Abstract:Deep reinforcement learning (RL) has achieved significant success, yet its application in real-world scenarios is often hindered by a lack of robustness to environmental uncertainties. To solve this challenge, some robust RL algorithms have been proposed, but most are limited to tabular settings. In this work, we propose Distributionally Robust Soft Actor-Critic (DR-SAC), a novel algorithm designed to enhance the robustness of the state-of-the-art Soft Actor-Critic (SAC) algorithm. DR-SAC aims to maximize the expected value with entropy against the worst possible transition model lying in an uncertainty set. A distributionally robust version of the soft policy iteration is derived with a convergence guarantee. For settings where nominal distributions are unknown, such as offline RL, a generative modeling approach is proposed to estimate the required nominal distributions from data. Furthermore, experimental results on a range of continuous control benchmark tasks demonstrate our algorithm achieves up to $9.8$ times the average reward of the SAC baseline under common perturbations. Additionally, compared with existing robust reinforcement learning algorithms, DR-SAC significantly improves computing efficiency and applicability to large-scale problems.
Abstract:Unlearning in large language models (LLMs) is becoming increasingly important due to regulatory compliance, copyright protection, and privacy concerns. However, a key challenge in LLM unlearning is unintended forgetting, where the removal of specific data inadvertently impairs the utility of the model and its retention of valuable, desired information. While prior work has primarily focused on architectural innovations, the influence of data-level factors on unlearning performance remains underexplored. As a result, existing methods often suffer from degraded retention when forgetting high-impact data. To address this, we propose GUARD-a novel framework for Guided Unlearning And Retention via Data attribution. At its core, GUARD introduces a lightweight proxy data attribution metric tailored for LLM unlearning, which quantifies the "alignment" between the forget and retain sets while remaining computationally efficient. Building on this, we design a novel unlearning objective that assigns adaptive, nonuniform unlearning weights to samples, inversely proportional to their proxy attribution scores. Through such a reallocation of unlearning power, GUARD mitigates unintended losses in retention. We provide rigorous theoretical guarantees that GUARD significantly enhances retention while maintaining forgetting metrics comparable to prior methods. Extensive experiments on the TOFU benchmark across multiple LLM architectures demonstrate that GUARD substantially improves utility preservation while ensuring effective unlearning. Notably, GUARD reduces utility sacrifice on the Retain Set by up to 194.92% in terms of Truth Ratio when forgetting 10% of the training data.
Abstract:Neural network verifiers based on linear bound propagation scale impressively to massive models but can be surprisingly loose when neuron coupling is crucial. Conversely, semidefinite programming (SDP) verifiers capture inter-neuron coupling naturally, but their cubic complexity restricts them to only small models. In this paper, we propose SDP-CROWN, a novel hybrid verification framework that combines the tightness of SDP relaxations with the scalability of bound-propagation verifiers. At the core of SDP-CROWN is a new linear bound, derived via SDP principles, that explicitly captures $\ell_{2}$-norm-based inter-neuron coupling while adding only one extra parameter per layer. This bound can be integrated seamlessly into any linear bound-propagation pipeline, preserving the inherent scalability of such methods yet significantly improving tightness. In theory, we prove that our inter-neuron bound can be up to a factor of $\sqrt{n}$ tighter than traditional per-neuron bounds. In practice, when incorporated into the state-of-the-art $\alpha$-CROWN verifier, we observe markedly improved verification performance on large models with up to 65 thousand neurons and 2.47 million parameters, achieving tightness that approaches that of costly SDP-based methods.
Abstract:Reinforcement learning (RL) has become an effective approach for fine-tuning large language models (LLMs), particularly to enhance their reasoning capabilities. However, RL fine-tuning remains highly resource-intensive, and existing work has largely overlooked the problem of data efficiency. In this paper, we propose two techniques to improve data efficiency in LLM RL fine-tuning: difficulty-targeted online data selection and rollout replay. We introduce the notion of adaptive difficulty to guide online data selection, prioritizing questions of moderate difficulty that are more likely to yield informative learning signals. To estimate adaptive difficulty efficiently, we develop an attention-based framework that requires rollouts for only a small reference set of questions. The adaptive difficulty of the remaining questions is then estimated based on their similarity to this set. To further reduce rollout cost, we introduce a rollout replay mechanism that reuses recent rollouts, lowering per-step computation while maintaining stable updates. Extensive experiments across 6 LLM-dataset combinations show that our method reduces RL fine-tuning time by 25% to 65% to reach the same level of performance as the original GRPO algorithm.
Abstract:This paper presents AlphaOne ($\alpha$1), a universal framework for modulating reasoning progress in large reasoning models (LRMs) at test time. $\alpha$1 first introduces $\alpha$ moment, which represents the scaled thinking phase with a universal parameter $\alpha$. Within this scaled pre-$\alpha$ moment phase, it dynamically schedules slow thinking transitions by modeling the insertion of reasoning transition tokens as a Bernoulli stochastic process. After the $\alpha$ moment, $\alpha$1 deterministically terminates slow thinking with the end-of-thinking token, thereby fostering fast reasoning and efficient answer generation. This approach unifies and generalizes existing monotonic scaling methods by enabling flexible and dense slow-to-fast reasoning modulation. Extensive empirical studies on various challenging benchmarks across mathematical, coding, and scientific domains demonstrate $\alpha$1's superior reasoning capability and efficiency. Project page: https://alphaone-project.github.io/
Abstract:Continual Learning with Pre-trained Models holds great promise for efficient adaptation across sequential tasks. However, most existing approaches freeze PTMs and rely on auxiliary modules like prompts or adapters, limiting model plasticity and leading to suboptimal generalization when facing significant distribution shifts. While full fine-tuning can improve adaptability, it risks disrupting crucial pre-trained knowledge. In this paper, we propose Mutual Information-guided Sparse Tuning (MIST), a plug-and-play method that selectively updates a small subset of PTM parameters, less than 5%, based on sensitivity to mutual information objectives. MIST enables effective task-specific adaptation while preserving generalization. To further reduce interference, we introduce strong sparsity regularization by randomly dropping gradients during tuning, resulting in fewer than 0.5% of parameters being updated per step. Applied before standard freeze-based methods, MIST consistently boosts performance across diverse continual learning benchmarks. Experiments show that integrating our method into multiple baselines yields significant performance gains. Our code is available at https://github.com/zhwhu/MIST.
Abstract:Existing benchmarks that assess Language Models (LMs) as Language Agents (LAs) for tool use primarily focus on stateless, single-turn interactions or partial evaluations, such as tool selection in a single turn, overlooking the inherent stateful nature of interactions in multi-turn applications. To fulfill this gap, we propose \texttt{DialogTool}, a multi-turn dialogue dataset with stateful tool interactions considering the whole life cycle of tool use, across six key tasks in three stages: 1) \textit{tool creation}; 2) \textit{tool utilization}: tool awareness, tool selection, tool execution; and 3) \textit{role-consistent response}: response generation and role play. Furthermore, we build \texttt{VirtualMobile} -- an embodied virtual mobile evaluation environment to simulate API calls and assess the robustness of the created APIs\footnote{We will use tools and APIs alternatively, there are no significant differences between them in this paper.}. Taking advantage of these artifacts, we conduct comprehensive evaluation on 13 distinct open- and closed-source LLMs and provide detailed analysis at each stage, revealing that the existing state-of-the-art LLMs still cannot perform well to use tools over long horizons.
Abstract:Recent advances in uncertainty estimation for Large Language Models (LLMs) during downstream adaptation have addressed key challenges of reliability and simplicity. However, existing Bayesian methods typically require multiple sampling iterations during inference, creating significant efficiency issues that limit practical deployment. In this paper, we investigate the possibility of eliminating the need for test-time sampling for LLM uncertainty estimation. Specifically, when given an off-the-shelf Bayesian LLM, we distill its aligned confidence into a non-Bayesian student LLM by minimizing the divergence between their predictive distributions. Unlike typical calibration methods, our distillation is carried out solely on the training dataset without the need of an additional validation dataset. This simple yet effective approach achieves N-times more efficient uncertainty estimation during testing, where N is the number of samples traditionally required by Bayesian LLMs. Our extensive experiments demonstrate that uncertainty estimation capabilities on training data can successfully generalize to unseen test data through our distillation technique, consistently producing results comparable to (or even better than) state-of-the-art Bayesian LLMs.
Abstract:While Large Language Models (LLMs) have demonstrated impressive capabilities, their output quality remains inconsistent across various application scenarios, making it difficult to identify trustworthy responses, especially in complex tasks requiring multi-step reasoning. In this paper, we propose a token-level uncertainty estimation framework to enable LLMs to self-assess and self-improve their generation quality in mathematical reasoning. Specifically, we introduce low-rank random weight perturbation to LLM decoding, generating predictive distributions that we use to estimate token-level uncertainties. We then aggregate these uncertainties to reflect semantic uncertainty of the generated sequences. Experiments on mathematical reasoning datasets of varying difficulty demonstrate that our token-level uncertainty metrics strongly correlate with answer correctness and model robustness. Additionally, we explore using uncertainty to directly enhance the model's reasoning performance through multiple generations and the particle filtering algorithm. Our approach consistently outperforms existing uncertainty estimation methods, establishing effective uncertainty estimation as a valuable tool for both evaluating and improving reasoning generation in LLMs.
Abstract:Evaluating generative models remains a fundamental challenge, particularly when the goal is to reflect human preferences. In this paper, we use music generation as a case study to investigate the gap between automatic evaluation metrics and human preferences. We conduct comparative experiments across five state-of-the-art music generation approaches, assessing both perceptual quality and distributional similarity to human-composed music. Specifically, we evaluate synthesis music from various perceptual dimensions and examine reference-based metrics such as Mauve Audio Divergence (MAD) and Kernel Audio Distance (KAD). Our findings reveal significant inconsistencies across the different metrics, highlighting the limitation of the current evaluation practice. To support further research, we release a benchmark dataset comprising samples from multiple models. This study provides a broader perspective on the alignment of human preference in generative modeling, advocating for more human-centered evaluation strategies across domains.