NeurIPS 2025 Papers

Policy exploration is critical in reinforcement learning (RL), where existing approaches include $\epsilon$-greedy, Gaussian process, etc. However, these approaches utilize preset stochastic processes and are indiscriminately applied in all kinds of RL tasks without considering task-specific features that influence policy exploration. Moreover, during RL training, the evolution of such stochastic processes is rigid, which typically only incorporates a decay in the variance, failing to adjust flexibly according to the agent's real-time learning status. Inspired by the analyzing and reasoning capability of large language models (LLMs), we design **LLM-Explorer** to adaptively generate task-specific exploration strategies with LLMs, enhancing the policy exploration in RL. In our design, we sample the learning trajectory of the agent during the RL training in a given task and prompt the LLM to analyze the agent's current policy learning status and then generate a probability distribution for future policy exploration. Updating the probability distribution periodically, we derive a stochastic process specialized for the particular task and dynamically adjusted to adapt to the learning process. Our design is a plug-in module compatible with various widely applied RL algorithms, including the DQN series, DDPG, TD3, and any possible variants developed based on them. Through extensive experiments on the Atari and MuJoCo benchmarks, we demonstrate LLM-Explorer's capability to enhance RL policy exploration, achieving an average performance improvement up to 37.27%. Our code is open-source at https://github.com/tsinghua-fib-lab/LLM-Explorer for reproducibility.

We introduce REAL, a benchmark and framework for multi-turn agent evaluations on deterministic simulations of real-world websites. REAL comprises high-fidelity, deterministic replicas of 11 widely-used websites across domains such as e-commerce, travel, communication, and professional networking. We also release a benchmark consisting of 112 practical tasks that mirror everyday complex user interactions requiring both accurate information retrieval and state-changing actions. All interactions occur within this fully controlled setting, eliminating safety risks and enabling robust, reproducible evaluation of agent capability and reliability. Our novel evaluation framework combines programmatic checks of website state for action-based tasks with rubric-guided LLM-based judgments for information retrieval. The framework supports both open-source and proprietary agent systems through a flexible evaluation harness that accommodates black-box commands within browser environments, allowing research labs to test agentic systems without modification. Our empirical results show that frontier language models achieve at most a 41% success rate on REAL, highlighting critical gaps in autonomous web navigation and task completion capabilities. Our framework supports easy integration of new tasks, reproducible evaluation, and scalable post-training data generation, marking a significant step forward in evaluating and advancing agent capabilities.

Compiler optimization is essential for improving program performance, yet modern compilers still depend on manually crafted transformation rules over intermediate representations (IRs). As compilers grow in complexity, maintaining these rule-based optimizations becomes increasingly labor-intensive and difficult to scale. Recent advances in large language models (LLMs) offer a promising alternative, but their effectiveness in compiler optimization remains limited—primarily due to the lack of IR-oriented datasets that expose models to diverse transformation samples in real-world scenarios (*optimization-sensitive samples*), hindering LLMs from learning rich and generalizable optimization strategies. In this paper, we introduce IR-OptSet, the first public optimization-sensitive dataset for advancing LLM-based IR optimizers. It comprises 170K LLVM IR samples from open-source repositories across 8 representative optimization domains. IR-OptSet defines two core tasks: Code Analysis and Optimized Code Generation, and provides tools for correctness verification, performance evaluation, and dataset expansion. In our experiments, fine-tuning three representative LLMs on IR-OptSet leads to significant accuracy improvements across both tasks. Moreover, the LLM fine-tuned with IR-OptSet *outperforms traditional compiler with the -O3 option* in 64 test cases in terms of performance. Further analysis reveals that IR-OptSet provides greater transformation diversity and representativeness than three widely used IR-oriented datasets, highlighting its potential to drive model-based IR optimization. IR-OptSet is publicly available at [https://huggingface.co/datasets/YangziResearch/IR-OptSet](https://huggingface.co/datasets/YangziResearch/IR-OptSet).

Post-training compression has been a widely employed approach to scale down large language model (LLM) and facilitate efficient inference. In various proposed compression methods, including pruning and quantization, calibration data plays a vital role by informing the weight importance and activation dynamic ranges. However, how calibration data impacts the LLM capability after compression is less explored. Few of the existing works, though recognizing the significance of this study, only investigate the language modeling or commonsense reasoning performance degradation from limited angles, like the data sources or sample amounts. More systematic research is still needed to examine the impacts on different LLM capabilities in terms of compositional properties and domain correspondence of calibration data. In this work, we aim at bridging this gap and further analyze underlying influencing mechanisms from the activation pattern perspective. Especially, we explore the calibration data's impacts on high-level complex reasoning capabilities, like math problem solving and code generation. Delving into the underlying mechanism, we find that the representativeness and diversity in activation space more fundamentally determine the quality of calibration data. Finally, we propose a calibration data curation framework based on such observations and analysis, enhancing the performance of existing post-training compression methods on preserving critical LLM capabilities. Our code is provided in [Link](https://github.com/BokwaiHo/COLA.git).

Key-Value (KV) cache eviction---which retains the KV pairs of the most important tokens while discarding less important ones---is a critical technique for optimizing both memory usage and inference latency in large language models (LLMs). However, existing approaches often rely on simple heuristics---such as attention weights---to measure token importance, overlooking the spatial relationships between token value states in the vector space. This often leads to suboptimal token selections and thus performance degradation. To tackle this problem, we propose a novel method, namely **AnDPro** (**An**chor **D**irection **Pro**jection), which introduces a projection-based scoring function to more accurately measure token importance. Specifically, AnDPro operates in the space of value vectors and leverages the projections of these vectors onto an *``Anchor Direction''*---the direction of the pre-eviction output---to measure token importance and guide more accurate token selection. Experiments on $16$ datasets from the LongBench benchmark demonstrate that AnDPro can maintain $96.07\\%$ of the full cache accuracy using only $3.44\\%$ KV cache budget, reducing KV cache budget size by $46.0\\%$ without compromising quality compared to previous state-of-the-arts.

Current evaluation paradigms for large language models (LLMs) represent a critical blind spot in AI research—relying on opaque numerical metrics that conceal fundamental limitations in spatial reasoning while providing no intuitive understanding of model capabilities. This deficiency creates a dangerous disconnect between reported performance and practical abilities, particularly for applications requiring physical world understanding. We introduce LTD-Bench, a breakthrough benchmark that transforms LLM evaluation from abstract scores to directly observable visual outputs by requiring models to generate drawings through dot matrices or executable code. This approach makes spatial reasoning limitations immediately apparent even to non-experts, bridging the fundamental gap between statistical performance and intuitive assessment. LTD-Bench implements a comprehensive methodology with complementary generation tasks (testing spatial imagination) and recognition tasks (assessing spatial perception) across three progressively challenging difficulty levels, methodically evaluating both directions of the critical language-spatial mapping. Our extensive experiments with state-of-the-art models expose an alarming capability gap: even LLMs achieving impressive results on traditional benchmarks demonstrate profound deficiencies in establishing bidirectional mappings between language and spatial concepts—a fundamental limitation that undermines their potential as genuine world models. Furthermore, LTD-Bench's visual outputs enable powerful diagnostic analysis, offering a potential approach to investigate model similarity. Our dataset and codes are available at https://github.com/walktaster/LTD-Bench.

Recent advancements in multimodal large language models (MLLMs) have focused on integrating multiple modalities, yet their ability to simultaneously process and reason across different inputs remains underexplored. We introduce OmniBench, a novel benchmark designed to evaluate models’ ability to recognize, interpret, and reason across visual, acoustic, and textual inputs simultaneously. We define language models capable of such tri-modal processing as omni-language models (OLMs). OmniBench features high-quality human annotations that require integrated understanding across all modalities. Our evaluation reveals that: i) open-source OLMs show significant limitations in instruction-following and reasoning in tri-modal contexts; and ii) most baseline models perform poorly (below 50% accuracy) even with textual alternatives to image/audio inputs. To address these limitations, we develop OmniInstruct, an 96K-sample instruction tuning dataset for training OLMs. We advocate for developing more robust tri-modal integration techniques and training strategies to enhance OLM performance. Codes and data could be found at https://m-a-p.ai/OmniBench/.

Tabular data have been playing a vital role in diverse real-world fields, including healthcare, finance, etc. With the recent success of Large Language Models (LLMs), early explorations of extending LLMs to the domain of tabular data have been developed. Most of these LLM-based methods typically first serialize tabular data into natural language descriptions, and then tune LLMs or directly infer on these serialized data. However, these methods suffer from two key inherent issues: (i) data perspective: existing data serialization methods lack universal applicability for structured tabular data, and may pose privacy risks through direct textual exposure, and (ii) model perspective: LLM fine-tuning methods struggle with tabular data, and in-context learning scalability is bottle-necked by input length constraints (suitable for few-shot learning). This work explores a novel direction of integrating LLMs into tabular data through logical decision tree rules as intermediaries, proposing a decision tree enhancer with LLM-derived rule for tabular prediction, DeLTa. The proposed DeLTa avoids tabular data serialization, and can be applied to full data learning setting without LLM fine-tuning. Specifically, we leverage the reasoning ability of LLMs to redesign an improved rule given a set of decision tree rules. Furthermore, we provide a calibration method for original decision trees via new generated rule by LLM, which approximates the error correction vector to steer the original decision tree predictions in the direction of ``errors'' reducing. Finally, extensive experiments on diverse tabular benchmarks show that our method achieves state-of-the-art performance.

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

The combination of linear transformations and nonlinear activation functions forms the foundation of most modern deep neural networks, enabling them to approximate highly complex functions. This paper explores the introduction of quadratic transformations to further increase the nonlinearity of the model, with the aim of enhancing the performance of existing architectures. To minimize the additional parameters and computational burden, we propose a lightweight quadratic enhancer that leverages matrix decomposition, weight sharing, and sparsification techniques. This approach introduces only a minimal and negligible increase in parameters and forward computation, while still yielding substantial improvements in model performance. We evaluate the effectiveness of the proposed method across three tasks: text classification, image classification, and fine-tuning large language models (LLMs). In all tasks, our approach demonstrates significant performance gains.

Recent advancements in large language models (LLMs) have significantly enhanced text generation capabilities, yet evaluating their performance in generative writing remains a challenge. Existing benchmarks primarily focus on generic text generation or limited in writing tasks, failing to capture the diverse requirements of high-quality written contents across various domains. To bridge this gap, we present WritingBench, a comprehensive benchmark designed to evaluate LLMs across 6 core writing domains and 100 subdomains. We further propose a query-dependent evaluation framework that empowers LLMs to dynamically generate instance-specific assessment criteria. This framework is complemented by a fine-tuned critic model for criteria-aware scoring, enabling evaluations in style, format and length. The framework's validity is further demonstrated by its data curation capability, which enables a 7B-parameter model to outperform the performance of GPT-4o in writing. We open-source the benchmark, along with evaluation tools and modular framework components, to advance the development of LLMs in writing.

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

One persistent challenge in LLM research is the development of attention mechanisms that are able to generalise from training on shorter contexts to inference on longer contexts. We propose two conditions that we expect all effective long-context attention mechanisms to have: scale-invariant total attention, and scale-invariant attention sparsity. Under a Gaussian assumption, we show that a simple position-dependent transformation of the attention logits is sufficient for these conditions to hold. Experimentally we find that the resulting scale-invariant attention scheme gives considerable benefits in terms of validation loss when zero-shot generalising from training on short contexts to validation on longer contexts, and is effective at long-context retrieval.

The issue-resolving task, where a model generates patches to fix real-world bugs, has emerged as a key benchmark for evaluating the capabilities of large language models (LLMs). While SWE-bench has become the dominant benchmark in this domain, it suffers from several limitations: it has not been updated since its release, is restricted to only 12 repositories, and relies heavily on manual effort for constructing test instances and setting up executable environments, significantly limiting its scalability. We present SWE-bench-Live, a live-updatable benchmark designed to address these limitations. SWE-bench-Live currently includes 1,890 tasks derived from real GitHub issues created since 2024, spanning 223 repositories. Each task is accompanied by a dedicated Docker image to ensure reproducible execution. Additionally, we introduce an automated curation pipeline that streamlines the entire process from instance creation to environment setup, removing manual bottlenecks and enabling scalability and continuous updates. We evaluate a range of state-of-the-art models and agent frameworks on SWE-bench-Live, offering detailed empirical insights into their real-world bug-fixing capabilities. By providing a fresh, diverse, and executable benchmark grounded in live repository activity, SWE-bench-Live supports reliable, large-scale assessment of code LLMs and code agents in realistic development settings.

Reinforcement learning (RL) post-training is crucial for LLM alignment and reasoning, but existing policy-based methods, such as PPO and DPO, can fall short of fixing shortcuts inherited from pre-training. In this work, we introduce $Q\sharp$, a value-based algorithm for KL-regularized RL that guides the reference policy using the optimal regularized $Q$ function. We propose to learn the optimal $Q$ function using distributional RL on an aggregated online dataset. Unlike prior value-based baselines that guide the model using unregularized $Q$-values, our method is theoretically principled and provably learns the optimal policy for the KL-regularized RL problem. Empirically, $Q\sharp$ outperforms prior baselines in math reasoning benchmarks while maintaining a smaller KL divergence to the reference policy. Theoretically, we establish a reduction from KL-regularized RL to no-regret online learning, providing the first bounds for deterministic MDPs under only realizability. Thanks to distributional RL, our bounds are also variance-dependent and converge faster when the reference policy has small variance. In sum, our results highlight $Q\sharp$ as an effective approach for post-training LLMs, offering both improved performance and theoretical guarantees. The code can be found at \url{https://github.com/jinpz/q_sharp}.

Large language models are typically adapted to downstream tasks through supervised fine-tuning on domain-specific data. While standard fine-tuning focuses on minimizing generation loss to optimize model parameters, we take a deeper step by retaining and leveraging the model’s own learning signals, analogous to how human learners reflect on past mistakes to improve future performance. We first introduce the concept of Mistake Log to systematically track the model’s learning behavior and recurring errors throughout fine-tuning. Treating the original transformer-based model as the Pilot, we correspondingly design a Copilot model to refine the Pilot’s inference performance via logits rectification. We name the overall Pilot-Copilot framework the Transformer Copilot, which introduces (i) a novel Copilot model design, (ii) a joint training paradigm where the Copilot continuously learns from the evolving Mistake Log alongside the Pilot, and (iii) a fused inference paradigm where the Copilot rectifies the Pilot’s logits for enhanced generation. We provide both theoretical and empirical analyses on our new learning framework. Experiments on 12 benchmarks spanning commonsense, arithmetic, and recommendation tasks demonstrate that Transformer Copilot consistently improves performance by up to 34.5%, while introducing marginal computational overhead to Pilot models and exhibiting strong scalability and transferability. Our code is released at https://github.com/jiaruzouu/TransformerCopilot.

Quantization plays a crucial role in accelerating the inference of large-scale models, and rotational matrices have been shown to effectively improve quantization performance by smoothing outliers. However, end-to-end fine-tuning of rotational optimization algorithms incurs high computational costs and is prone to overfitting. To address this challenge, we propose an efficient distribution-aware rotational calibration method, DartQuant, which reduces the complexity of rotational optimization by constraining the distribution of the activations after rotation. This approach also effectively reduces reliance on task-specific losses, thereby mitigating the risk of overfitting. Additionally, we introduce the QR-Orth optimization scheme, which replaces expensive alternating optimization with a more efficient solution. In a variety of model quantization experiments, DartQuant demonstrates superior performance. Compared to existing methods, it achieves 47$\times$ acceleration and 10$\times$ memory savings for rotational optimization on a 70B model. Furthermore, it is the first to successfully complete rotational calibration for a 70B model on a single 3090 GPU, making quantization of large language models feasible in resource-constrained environments.

We interact with computers on an everyday basis, be it in everyday life or work, and many aspects of work can be done entirely with access to a computer and the Internet. At the same time, thanks to improvements in large language models (LLMs), there has also been a rapid development in AI agents that interact with and affect change in their surrounding environments. But how performant are AI agents at helping to accelerate or even autonomously perform work-related tasks? The answer to this question has important implications for both industry looking to adopt AI into their workflows, and for economic policy to understand the effects that adoption of AI may have on the labor market. To measure the progress of these LLM agents' performance on performing real-world professional tasks, in this paper, we introduce TheAgentCompany, an extensible benchmark for evaluating AI agents that interact with the world in similar ways to those of a digital worker: by browsing the Web, writing code, running programs, and communicating with other coworkers. We build a self-contained environment with internal web sites and data that mimics a small software company environment, and create a variety of tasks that may be performed by workers in such a company. We test baseline agents powered by both closed API-based and open-weights language models (LMs), and find that with the most competitive agent, 30% of the tasks can be completed autonomously. This paints a nuanced picture on task automation with LM agents -- in a setting simulating a real workplace, a good portion of simpler tasks could be solved autonomously, but more difficult long-horizon tasks are still beyond the reach of current systems. For more information and demos, refer to https://the-agent-company.com.

Modern LLMs are increasingly deep, and depth correlates with performance, albeit with diminishing returns. However, do these models use their depth efficiently? Do they compose more features to create higher-order computations that are impossible in shallow models, or do they merely spread the same kinds of computation out over more layers? To address these questions, we analyze the residual stream of the Llama 3.1, Qwen 3, and OLMo 2 family of models. We find: First, comparing the output of the sublayers to the residual stream reveals that layers in the second half contribute much less than those in the first half, with a clear phase transition between the two halves. Second, skipping layers in the second half has a much smaller effect on future computations and output predictions. Third, for multihop tasks, we are unable to find evidence that models are using increased depth to compose subresults in examples involving many hops. Fourth, we seek to directly address whether deeper models are using their additional layers to perform new kinds of computation. To do this, we train linear maps from the residual stream of a shallow model to a deeper one. We find that layers with the same relative depth map best to each other, suggesting that the larger model simply spreads the same computations out over its many layers. All this evidence suggests that deeper models are not using their depth to learn new kinds of computation, but only using the greater depth to perform more fine-grained adjustments to the residual. This may help explain why increasing scale leads to diminishing returns for stacked Transformer architectures.