NeurIPS 2025 Papers

While large language models (LLMs) are driving the rapid advancement of artificial intelligence, effectively and reliably training these large models remains one of the field's most significant challenges. To address this challenge, we propose POET, a novel reParameterized training algorithm that uses Orthogonal Equivalence Transformation to optimize neurons. Specifically, POET reparameterizes each neuron with two learnable orthogonal matrices and a fixed random weight matrix. Because of its provable preservation of spectral properties of weight matrices, POET can stably optimize the objective function with improved generalization. We further develop efficient approximations that make POET flexible and scalable for training large-scale neural networks. Extensive experiments validate the effectiveness and scalability of POET in training LLMs.

Adapting billion-parameter language models to a downstream task is still costly, even with parameter-efficient fine-tuning (PEFT). We re-cast task adaptation as output-distribution alignment: the objective is to steer the output distribution toward the task distribution directly during decoding rather than indirectly through weight updates. Building on this view, we introduce Steering Vector Decoding (SVDecode), a lightweight, PEFT-compatible, and theoretically grounded method. We start with a short warm-start fine-tune and extract a task-aware steering vector from the Kullback-Leibler (KL) divergence gradient between the output distribution of the warm-started and pre-trained models. This steering vector is then used to guide the decoding process to steer the model's output distribution towards the task distribution. We theoretically prove that SVDecode is first-order equivalent to the gradient step of full fine-tuning and derive a globally optimal solution for the strength of the steering vector. Across three tasks and nine benchmarks, SVDecode paired with four standard PEFT methods improves multiple-choice accuracy by up to 5 percentage points and open-ended truthfulness by 2 percentage points, with similar gains (1-2 percentage points) on commonsense datasets without adding trainable parameters beyond the PEFT adapter. SVDecode thus offers a lightweight, theoretically grounded path to stronger task adaptation for large language models.

Post-training of large language models is essential for adapting pre-trained language models (PLMs) to align with human preferences and downstream tasks. While PLMs typically exhibit well-calibrated confidence, post-trained language models (PoLMs) often suffer from over-confidence, assigning high confidence to both correct and incorrect outputs, which can undermine reliability in critical applications. A major obstacle in calibrating PoLMs is the scarcity of labeled data for individual downstream tasks. To address this, we propose Disagreement-Aware Confidence Alignment (DACA), a novel unsupervised method to optimize the parameters (e.g., temperature $\tau$) in post-hoc confidence calibration. Our method is motivated by the under-confidence issue caused by prediction disagreement between the PLM and PoLM while aligning their confidence via temperature scaling. Theoretically, the PLM's confidence underestimates PoLM's prediction accuracy on disagreement examples, causing a larger $\tau$ and producing under-confident predictions. DACA mitigates this by selectively using only agreement examples for calibration, effectively decoupling the influence of disagreement. In this manner, our method avoids an overly large $\tau$ in temperature scaling caused by disagreement examples, improving calibration performance. Extensive experiments demonstrate the effectiveness of our method, improving the average ECE of open-sourced and API-based LLMs (e.g. GPT-4o) by up to 15.08$\%$ on common benchmarks.

Robust unlearning is crucial for safely deploying large language models (LLMs) in environments where data privacy, model safety, and regulatory compliance must be ensured. Yet the task is inherently challenging, partly due to difficulties in reliably measuring whether unlearning has truly occurred. Moreover, fragmentation in current methodologies and inconsistent evaluation metrics hinder comparative analysis and reproducibility. To unify and accelerate research efforts, we introduce OpenUnlearning, a standardized and extensible framework designed explicitly for benchmarking both LLM unlearning methods and metrics. OpenUnlearning integrates 13 state-of-the-art unlearning algorithms and 16 diverse evaluations across 3 leading benchmarks (TOFU, MUSE, and WMDP) and also enables analyses of forgetting behaviors across 450+ publicly released checkpoints. Leveraging OpenUnlearning, we propose a novel meta-evaluation benchmark focused specifically on assessing the faithfulness and robustness of evaluation metrics themselves. We also benchmark diverse unlearning methods and provide a comparative analysis against an extensive evaluation suite. Overall, we establish a clear, community-driven pathway toward rigorous development in LLM unlearning research.

While large language models (LLMs) demonstrate impressive capabilities across numerous applications, their robustness remains a critical concern. This paper is motivated by a specific vulnerability: the order sensitivity of LLMs. This vulnerability manifests itself as the order bias observed when LLMs decide between possible options (for example, a preference for the first option) and the tendency of LLMs to provide different answers when options are reordered. The use cases for this scenario extend beyond the classical case of multiple-choice question answering to the use of LLMs for multidocument tasks and as automated evaluators in AI pipelines. We introduce Set-LLM, a novel architectural adaptation for pretrained LLMs that enables the processing of mixed set-text inputs with permutation invariance guarantees. The adaptations involve a new attention mask and new positional encodings specifically designed for sets. We provide a theoretical proof of invariance and demonstrate through experiments that Set-LLM can be trained effectively, achieving comparable or improved performance and maintaining the runtime of the original model, while altogether eliminating order sensitivity.

Can we build accurate world models out of large language models (LLMs)? How can world models benefit LLM agents? The gap between the prior knowledge of LLMs and the specified environment's dynamics usually bottlenecks LLMs' performance as world models. To bridge the gap, we propose a training-free "world alignment" that learns an environment's symbolic knowledge complementary to LLMs. The symbolic knowledge covers action rules, knowledge graphs, and scene graphs, which are extracted by LLMs from exploration trajectories and encoded into executable codes to regulate LLM agents' policies. We further propose an RL-free, model-based agent "WALL-E" through the model-predictive control (MPC) framework. Unlike classical MPC requiring costly optimization on the fly, we adopt an LLM agent as an efficient look-ahead optimizer of future steps' actions by interacting with the neurosymbolic world model. While the LLM agent's strong heuristics make it an efficient planner in MPC, the quality of its planned actions is also secured by the accurate predictions of the aligned world model. They together considerably improve learning efficiency in a new environment. On open-world challenges in Mars (Minecraft like) and ALFWorld (embodied indoor environments), WALL-E significantly outperforms existing methods, e.g., surpassing baselines in Mars by 16.1%–51.6% of success rate and by at least 61.7% in score. In ALFWorld, it achieves a new record 98% success rate after only 4 iterations.

Large language models (LLMs) often produce reasoning steps that are superficially coherent yet internally inconsistent, leading to unreliable outputs. Since such failures typically arise from implicit or poorly-grounded knowledge, we introduce \emph{Grounded Reasoning in Dependency (GRiD)}, a novel dependency-aware reasoning framework that explicitly grounds reasoning steps in structured knowledge. GRiD represents reasoning as a graph consisting of interconnected knowledge extraction nodes and reasoning nodes, enforcing logical consistency through explicit dependencies. Each reasoning step is validated via a lightweight, step-wise verifier that ensures logical correctness relative to its premises. Extensive experiments across diverse reasoning benchmarks—including StrategyQA, CommonsenseQA, GPQA, and TruthfulQA—demonstrate that GRiD substantially improves reasoning accuracy, consistency, and faithfulness compared to recent state-of-the-art structured reasoning methods. Notably, GRiD enhances performance even when applied purely as a lightweight verification module at inference time, underscoring its generalizability and practical utility. Code is available at: https://github.com/cure-lab/GRiD.

Large language model (LLM) routers improve the efficiency of multi-model systems by directing each query to the most appropriate model while leveraging the diverse strengths of heterogeneous LLMs. Most existing approaches frame routing as a classification problem based solely on the input query. While this reduces overhead by avoiding inference across all models, it overlooks valuable information that could be gleaned from potential outputs and fails to capture implicit intent or contextual nuances that often emerge only during response generation. These limitations can result in suboptimal routing decisions, particularly for complex or ambiguous queries that require deeper semantic understanding. To address this challenge, we propose Lookahead, a routing framework that "foresees" potential model outputs by predicting their latent representations and uses these predictions to guide model selection, thus enabling more informed routing without full inference. Within this framework, we implement two approaches based on causal and masked language models. Empirical evaluations across seven public benchmarks—spanning instruction following, mathematical reasoning, and code generation—show that Lookahead consistently outperforms existing routing baselines, achieving an average performance gain of 7.7\% over the state-of-the-art. Our code is available at https://github.com/huangcb01/lookahead-routing.

Finetuning large language models (LLMs) is a resource-intensive task for researchers in academia, with memory constraints posing a key bottleneck. A classic optimization method, block coordinate descent (BCD), significantly reduces memory cost by segmenting the trainable parameters into multiple blocks and optimizing one active block at a time while freezing the others. However, we identify that blindly applying BCD to train LLMs can be inefficient for two reasons. First, optimizing only the active block requires backpropagating through multiple deeper yet inactive blocks, resulting in wasteful computations. Second, the frozen blocks, when they are not quite close to optimality, can narrow the optimization landscape, potentially misguiding the training of the active block. To address these issues simultaneously, we propose integrating BCD with *landscape expansion*, which unfreezes the inactive blocks and updates them in a cost-efficient manner during the same backpropagation as the update to the active block. Experiments on 8B and 70B models demonstrate that our proposed method surpasses memory-efficient baselines and matches Adam's downstream performance while requiring only 24 GB of memory for the 8B model and 300 GB for the 70B model.

Large language models (LLMs) have recently transformed from text-based assistants to autonomous agents capable of planning, reasoning, and iteratively improving their actions. While numerical reward signals and verifiers can effectively rank candidate actions, they often provide limited contextual guidance. In contrast, natural language feedback better aligns with the generative capabilities of LLMs, providing richer and more actionable suggestions. However, parsing and implementing this feedback effectively can be challenging for LLM-based agents. In this work, we introduce Critique-Guided Improvement (CGI), a novel two-player framework, comprising an actor model that explores an environment and a critic model that generates detailed nature language feedback. By training the critic to produce fine-grained assessments and actionable revisions, and the actor to utilize these critiques, our approach promotes more robust exploration of alternative strategies while avoiding local optima. Experiments in three interactive environments show that CGI outperforms existing baselines by a substantial margin. Notably, even a small critic model surpasses GPT-4 in feedback quality. The resulting actor achieves state-of-the-art performance, demonstrating the power of explicit iterative guidance to enhance decision-making in LLM-based agents.

Large Language Models (LLMs) have demonstrated remarkable capabilities, yet their extensive parameter scales pose significant challenges for practical deployment. Unstructured pruning has emerged as an effective model compression strategy with minimal performance loss, which introduces fine-grained sparsity for weight parameters. While existing methods employ a layer-wise pruning strategy to avoid the complexity of global pruning for billion-scale LLMs, they require appropriate sparsity allocation for the layer-wise pruning objectives and often lead to suboptimal solutions for the overall model. In this paper, we propose Lua-LLM ($\textbf{L}$earning $\textbf{u}$nstructured-sparsity $\textbf{a}$llocation in LLMs), a learning-based global pruning framework that explores the optimal unstructured sparsity allocation. Unlike existing pruning methods, which primarily focus on allocating per-layer sparsity, Lua-LLM achieves flexible allocation for both layer-wise and intra-layer sparsity. Furthermore, Lua-LLM leverages a soft Top-K operator to approximate the importance-based mask selection mechanism, enabling efficient binary mask learning. Experimental results on LLaMA and OPT families demonstrate significant performance improvements over existing methods.

Measuring parameter importance is crucial for understanding and optimizing large language models (LLMs). Existing work predominantly focuses on pruning or probing at neuron/feature levels without fully considering the cooperative behaviors of model parameters. In this paper, we introduce a novel approach--Model Shapley to quantify parameter importance based on the Shapley value, a principled method from cooperative game theory that captures both individual and synergistic contributions among parameters, via only one gradient backpropagation. We derive a scalable second-order approximation to compute Shapley values at the parameter level, leveraging blockwise Fisher information for tractability in large-scale settings. Our method enables fine-grained differentiation of parameter importance, facilitating targeted knowledge injection and model compression. Through mini-batch Monte Carlo updates and efficient approximation of the Hessian structure, we achieve robust Shapley-based attribution with only modest computational overhead. Experimental results indicate that this cooperative game perspective enhances interpretability, guides more effective parameter-specific fine-tuning and model compressing, and paves the way for continuous model improvement in various downstream tasks.

Prompt caching is critical for reducing latency and cost in LLM inference---OpenAI and Anthropic report up to 50–90\% cost savings through prompt reuse. Despite its widespread success, little is known about what constitutes an optimal prompt caching policy, particularly when optimizing tail latency—a metric of central importance to practitioners. The widely used Least Recently Used (LRU) policy can perform arbitrarily poor on this metric, as it is oblivious to the heterogeneity of conversation lengths. To address this gap, we propose Tail-Optimized LRU, a simple two-line modification that reallocates KV cache capacity to prioritize high-latency conversations by evicting cache entries that are unlikely to affect future turns. Though the implementation is simple, we prove its optimality under a natural stochastic model of conversation dynamics, providing the first theoretical justification for LRU in this setting---a result that may be of independent interest to the caching community. Experimentally, on real conversation data WildChat~\citep{zhao2024wildchat}, Tail-Optimized LRU achieves up to 27.5\% reduction in P90 tail Time to First Token latency and 23.9\% in P95 tail latency compared to LRU, along with up to 38.9\% decrease in SLO violations of 200ms. We believe this provides a practical and theoretically grounded option for practitioners seeking to optimize tail latency in real-world LLM deployments.

Policy-based methods currently dominate reinforcement learning (RL) pipelines for large language model (LLM) reasoning, leaving value-based approaches largely unexplored. We revisit the classical paradigm of Bellman Residual Minimization and introduce Trajectory Bellman Residual Minimization (TBRM), an algorithm that naturally adapts this idea to LLMs, yielding a simple yet effective off-policy algorithm that optimizes a single trajectory-level Bellman objective using the model's own logits as $Q$-values. TBRM removes the need for critics, importance-sampling ratios, or clipping, and can operate with only one rollout per prompt. We prove convergence to the near-optimal KL-regularized policy from arbitrary off-policy data via an improved change-of-trajectory-measure analysis. Experiments on standard mathematical-reasoning benchmarks show that TBRM matches or surpasses policy-based baselines, like PPO and GRPO, with comparable or lower computational and memory overhead. Our results indicate that value-based RL might be a principled and efficient alternative for enhancing reasoning capabilities in LLMs. The codebase for TBRM is publicly available at [https://github.com/rlx-lab/TBRM](https://github.com/rlx-lab/TBRM).

Recent methods in language model interpretability employ techniques such as sparse autoencoders to decompose residual stream contributions into linear, semantically-meaningful features. Our work demonstrates that an underlying assumption of these methods—that residual stream contributions build additively upon each other—is insufficient to fully explain model behavior. Specifically, we identify the Transformer Layer Correction Mechanism (TLCM), wherein adjacent transformer layers systematically counteract each other's contributions to the residual stream. TLCM appears in 5 out of 7 major open-source model families and activates across nearly all tokens in diverse texts. To understand TLCM, we show that it emerges during pretraining, operates most strongly on punctuation and numbers, and adaptively calibrates its correction strength based on the preceding layer's output. We further show that TLCM actively corrects a small subspace and promotes other subspaces, different from standard model behavior. We advance the ``propose-and-reject'' hypothesis: layers may propose multiple candidate features, while subsequent layers selectively filter out inappropriate ones. Finally, we discuss how our findings help explain three persistent challenges in feature-based interpretability: why extracted features descriptions often suffer from low specificity; why feature-based interventions for model steering fail at low magnitude; why recent work finds cross-layer transcoders outperform SAEs.

Large language models (LLMs) have demonstrated remarkable proficiency in mainstream academic disciplines such as mathematics, physics, and computer science. However, human knowledge encompasses over 200 specialized disciplines, far exceeding the scope of existing benchmarks. The capabilities of LLMs in many of these specialized fields-particularly in light industry, agriculture, and service-oriented disciplines-remain inadequately evaluated. To address this gap, we present SuperGPQA, a comprehensive benchmark that evaluates graduate-level knowledge and reasoning capabilities across 285 disciplines. Our benchmark employs a novel Human-LLM collaborative filtering mechanism to eliminate trivial or ambiguous questions through iterative refinement based on both LLM responses and expert feedback. Our experimental results reveal significant room for improvement in the performance of current state-of-the-art LLMs across diverse knowledge domains (e.g., the reasoning-focused model Gemini-2.5-Pro achieved the highest accuracy of 63.56% on SuperGPQA), highlighting the considerable gap between current model capabilities and artificial general intelligence. Additionally, we present comprehensive insights from our management of a large-scale annotation process, involving over 80 expert annotators and an interactive Human-LLM collaborative system, offering valuable methodological guidance for future research initiatives of comparable scope.

Recent studies have shown that making a model spend more time thinking through longer Chain of Thoughts (CoTs) enables it to gain significant improvements in complex reasoning tasks. While current researches continue to explore the benefits of increasing test-time compute by extending the CoT lengths of Large Language Models (LLMs), we are concerned about a potential issue hidden behind the current pursuit of test-time scaling: Would excessively scaling the CoT length actually bring adverse effects to a model's reasoning performance? Our explorations on mathematical reasoning tasks reveal an unexpected finding that scaling with longer CoTs can indeed impair the reasoning performance of LLMs in certain domains. Moreover, we discover that there exists an optimal scaled length distribution that differs across different domains. Based on these insights, we propose a Thinking-Optimal Scaling strategy. Our method first uses a small set of seed data with varying response length distributions to teach the model to adopt different reasoning efforts for deep thinking. Then, the model selects its shortest correct response under different reasoning efforts on additional problems for self-improvement. Our self-improved models built upon Qwen2.5-32B-Instruct outperform other distillation-based 32B o1-like models across various math benchmarks, and achieve performance on par with the teacher model QwQ-32B-Preview that produces the seed data.

Large Language Models (LLMs) are rapidly saturating existing benchmarks, necessitating new open-ended evaluations. We introduce the Factorio Learning Environment (FLE), based on the game of Factorio, that tests agents in long-term planning, spatial reasoning, program synthesis, and resource optimization. FLE provides exponentially scaling challenges -- from basic automation to complex factories processing millions of resource units per second. We provide two settings: (1) open-play with the open-ended task of building the largest factory on an procedurally generated map and (2) lab-play consisting of 33 bounded tasks accross three settings with fixed resources. We demonstrate across both settings that models still lack strong spatial reasoning. In lab-play, we find that LLMs exhibit promising short-horizon skills, yet are unable to operate effectively in constrained environments, reflecting limitations in error analysis. In open-play, while LLMs discover automation strategies that improve growth (e.g electric-powered drilling), they fail to achieve complex automation (e.g electronic-circuit manufacturing)

Gated Linear Units (GLUs) have become essential components in the feed-forward networks of state-of-the-art Large Language Models (LLMs). However, they require twice as many memory reads compared to feed-forward layers without gating, due to the use of separate weight matrices for the gate and value streams. To address this bottleneck, we introduce Masked Gated Linear Units (MGLUs), a novel family of GLUs with an efficient kernel implementation. The core contribution of MGLUs include: (1) the Mixture of Element-wise Gating (MoEG) architecture that learns multiple binary masks, each determining gate or value assignments at the element level on a single shared weight matrix resulting in reduced memory transfer, and (2) FlashMGLU, a hardware-friendly kernel that yields up to a 19.7$\times$ inference-time speed-up over a na\"ive PyTorch MGLU and is 47\% more memory-efficient and 34\% faster than standard GLUs despite added architectural complexity on an RTX5090 GPU. In LLM experiments, the Swish-activated variant SwiMGLU preserves its memory advantages while matching—or even surpassing—the downstream accuracy of the SwiGLU baseline.