common mitigation strategy is to update the model more strongly on examples deemed safe, while downweighting or excluding those flagged as unsafe. However, because safety context can shift within a single example, updating the model equally on both harmful and harmless parts of a response is suboptimal — an atomic treatment we term static safety shaping. In contrast, we propose dynamic safety shaping (DSS), a dynamic shaping framework that uses fine-grained safety signals to reinforce learning from safe segments of a response while suppressing unsafe content. To enable such fine-grained control during finetuning, we introduce a key insight: guardrail models, traditionally used for filtering, can be repurposed to evaluate partial responses, tracking how safety risk evolves throughout the response, segment by segment. This leads to the Safety Trajectory Assessment of Response (STAR), a token-level signal that enables shaping to operate dynamically over the training sequence. Building on this, we present ★DSS, a DSS method guided by STAR scores that robustly mitigates finetuning risks and delivers substantial safety improvements across diverse threats, datasets, and model families, all without compromising capability on intended tasks. We encourage future safety research to build on dynamic shaping principles for stronger mitigation against evolving finetuning risks. Our code is publicly available at https://github.com/poloclub/star-dss

Generating an interactive visual world from a single image is both challenging and practically valuable, as single-view inputs are easy to acquire and align well with prompt-driven applications such as gaming and virtual reality. This paper introduces a novel unified framework, Image as a World (**IaaW**), which synthesizes high-quality 360-degree videos from a single image that are both controllable and temporally continuable. Our framework consists of three stages: world initialization, which jointly synthesizes spatially complete and temporally dynamic scenes from a single view; world exploration, which supports user-specified viewpoint rotation; and world continuation, which extends the generated scene forward in time with temporal consistency. To support this pipeline, we design a visual world model based on generative diffusion models modulated with spherical 3D positional encoding and multi-view composition to represent geometry and view semantics. Additionally, a vision-language model (IaaW-VLM) is fine-tuned to produce both global and view-specific prompts, improving semantic alignment and controllability. Extensive experiments demonstrate that our method produces panoramic videos with superior visual quality, minimal distortion and seamless continuation in both qualitative and quantitative evaluations. To the best of our knowledge, this is the first work to generate a controllable, consistent, and temporally expandable 360-degree world from a single image.

Model editing enables targeted updates to the knowledge of large language models (LLMs) with minimal retraining. Among existing approaches, locate-then-edit methods constitute a prominent paradigm: they first identify critical layers, then compute residuals at the final critical layer based on the target edit, and finally apply least-squares-based multi-layer updates via $\textbf{residual distribution}$. While empirically effective, we identify a counterintuitive failure mode: residual distribution, a core mechanism in these methods, introduces weight shift errors that undermine editing precision. Through theoretical and empirical analysis, we show that such errors increase with the distribution distance, batch size, and edit sequence length, ultimately leading to inaccurate or suboptimal edits. To address this, we propose the $\textbf{B}$oundary $\textbf{L}$ayer $\textbf{U}$pdat$\textbf{E (BLUE)}$ strategy to enhance locate-then-edit methods. Sequential batch editing experiments on three LLMs and two datasets demonstrate that BLUE not only delivers an average performance improvement of 35.59\%, significantly advancing the state of the art in model editing, but also enhances the preservation of LLMs' general capabilities. Our code is available at https://github.com/xpq-tech/BLUE.

Knowledge Distillation (KD) aims to train a lightweight student model by transferring knowledge from a large, high-capacity teacher. Recent studies have shown that leveraging diverse teacher perspectives can significantly improve distillation performance; however, achieving such diversity typically requires multiple teacher networks, leading to high computational costs. In this work, we propose a novel cost-efficient knowledge augmentation method for KD that generates diverse multi-views by attaching multiple branches to a single teacher. To ensure meaningful semantic variation across multi-views, we introduce two angular diversity objectives: 1) $\textit{constrained inter-angle diversify loss}$, which maximizes angles between augmented views while preserving proximity to the original teacher output, and 2) $\textit{intra-angle diversify loss}$, which encourages an even distribution of views around the original output. The ensembled knowledge from these angularly diverse views, along with the original teacher, is distilled into the student. We further theoretically demonstrate that our objectives increase the diversity among ensemble members and thereby reduce the upper bound of the ensemble's expected loss, leading to more effective distillation. Experimental results show that our method surpasses an existing knowledge augmentation method across diverse configurations. Moreover, the proposed method is compatible with other KD frameworks in a plug-and-play fashion, providing consistent improvements in generalization performance.

While diffusion models have achieved remarkable success in text-to-image generation, they encounter significant challenges with instruction-driven image editing. Our research highlights a key challenge: these models particularly struggle with structurally-inconsistent edits that involve substantial layout changes. To address this gap, we introduce Image Editing As Programs (IEAP), a unified image editing framework built upon the Diffusion Transformer (DiT) architecture. Specifically, IEAP deals with complex instructions by decomposing them into a sequence of programmable atomic operations. Each atomic operation manages a specific type of structurally consistent edit; when sequentially combined, IEAP enables the execution of arbitrary and structurally-inconsistent transformations. This reductionist approach enables IEAP to robustly handle a wide spectrum of edits, encompassing both structurally-consistent and inconsistent changes. Extensive experiments demonstrate that IEAP significantly outperforms state-of-the-art methods on standard benchmarks across various editing scenarios. In these evaluations, our framework delivers superior accuracy and semantic fidelity, particularly for complex, multi-step instructions. Codes are available at https://github.com/YujiaHu1109/IEAP.

Diffusion models have recently emerged as powerful neural solvers for combinatorial optimization (CO). However, existing approaches fail to reveal how variables are progressively determined during inference, making the final solution opaque until the last step. To address this limitation, we propose a structured denoising diffusion model, StruDiCO, which incrementally constructs solutions through step-wise variable selection. This is achieved via a variable-absorption noising model, wherein the forward process simulates gradual variable deactivation, converging to an empty solution, while the reverse process incrementally selects variables to reconstruct the final solution. This design induces structural continuity across intermediate states, enabling interpretable and trajectory-consistent partial solutions throughout inference. To further improve the reliability of reverse inference, we introduce a constrained consistency sampling strategy, which suppresses low-confidence variable selection at each step to stabilize the reverse process. Leveraging the structure-preserving reverse process, we further propose a lightweight, gradient-free, objective-aware refinement framework, which iteratively improves solution quality by applying structure-aware perturbations to the current solution, performing reverse inference through the constraint consistency model, and decoding with an objective-guided scoring scheme. Extensive experiments on two canonical CO tasks, the Traveling Salesman Problem (TSP) and Maximal Independent Set (MIS), show that StruDiCO outperforms state-of-the-art diffusion-based solvers, achieving up to $3.5\times$ faster inference, 70\% lower GPU memory usage, and significantly improved solution quality, with up to 37.7\% drop reduction on TSP and an average 38.1\% improvement on MIS. The codes are publicly available at https://github.com/yuuuuwang/StruDiCO.

Recurrent neural networks (RNNs) trained on neuroscience-inspired tasks offer powerful models of brain computation. However, typical training paradigms rely on open-loop, supervised settings, whereas real-world learning unfolds in closed-loop environments. Here, we develop a mathematical theory describing the learning dynamics of linear RNNs trained in closed-loop contexts. We first demonstrate that two otherwise identical RNNs, trained in either closed- or open-loop modes, follow markedly different learning trajectories. To probe this divergence, we analytically characterize the closed-loop case, revealing distinct stages aligned with the evolution of the training loss. Specifically, we show that the learning dynamics of closed-loop RNNs, in contrast to open-loop ones, are governed by an interplay between two competing objectives: short-term policy improvement and long-term stability of the agent-environment interaction. Finally, we apply our framework to a realistic motor control task, highlighting its broader applicability. Taken together, our results underscore the importance of modeling closed-loop dynamics in a biologically plausible setting.

Numerous models have been developed for scanpath and saliency prediction, which are typically trained on scanpaths, which model eye movement as a sequence of discrete fixation points connected by saccades, while the rich information contained in the raw trajectories is often discarded. Moreover, most existing approaches fail to capture the variability observed among human subjects viewing the same image. They generally predict a single scanpath of fixed, pre-defined length, which conflicts with the inherent diversity and stochastic nature of real-world visual attention. To address these challenges, we propose DiffEye, a diffusion-based training framework designed to model continuous and diverse eye movement trajectories during free viewing of natural images. Our method builds on a diffusion model conditioned on visual stimuli and introduces a novel component, namely Corresponding Positional Embedding (CPE), which aligns spatial gaze information with the patch-based semantic features of the visual input. By leveraging raw eye-tracking trajectories rather than relying on scanpaths, DiffEye captures the inherent variability in human gaze behavior and generates high-quality, realistic eye movement patterns, despite being trained on a comparatively small dataset. The generated trajectories can also be converted into scanpaths and saliency maps, resulting in outputs that more accurately reflect the distribution of human visual attention. DiffEye is the first method to tackle this task on natural images using a diffusion model while fully leveraging the richness of raw eye-tracking data. Our extensive evaluation shows that DiffEye not only achieves state-of-the-art performance in scanpath generation but also enables, for the first time, the generation of continuous eye movement trajectories. Project webpage: https://diff-eye.github.io/

Self-play fine-tuning has demonstrated promising abilities in adapting large language models (LLMs) to downstream tasks with limited real-world data. The basic principle is to iteratively refine the model with real samples and synthetic ones generated from itself. However, the existing methods primarily focus on the relative gaps between the rewards for two types of data, neglecting their absolute values. Through theoretical analysis, we identify that the gap-based methods suffer from unstable evolution, due to the potentially degenerated objectives. To address this limitation, we introduce a novel self-play fine-tuning method, namely \underline{S}elf-\underline{P}l\underline{A}y via Noise \underline{C}ontrastive \underline{E}stimation (SPACE), which leverages noise contrastive estimation to capture the real-world data distribution. Specifically, SPACE treats synthetic samples as auxiliary components, and discriminates them from the real ones in a binary classification manner. As a result, SPACE independently optimizes the absolute reward values for each type of data, ensuring a consistently meaningful objective and thereby avoiding the instability issue. Theoretically, we show that the optimal solution of the objective in SPACE aligns with the underlying distribution of real-world data, and SPACE guarantees a provably stable convergence to the optimal distribution. Empirically, we show that SPACE significantly improves the performance of LLMs over various tasks, and outperforms supervised fine-tuning that employs much more real-world samples. Compared to gap-based self-play fine-tuning methods, SPACE exhibits remarkable superiority and stable evolution.

Alzheimer's Disease (AD) poses a significant health threat to the aging population, underscoring the critical need for early diagnosis to delay disease progression and improve patient quality of life. Recent advances in heterogeneous multimodal artificial intelligence (AI) have facilitated comprehensive joint diagnosis, yet practical clinical scenarios frequently encounter incomplete modalities due to factors like high acquisition costs or radiation risks. Moreover, traditional convolution-based architecture face inherent limitations in capturing long-range dependencies and handling heterogeneous medical data efficiently. To address these challenges, in our proposed heterogeneous multimodal diagnostic framework (HAD), we develop a multi-view Hilbert curve-based Mamba block and a hierarchical spatial feature extraction module to simultaneously capture local spatial features and global dependencies, effectively alleviating spatial discontinuities introduced by voxel serialization. Furthermore, to balance semantic consistency and modal specificity, we build a unified mutual information learning objective in the heterogeneous multimodal embedding space, which maintains effective learning of modality-specific information to avoid modality collapse caused by model preference. Extensive experiments demonstrate that our HAD significantly outperforms state-of-the-art methods in various modality-missing scenarios, providing an efficient and reliable solution for early-stage AD diagnosis.

Text-to-Speech (TTS) benchmarks often fail to capture how well models handle nuanced and semantically complex text. Building on $\textit{EmergentTTS}$, we introduce $\textit{EmergentTTS-Eval}$, a comprehensive benchmark covering six challenging TTS scenarios: emotions, paralinguistics, foreign words, syntactic complexity, complex pronunciation (e.g. URLs, formulas), and questions. Crucially, our framework automates both test-case generation and evaluation, making the benchmark easily extensible. Starting from a small set of human-written seed prompts, we iteratively extend them using LLMs to target specific structural, phonetic and prosodic challenges, resulting in 1,645 diverse test samples. Moreover, we employ a model-as-a-judge approach, using a Large Audio Language Model (LALM) to assess the speech across multiple dimensions such as expressed emotion, prosodic, intonational, and pronunciation accuracy. We evaluate state-of-the-art open-source and proprietary TTS systems, such as 11Labs, Deepgram, and OpenAI's 4o-mini-TTS, on EmergentTTS-Eval, demonstrating its ability to reveal fine-grained performance differences. Results show that the model-as-a-judge approach offers robust TTS assessment and a high correlation with human preferences. We open-source the code and the dataset.

Human-centered images often suffer from severe generic degradation during transmission and are prone to human motion blur (HMB), making restoration challenging. Existing research lacks sufficient focus on these issues, as both problems often coexist in practice. To address this, we design a degradation pipeline that simulates the coexistence of HMB and generic noise, generating synthetic degraded data to train our proposed HAODiff, a human-aware one-step diffusion. Specifically, we propose a triple-branch dual-prompt guidance (DPG), which leverages high-quality images, residual noise (LQ minus HQ), and HMB segmentation masks as training targets. It produces a positive–negative prompt pair for classifier‑free guidance (CFG) in a single diffusion step. The resulting adaptive dual prompts let HAODiff exploit CFG more effectively, boosting robustness against diverse degradations. For fair evaluation, we introduce MPII‑Test, a benchmark rich in combined noise and HMB cases. Extensive experiments show that our HAODiff surpasses existing state-of-the-art (SOTA) methods in terms of both quantitative metrics and visual quality on synthetic and real-world datasets, including our introduced MPII-Test. Code is available at: https://github.com/gobunu/HAODiff.

Approaches for goal-conditioned reinforcement learning (GCRL) often use learned state representations to extract goal-reaching policies. Two frameworks for representation structure have yielded particularly effective GCRL algorithms: (1) *contrastive representations*, in which methods learn "successor features" with a contrastive objective that performs inference over future outcomes, and (2) *temporal distances*, which link the (quasimetric) distance in representation space to the transit time from states to goals. We propose an approach that unifies these two frameworks, using the structure of a quasimetric representation space (triangle inequality) with the right additional constraints to learn successor representations that enable optimal goal-reaching. Unlike past work, our approach is able to exploit a **quasimetric** distance parameterization to learn **optimal** goal-reaching distances, even with **suboptimal** data and in **stochastic** environments. This gives us the best of both worlds: we retain the stability and long-horizon capabilities of Monte Carlo contrastive RL methods, while getting the free stitching capabilities of quasimetric network parameterizations. On existing offline GCRL benchmarks, our representation learning objective improves performance on stitching tasks where methods based on contrastive learning struggle, and on noisy, high-dimensional environments where methods based on quasimetric networks struggle.

LiDAR-based world models offer more structured and geometry-aware representations than their image-based counterparts. However, existing LiDAR world models are narrowly trained; each model excels only in the domain for which it was built. This raises a critical question: can we develop LiDAR world models that exhibit strong transferability across multiple domains? To answer this, we conduct the first systematic domain transfer study across three demanding scenarios: (i) outdoor to indoor generalization, (ii) sparse- to dense-beam adaptation, and (iii) non-semantic to semantic transfer. Given different amounts of fine-tuning data, our experiments show that a single pretrained model can achieve up to 11\% absolute improvement (83\% relative) over training from scratch and outperforms training from scratch in 30/36 of our comparisons. This transferability significantly reduces the reliance on manually annotated data for semantic occupancy forecasting: our method exceeds previous baselines with only 5\% of the labeled training data of prior work. We also observed inefficiencies of current generative-model-based LiDAR world models, mainly through their under-compression of LiDAR data and inefficient training objectives. To address these issues, we propose a latent conditional flow matching (CFM)-based framework that achieves state-of-the-art reconstruction accuracy using only half the training data and a compression ratio 6 times higher than that of prior methods. Our model also achieves SOTA performance on semantic occupancy forecasting while being 1.98x-23x more computationally efficient (a 1.1x-3.9x FPS speedup) than previous methods.

Automated Theorem Proving (ATP) in formal languages remains a formidable challenge in AI, demanding rigorous logical deduction and navigating vast search spaces. While large language models (LLMs) have shown promising performance, existing stepwise provers often suffer from biased search guidance, leading to inefficiencies and suboptimal proof strategies. This paper introduces the Multi-Perspective Search Prover (MPS-Prover), a novel stepwise ATP system designed to overcome these limitations. MPS-Prover incorporates two key innovations: a highly effective post-training data curation strategy that prunes approximately 40\% of redundant training data without sacrificing performance, and a multi-perspective tree search mechanism. This search integrates a learned critic model with strategically designed heuristic rules to diversify tactic selection, prevent getting trapped in unproductive states, and enhance search robustness. Extensive evaluations demonstrate that MPS-Prover achieves state-of-the-art performance on multiple challenging benchmarks, including miniF2F and ProofNet, outperforming prior 7B parameter models. Furthermore, our analyses reveal that MPS-Prover generates significantly shorter and more diverse proofs compared to existing stepwise and whole-proof methods, highlighting its efficiency and efficacy. Our work advances the capabilities of LLM-based formal reasoning and offers a robust framework and a comprehensive analysis for developing more powerful theorem provers.

Signal Temporal Logic (STL) is a powerful specification language for describing complex temporal behaviors of continuous signals, making it well-suited for high-level robotic task descriptions. However, generating executable plans for STL tasks is challenging, as it requires consideration of the coupling between the task specification and the system dynamics. Existing approaches either follow a model-based setting that explicitly requires knowledge of the system dynamics or adopt a task-oriented data-driven approach to learn plans for specific tasks. In this work, we address the problem of generating executable STL plans for systems with unknown dynamics. We propose a hierarchical planning framework that enables zero-shot generalization to new STL tasks by leveraging only task-agnostic trajectory data during offline training. The framework consists of three key components: (i) decomposing the STL specification into several progresses and time constraints, (ii) searching for timed waypoints that satisfy all progresses under time constraints, and (iii) generating trajectory segments using a pre-trained diffusion model and stitching them into complete trajectories. We formally prove that our method guarantees STL satisfaction, and simulation results demonstrate its effectiveness in generating dynamically feasible trajectories across diverse long-horizon STL tasks. Project Page: https://cps-sjtu.github.io/Zero-Shot-STL/

Model editing has become a *de-facto* practice to address hallucinations and outdated knowledge of large language models (LLMs). However, existing methods are predominantly evaluated in isolation, i.e., one edit at a time, failing to consider a critical scenario of compositional model editing, where multiple edits must be integrated and jointly utilized to answer real-world multifaceted questions. For instance, in medical domains, if one edit informs LLMs that COVID-19 causes "fever" and another that it causes "loss of taste", a qualified compositional editor should enable LLMs to answer the question "What are the symptoms of COVID-19?" with both "fever" and "loss of taste" (and potentially more). In this work, we define and systematically benchmark this compositional model editing (CME) task, identifying three key undesirable issues that existing methods struggle with: *knowledge loss*, *incorrect preceding* and *knowledge sinking*. To overcome these issues, we propose A$^3$E, a novel compositional editor that (1) ***a**daptively combines and **a**daptively regularizes* pre-trained foundation knowledge in LLMs in the stage of edit training and (2) ***a**daptively merges* multiple edits to better meet compositional needs in the stage of edit composing. Extensive experiments demonstrate that A$^3$E improves the composability by at least 22.45\% without sacrificing the performance of non-compositional model editing.

Matrix preconditioning is a critical technique to accelerate the solution of linear systems, where performance heavily depends on the selection of preconditioning parameters. Traditional parameter selection approaches often define fixed constants for specific scenarios. However, they rely on domain expertise and fail to consider the instance-wise features for individual problems, limiting their performance. In contrast, machine learning (ML) approaches, though promising, are hindered by high inference costs and limited interpretability. To combine the strengths of both approaches, we propose a symbolic discovery framework—namely, **Sym**bolic **Ma**trix **P**reconditioning (**SymMaP**)—to learn efficient symbolic expressions for preconditioning parameters. Specifically, we employ a neural network to search the high-dimensional discrete space for expressions that can accurately predict the optimal parameters. The learned expression allows for high inference efficiency and excellent interpretability (expressed in concise symbolic formulas), making it simple and reliable for deployment. Experimental results show that SymMaP consistently outperforms traditional strategies across various benchmarks.

Recent advances in large language models (LLMs) and audio language models have significantly improved music generation, particularly in lyrics-to-song generation. However, existing approaches still struggle with the complex composition of songs and the scarcity of high-quality data, leading to limitations in audio quality, musicality, instruction following, and vocal-instrument harmony. To address these challenges, we introduce LeVo, a language model based framework consisting of LeLM and Music Codec. LeLM is capable of parallel modeling of two types of tokens: mixed tokens, which represent the combined audio of vocals and accompaniment to achieve better vocal-instrument harmony, and dual-track tokens, which separately encode vocals and accompaniment for high-quality song generation. It employs two decoder-only transformers and a modular extension training strategy to prevent interference between different token types. To further enhance musicality and instruction following ability, we introduce a multi-preference alignment method based on Direct Preference Optimization (DPO). This method handles diverse human preferences through a semi-automatic data construction process and post-training. Experimental results demonstrate that LeVo significantly outperforms existing open-source methods in both objective and subjective metrics, while performing competitively with industry systems. Ablation studies further justify the effectiveness of our designs. Audio examples and source code are available at https://levo-demo.github.io and https://github.com/tencent-ailab/songgeneration.