NeurIPS 2025 Papers

Length generalization—the ability to solve problems longer than those seen during training—remains a critical challenge for large language models (LLMs). Previous work modifies positional encodings (PEs) and data formats to improve length generalization on specific symbolic tasks such as addition and sorting. However, these approaches are fundamentally limited to special tasks, often degrading general language performance. Furthermore, they are typically evaluated on small transformers trained from scratch on single tasks and can cause performance drop when applied during post-training stage of practical LLMs with general capabilities. Hu et al., (2024) proposed Rule-Following Fine-Tuning (RFFT) to improve length generalization in the post-training stage of LLMs. Despite its compatibility with practical models and strong performance, RFFT is proposed for single tasks too, requiring re-training for each individual task with extensive examples. In this paper, we study length generalization in multi-task settings and propose *Meta Rule-Following Fine-Tuning (Meta-RFFT)*, the first framework enabling robust *cross-task* length generalization. As our first contribution, we construct a large length generalization dataset containing **86 tasks** spanning code execution, number processing, symbolic and logical reasoning tasks, beyond the common addition or multiplication tasks. Secondly, we show that cross-task length generalization is possible with Meta-RFFT—after training on a large number of tasks and instances, the models achieve remarkable length generalization ability on *unseen* tasks with *minimal fine-tuning or one-shot prompting*. For example, after fine-tuning on 1 to 5 digit addition, our 32B model **achieves 95% accuracy on 30 digit addition**, significantly outperforming the state-of-the-art reasoning models (DeepSeek-R1-671B: 72%; QwQ-32B: 32%), despite never seeing this task during RF-pretraining.

We study the problem of controlling worst-case errors in misspecified linear regression under the random design setting, where the regression function is estimated via (penalized) least-squares. This setting arises naturally in value function approximation for bandit algorithms and reinforcement learning (RL). Our first main contribution is the observation that the amplification of the misspecification error when using least-squares is governed by the \emph{Lebesgue constant}, a classical quantity from approximation theory that depends on the choice of the feature subspace and the covariate distribution. We also show that this dependence on the misspecification error is tight for least-squares regression: in general, no method minimizing the empirical squared loss, including regularized least-squares, can improve it substantially. We argue this explains the empirical observation that some feature-maps (e.g., those derived from the Fourier bases) ``work better in RL'' than others (e.g., polynomials): given some covariate distribution, the Lebesgue constant is known to be highly sensitive to choice of the feature-map. As a second contribution, we propose a method that augments the original feature set with auxiliary features designed to reduce the error amplification. We then prove that the method successfully competes with an "oracle'' that knows the best way of using the auxiliary features to reduce this amplification. For example, when the domain is a real interval and the features are monomials, our method reduces the amplification factor to $O(1)$ as $d\to\infty$, while without our method, least-squares with the monomials (and in fact polynomials) will suffer a worst-case error amplification of order $\Omega(d)$. It follows that there are functions and feature maps for which our method is consistent, while least-squares is inconsistent.

Programming assistants powered by large language models have transformed software development, yet most benchmarks focus narrowly on code generation tasks. Recent efforts like InfiBench and StackEval attempt to address this gap using Stack Overflow data but remain limited to single-turn interactions in isolated contexts, require significant manual curation, and fail to represent complete project environments. We introduce CodeAssistBench (CAB), the first benchmark framework for evaluating multi-turn programming assistance in realistic settings that address questions grounded in actual codebases. Unlike existing programming Q&A benchmarks, CAB automatically generates scalable datasets from GitHub issues tagged with questions using configurable parameters (e.g., repository creation date, star count, programming languages), and includes automatic containerization of codebases for evaluation. It then evaluates models through simulated users in these containerized environments with full codebase access. Using this framework, we constructed a test set of 3,286 real-world programming questions across 214 repositories, spanning seven programming languages and diverse problem domains. Our evaluation of leading LLMs reveals a substantial capability gap: while models perform well on Stack Overflow questions with success rates of 70-83%, they resolve only up to 16.49% of CAB's issues from recent repositories (post-training cutoff). This discrepancy highlights the challenges of providing assistance in complex, project-specific contexts versus answering standalone questions. Our fully automated framework enables continuous benchmark expansion and is available at https://github.com/amazon-science/CodeAssistBench/.

Recent advancements in Large Language Models (LLMs) have emphasized the critical role of fine-tuning (FT) techniques in adapting LLMs to specific tasks, especially when retraining from scratch is computationally infeasible. Fine-tuning enables LLMs to leverage task- or domain-specific data, producing models that more effectively meet the requirements of targeted applications. However, con- ventional FT approaches often suffer from catastrophic forgetting and suboptimal data efficiency, limiting their real-world applicability. To address these challenges, this paper proposes DEAL, a novel framework that integrates Low-Rank Adapta- tion (LoRA) with a continuous fine-tuning strategy. By incorporating knowledge retention and adaptive parameter update modules, the framework mitigates the limitations of existing FT methods while maintaining efficiency. Experiments on 15 diverse datasets show that DEAL consistently outperforms baseline methods, yielding substantial gains in task accuracy and resource efficiency. These find- ings demonstrate the potential of our approach to advance continual adaptation in LLMs by enhancing task performance while improving resource efficiency. The source code is publicly available at https://github.com/Applied-Machine-Learning-Lab/DEAL.

Series Section Electron Microscopy (ssEM) has emerged as a pivotal technology for deciphering nanoscale biological architectures. Three-dimensional (3D) registration is a critical step in ssEM, tasked with rectifying axial misalignments and nonlinear distortions introduced during serial sectioning. The core scientific challenge lies in achieving distortion mitigation without erasing the natural morphological deformations of biological tissues, thereby enabling faithful reconstruction of 3D ultrastructural organization. In this study, we present a paradigm-shifting optimization framework that rethinks 3D registration through the lens of manifold trajectory optimization. We propose the first continuous trajectory dynamics formulation for 3D registration and introduce a novel optimization strategy. Specifically, we introduce a dual optimization objective that inherently balances global trajectory smoothness with local structural preservation, while developing a solver that combines Gauss-Seidel iteration with Neural ODEs to systematically integrate biophysical priors with data-driven deformation compensation. A key strength of our method is its fully unsupervised training, which avoids reliance on ground truth and suits ssEM scenarios where annotations are difficult to obtain. Extensive experiments on multiple datasets spanning diverse tissue types demonstrate our method's superior performance in structural restoration accuracy and cross-tissue robustness.

The advancement of Embodied AI heavily relies on large-scale, simulatable 3D scene datasets characterized by scene diversity and realistic layouts. However, existing datasets typically suffer from limitations in data scale or diversity, sanitized layouts lacking small items, and severe object collisions. To address these shortcomings, we introduce \textbf{InternScenes}, a novel large-scale simulatable indoor scene dataset comprising approximately 40,000 diverse scenes by integrating three disparate scene sources, \ie, real-world scans, procedurally generated scenes, and designer-created scenes, including 1.96M 3D objects and covering 15 common scene types and 288 object classes. We particularly preserve massive small items in the scenes, resulting in realistic and complex layouts with an average of 41.5 objects per region. Our comprehensive data processing pipeline ensures simulatability by creating real-to-sim replicas for real-world scans, enhances interactivity by incorporating interactive objects into these scenes, and resolves object collisions by physical simulations. We demonstrate the value of InternScenes with two benchmark applications: scene layout generation and point-goal navigation. Both show the new challenges posed by the complex and realistic layouts. More importantly, InternScenes paves the way for scaling up the model training for both tasks, making the generation and navigation in such complex scenes possible. We commit to open-sourcing the data and benchmarks to benefit the whole community.

Recommender systems are among the most impactful AI applications, interacting with billions of users every day, guiding them to relevant products, services, or information tailored to their preferences. However, the research and development of recommender systems are hindered by existing datasets that fail to capture realistic user behaviors and inconsistent evaluation settings that lead to ambiguous conclusions. This paper introduces the \textbf{O}pen \textbf{R}ecommendation \textbf{B}enchmark for Reproducible Research with H\textbf{I}dden \textbf{T}ests (\textbf{ORBIT}), a unified benchmark for consistent and realistic evaluation of recommendation models. ORBIT offers a standardized evaluation framework of public datasets with reproducible splits and transparent settings for its public leaderboard. Additionally, ORBIT introduces a new webpage recommendation task, ClueWeb-Reco, featuring web browsing sequences from 87 million public, high-quality webpages. ClueWeb-Reco is a synthetic dataset derived from real, user-consented, and privacy-guaranteed browsing data. It aligns with modern recommendation scenarios and is reserved as the hidden test part of our leaderboard to challenge recommendation models' generalization ability. ORBIT measures 12 representative recommendation models on its public benchmark and introduces a prompted LLM baseline on the ClueWeb-Reco hidden test. Our benchmark results reflect general improvements of recommender systems on the public datasets, with variable individual performances. The results on the hidden test reveal the limitations of existing approaches in large-scale webpage recommendation and highlight the potential for improvements with LLM integrations. ORBIT benchmark, leaderboard, and codebase are available at \url{https://www.open-reco-bench.ai}.

Reasoning methods that adaptively allocate test-time compute have advanced LLM performance on easy to verify domains such as math and code. In this work, we study how to utilize this approach to train models that exhibit a degree of robustness to safety vulnerabilities, and show that doing so can provide benefits. We build a recipe called $\textit{\textbf{TARS}}$ (Training Adaptive Reasoners for Safety), a reinforcement learning (RL) approach that trains models to reason about safety using chain-of-thought traces and a reward signal that balances safety with task completion. To build TARS, we identify three critical design choices: (1) a ``lightweight'' warmstart SFT stage, (2) a mix of harmful, harmless, and ambiguous prompts to prevent shortcut behaviors such as too many refusals, and (3) a reward function to prevent degeneration of reasoning capabilities during training. Models trained with TARS exhibit adaptive behaviors by spending more compute on ambiguous queries, leading to better safety-refusal trade-offs. They also internally learn to better distinguish between safe and unsafe prompts and attain greater robustness to both white-box (e.g., GCG) and black-box attacks (e.g., PAIR). Overall, our work provides an effective, open recipe for training LLMs against jailbreaks and harmful requests by reasoning per prompt.

Machine unlearning (MU) aims to remove the influence of specific ''forget'' data from a trained model while preserving its knowledge of the remaining ''retain'' data. Existing MU methods based on label manipulation or model weight perturbations often achieve limited unlearning effectiveness. To address this, we introduce CoUn, a novel MU framework inspired by the observation that a model retrained from scratch using only retain data classifies forget data based on their semantic similarity to the retain data. CoUn emulates this behavior by adjusting learned data representations through contrastive learning (CL) and supervised learning, applied exclusively to retain data. Specifically, CoUn (1) leverages semantic similarity between data samples to indirectly adjust forget representations using CL, and (2) maintains retain representations within their respective clusters through supervised learning. Extensive experiments across various datasets and model architectures show that CoUn consistently outperforms state-of-the-art MU baselines in unlearning effectiveness. Additionally, integrating our CL module into existing baselines empowers their unlearning effectiveness.

Domain Adaptation of Black-box Predictors (DABP) transfers knowledge from a labeled source domain to an unlabeled target domain, without requiring access to either source data or source model. Common practices of DABP leverage reliable samples to suppress negative information about unreliable samples. However, there are still some problems: i) Excessive attention to reliable sample aggregation leads to premature overfitting; ii) Valuable information in unreliable samples is often overlooked. To address them, we propose a novel spatial learning approach, called Controlled Visual Hallucination via Thalamus-driven Decoupling Network (CVH-TDN). Specifically, CVH-TDN is the first work that introduces the thalamus-driven decoupling network in the visual task, relying on its connection with hallucination to control the direction of sample generation in feature space. CVH-TDN is composed of Hallucination Generation (HG), Hallucination Alignment (HA), and Hallucination Calibration (HC), aiming to explore the spatial relationship information between samples and hallucinations. Extensive experiments confirm that CVH-TDN achieves SOTA performance on four standard benchmarks.

Despite the rapid advancement of LLM-based agents, the reliable evaluation of their safety and security remains a significant challenge. Existing rule-based or LLM-based evaluators often miss dangers in agents' step-by-step actions, overlook subtle meanings, fail to see how small issues compound, and get confused by unclear safety or security rules. To overcome this evaluation crisis, we introduce AgentAuditor, a universal, training-free, memory-augmented reasoning framework that empowers LLM evaluators to emulate human expert evaluators. AgentAuditor constructs an experiential memory by having an LLM adaptively extract structured semantic features (e.g., scenario, risk, behavior) and generate associated chain-of-thought reasoning traces for past interactions. A multi-stage, context-aware retrieval-augmented generation process then dynamically retrieves the most relevant reasoning experiences to guide the LLM evaluator's assessment of new cases. Moreover, we developed ASSEBench, the first benchmark designed to check how well LLM-based evaluators can spot both safety risks and security threats. ASSEBench comprises 2293 meticulously annotated interaction records, covering 15 risk types across 29 application scenarios. A key feature of ASSEBench is its nuanced approach to ambiguous risk situations, employing "Strict" and "Lenient" judgment standards. Experiments demonstrate that AgentAuditor not only consistently improves the evaluation performance of LLMs across all benchmarks but also sets a new state-of-the-art in LLM-as-a-judge for agent safety and security, achieving human-level accuracy. Our work is openly accessible at https://github.com/Astarojth/AgentAuditor.

Reinforcement learning (RL) has become popular in enhancing the reasoning capabilities of large language models (LLMs), with Group Relative Policy Optimization (GRPO) emerging as a widely used algorithm in recent systems. Despite GRPO's widespread adoption, we identify a previously unrecognized phenomenon we term Lazy Likelihood Displacement (LLD), wherein the likelihood of correct responses marginally increases or even decreases during training. This behavior mirrors a recently discovered misalignment issue in Direct Preference Optimization (DPO), attributed to the influence of negative gradients. We provide a theoretical analysis of GRPO’s learning dynamic, identifying the source of LLD as the naive penalization of all tokens in incorrect responses with the same strength. To address this, we develop a method called NTHR, which downweights penalties on tokens contributing to the LLD. Unlike prior DPO-based approaches, NTHR takes advantage of GRPO’s group-based structure, using correct responses as anchors to identify influential tokens. Experiments on math reasoning benchmarks demonstrate that NTHR effectively mitigates LLD, yielding consistent performance gains across models ranging from 0.5B to 3B parameters.

Motivated by the geometric advantages of quaternions in representing rotations and postures, we propose a quaternion-valued supervised learning Hopfield-structured neural network (QSHNN) with a fully connected structure inspired by the classic Hopfield neural network (HNN). Starting from a continuous-time dynamical model of HNNs, we extend the formulation to the quaternionic domain and establish the existence and uniqueness of fixed points with asymptotic stability. For the learning rules, we introduce a periodic projection strategy that modifies standard gradient descent by periodically projecting each $4\times 4$ block of the weight matrix onto the closest quaternionic structure in the least-squares sense. This approach preserves both convergence and quaternionic consistency throughout training. Benefiting from this rigorous mathematical foundation, the experimental model implementation achieves high accuracy, fast convergence, and strong reliability across randomly generated target sets. Moreover, the evolution trajectories of the QSHNN exhibit well-bounded curvature, i.e., sufficient smoothness, which is crucial for applications such as control systems or path planning modules in robotic arms, where joint postures are parameterized by quaternion neurons. Beyond these application scenarios, the proposed model offers a practical implementation framework and a general mathematical methodology for designing neural networks under hypercomplex or non-commutative algebraic structures.

Graph domain adaptation (GDA) is a fundamental task in graph machine learning, with techniques like shift-robust graph neural networks (GNNs) and specialized training procedures to tackle the distribution shift problem. Although these model-centric approaches show promising results, they often struggle with severe shifts and constrained computational resources. To address these challenges, we propose a novel model-free framework, GRADATE (GRAph DATa sElection), that selects the best training data from the source domain for the classification task on the target domain. GRADATE picks training samples without relying on any GNN model’s predictions or training recipes, leveraging optimal transport theory to capture and adapt to distribution changes. GRADATE is data-efficient, scalable and meanwhile complements existing model-centric GDA approaches. Through comprehensive empirical studies on several real-world graph-level datasets and multiple covariate shift types, we demonstrate that GRADATE outperforms existing selection methods and enhances off-the-shelf GDA methods with much fewer training data.

Despite significant advances in Multimodal Large Language Models (MLLMs), understanding complex temporal dynamics in videos remains a major challenge. Our experiments show that current Video Large Language Model (Video-LLM) architectures have critical limitations in temporal understanding, struggling with tasks that require detailed comprehension of action sequences and temporal progression. In this work, we propose a Video-LLM architecture that introduces stacked temporal attention modules directly within the vision encoder. This design incorporates a temporal attention in vision encoder, enabling the model to better capture the progression of actions and the relationships between frames before passing visual tokens to the LLM. Our results show that this approach significantly improves temporal reasoning and outperforms existing models in video question answering tasks, specifically in action recognition. We improve on benchmarks including VITATECS, MVBench, and Video-MME by up to +5.5%. By enhancing the vision encoder with temporal structure, we address a critical gap in video understanding for Video-LLMs. Project page and code are available at: https://alirasekh.github.io/STAVEQ2/

We introduce Dr. RAW, a unified and tuning-efficient framework for high-level computer vision tasks directly operating on camera RAW data. Unlike previous approaches that optimize image signal processing (ISP) pipelines and fully fine-tune networks for each task, Dr. RAW achieves state-of-the-art performance with minimal parameter updates. At the input stage, we apply lightweight pre-processing modules, sensor and illumination mapping, followed by re-mosaicing, to mitigate data inconsistencies stemming from sensor variation and lighting. At the network level, we introduce task-specific adaptation through two modules: Sensor Prior Prompts (SPP) and Low-Rank Adaptation (LoRA). SPP injects sensor-aware conditioning into the network via learnable prompts derived from imaging priors, while LoRA enables efficient task-specific tuning by updating only low-rank matrices in key backbone layers. Despite minimal tuning, our method delivers superior results across four RAW-based tasks (object detection, semantic segmentation, instance segmentation, and pose estimation) on nine datasets encompassing low-light and over-exposed conditions. By harnessing the intrinsic physical cues of RAW data alongside parameter-efficient techniques, our method advances RAW-based vision systems, achieving both high accuracy and computational economy. We will release our source code.

Self-supervised learning (SSL) has achieved major advances in natural images and video understanding, but challenges remain in domains like echocardiography (heart ultrasound) due to subtle anatomical structures, complex temporal dynamics, and the current lack of domain-specific pre-trained models. Existing SSL approaches such as contrastive, masked modeling, and clustering-based methods struggle with high intersample similarity, sensitivity to low PSNR inputs common in ultrasound, or aggressive augmentations that distort clinically relevant features. We present DISCOVR (Distilled Image Supervision for Cross Modal Video Representation), a self-supervised dual-branch framework for cardiac ultrasound video representation learning. DISCOVR combines a clustering-based video encoder that models temporal dynamics with an online image encoder that extracts fine-grained spatial semantics. These branches are connected through a semantic cluster distillation loss that transfers anatomical knowledge from the evolving image encoder to the video encoder, enabling temporally coherent representations enriched with fine-grained semantic understanding. Evaluated on six echocardiography datasets spanning fetal, pediatric, and adult populations, DISCOVR outperforms both specialized video anomaly detection methods and state-of-the-art video-SSL baselines in zero-shot and linear probing setups, achieving superior segmentation transfer and strong downstream performance on clinically relevant tasks such as LVEF prediction. **Code available at:** [https://github.com/mdivyanshu97/DISCOVR](https://github.com/mdivyanshu97/DISCOVR)

While model serving has unlocked unprecedented capabilities, the high cost of serving large-scale models continues to be a significant barrier to widespread accessibility and rapid innovation. Compiler optimizations have long driven substantial performance improvements, but existing compilers struggle with neural workloads due to the exponentially large and highly interdependent space of possible transformations. Although existing stochastic search techniques can be effective, they are often sample-inefficient and fail to leverage the structural context underlying compilation decisions. We set out to investigate the research question of whether reasoning with large language models (LLMs), without any retraining, can leverage the context-aware decision space of compiler optimizations to significantly improve sample efficiency. To that end, we introduce a novel compilation framework (dubbed Reasoning Compiler) that formulates optimization as a sequential, context-aware decision process guided by a large language model and structured Monte Carlo tree search (MCTS). The LLM acts as a proposal mechanism, suggesting hardware-informed transformations that reflect the current program state and accumulated performance feedback. MCTS incorporates the LLM-generated proposals to balance exploration and exploitation, facilitating structured, context-sensitive traversal of the expansive compiler optimization space. By achieving substantial speedups with markedly fewer samples than leading neural compilers, our approach demonstrates the potential of LLM-guided reasoning to transform the landscape of compiler optimization.

Correct model interpretation in high-stakes settings is critical, yet both post-hoc feature attribution methods and so-called intrinsically interpretable models can systematically attribute false-positive importance to non-informative features such as suppressor variables. Specifically, both linear models and their powerful non-linear generalisation such as General Additive Models (GAMs) are susceptible to spurious attributions to suppressors. We present a principled generalisation of activation patterns - originally developed to make linear models interpretable - to additive models, correctly rejecting suppressor effects for non-linear features. This yields PatternGAM, an importance attribution method based on univariate generative surrogate models for the broad family of additive models, and PatternQLR for polynomial models. Empirical evaluations on the XAI-TRIS benchmark with a novel false-negative invariant formulation of the earth mover's distance accuracy metric demonstrates significant improvements over popular feature attribution methods and the traditional interpretation of additive models. Finally, real-world case studies on the COMPAS and MIMIC-IV datasets provide new insights into the role of specific features by disentangling genuine target-related information from suppression effects that would mislead conventional GAM interpretations.