Step-Level Sparse Autoencoder for Reasoning Process Interpretation

Large Language Models (LLMs) have achieved strong complex reasoning capabilities through Chain-of-Thought (CoT) reasoning. However, their reasoning patterns remain too complicated to analyze. While Sparse Autoencoders (SAEs) have emerged as a powerful tool for interpretability, existing approaches predominantly operate at the token level, creating a granularity mismatch when capturing more critical step-level information, such as reasoning direction and semantic transitions. In this work, we propose step-level sparse autoencoder (SSAE), which serves as an analytical tool to disentangle different aspects of LLMs' reasoning steps into sparse features. Specifically, by precisely controlling the sparsity of a step feature conditioned on its context, we form an information bottleneck in step reconstruction, which splits incremental information from background information and disentangles it into several sparsely activated dimensions. Experiments on multiple base models and reasoning tasks show the effectiveness of the extracted features. By linear probing, we can easily predict surface-level information, such as generation length and first token distribution, as well as more complicated properties, such as the correctness and logicality of the step. These observations indicate that LLMs should already at least partly know about these properties during generation, which provides the foundation for the self-verification ability of LLMs. The code is available at https://github.com/Miaow-Lab/SSAE

SkillCraft: Can LLM Agents Learn to Use Tools Skillfully?

Real-world tool-using agents operate over long-horizon workflows with recurring structure and diverse demands, where effective behavior requires not only invoking atomic tools but also abstracting, and reusing higher-level tool compositions. However, existing benchmarks mainly measure instance-level success under static tool sets, offering limited insight into agents' ability to acquire such reusable skills. We address this gap by introducing SkillCraft, a benchmark explicitly stress-test agent ability to form and reuse higher-level tool compositions, where we call Skills. SkillCraft features realistic, highly compositional tool-use scenarios with difficulty scaled along both quantitative and structural dimensions, designed to elicit skill abstraction and cross-task reuse. We further propose a lightweight evaluation protocol that enables agents to auto-compose atomic tools into executable Skills, cache and reuse them inside and across tasks, thereby improving efficiency while accumulating a persistent library of reusable skills. Evaluating state-of-the-art agents on SkillCraft, we observe substantial efficiency gains, with token usage reduced by up to 80% by skill saving and reuse. Moreover, success rate strongly correlates with tool composition ability at test time, underscoring compositional skill acquisition as a core capability.

Not All Steps are Informative: On the Linearity of LLMs' RLVR Training

Reinforcement learning with verifiable rewards (RLVR) has become a central component of large language model (LLM) post-training. Unlike supervised fine-tuning (SFT), RLVR lets an LLM generate multiple candidate solutions and reinforces those that lead to a verifiably correct final answer. However, in practice, RLVR often requires thousands of training steps to reach strong performance, incurring substantial computation largely attributed to prolonged exploration. In this work, we make a surprising observation: during RLVR, LLMs evolve in a strongly linear manner. Specifically, both model weights and model output log-probabilities exhibit strong linear correlations with RL training steps. This suggests that RLVR predominantly amplifies trends that emerge early in training, rather than continuously discovering new behaviors throughout the entire optimization trajectory. Motivated by this linearity, we investigate whether future model states can be predicted from intermediate checkpoints via extrapolation, avoiding continued expensive training. We show that Weight Extrapolation produces models with performance comparable to standard RL training while requiring significantly less computation. Moreover, Logits Extrapolation consistently outperforms continued RL training on all four benchmarks by extrapolating beyond the step range where RL training remains stable.

Enhancing Large Language Model Reasoning with Reward Models: An Analytical Survey

Reward models (RMs) play a critical role in enhancing the reasoning performance of LLMs. For example, they can provide training signals to finetune LLMs during reinforcement learning (RL) and help select the best answer from multiple candidates during inference. In this paper, we provide a systematic introduction to RMs, along with a comprehensive survey of their applications in LLM reasoning. We first review fundamental concepts of RMs, including their architectures, training methodologies, and evaluation techniques. Then, we explore their key applications: (1) guiding generation and selecting optimal outputs during LLM inference, (2) facilitating data synthesis and iterative self-improvement for LLMs, and (3) providing training signals in RL-based finetuning. Finally, we address critical open questions regarding the selection, generalization, evaluation, and enhancement of RMs, based on existing research and our own empirical findings. Our analysis aims to provide actionable insights for the effective deployment and advancement of RMs for LLM reasoning.

BED-LLM: Intelligent Information Gathering with LLMs and Bayesian Experimental Design

We propose a general-purpose approach for improving the ability of Large Language Models (LLMs) to intelligently and adaptively gather information from a user or other external source using the framework of sequential Bayesian experimental design (BED). This enables LLMs to act as effective multi-turn conversational agents and interactively interface with external environments. Our approach, which we call BED-LLM (Bayesian Experimental Design with Large Language Models), is based on iteratively choosing questions or queries that maximize the expected information gain (EIG) about the task of interest given the responses gathered previously. We show how this EIG can be formulated in a principled way using a probabilistic model derived from the LLM's belief distribution and provide detailed insights into key decisions in its construction. Further key to the success of BED-LLM are a number of specific innovations, such as a carefully designed estimator for the EIG, not solely relying on in-context updates for conditioning on previous responses, and a targeted strategy for proposing candidate queries. We find that BED-LLM achieves substantial gains in performance across a wide range of tests based on the 20-questions game and using the LLM to actively infer user preferences, compared to direct prompting of the LLM and other adaptive design strategies.

SelfCheck: Using LLMs to Zero-Shot Check Their Own Step-by-Step Reasoning

The recent progress in large language models (LLMs), especially the invention of chain-of-thoughts (CoT) prompting, makes it possible to solve reasoning problems. However, even the strongest LLMs are still struggling with more complicated problems that require non-linear thinking and multi-step reasoning. In this work, we explore whether LLMs have the ability to recognize their own errors, without resorting to external resources. In particular, we investigate whether they can be used to identify individual errors within a step-by-step reasoning. To this end, we propose a zero-shot verification scheme to recognize such errors. We then use this verification scheme to improve question-answering performance, by using it to perform weighted voting on different generated answers. We test the method on three math datasets-GSM8K, MathQA, and MATH-and find that it successfully recognizes errors and, in turn, increases final predictive performance.

Side Channel-Assisted Inference Leakage from Machine Learning-based ECG Classification

The Electrocardiogram (ECG) measures the electrical cardiac activity generated by the heart to detect abnormal heartbeat and heart attack. However, the irregular occurrence of the abnormalities demands continuous monitoring of heartbeats. Machine learning techniques are leveraged to automate the task to reduce labor work needed during monitoring. In recent years, many companies have launched products with ECG monitoring and irregular heartbeat alert. Among all classification algorithms, the time series-based algorithm dynamic time warping (DTW) is widely adopted to undertake the ECG classification task. Though progress has been achieved, the DTW-based ECG classification also brings a new attacking vector of leaking the patients' diagnosis results. This paper shows that the ECG input samples' labels can be stolen via a side-channel attack, Flush+Reload. In particular, we first identify the vulnerability of DTW for ECG classification, i.e., the correlation between warping path choice and prediction results. Then we implement an attack that leverages Flush+Reload to monitor the warping path selection with known ECG data and then build a predictor for constructing the relation between warping path selection and labels of input ECG samples. Based on experiments, we find that the Flush+Reload-based inference leakage can achieve an 84.0\% attacking success rate to identify the labels of the two samples in DTW.

Learning Instance-Specific Data Augmentations

Existing data augmentation methods typically assume independence between transformations and inputs: they use the same transformation distribution for all input instances. We explain why this can be problematic and propose InstaAug, a method for automatically learning input-specific augmentations from data. This is achieved by introducing an augmentation module that maps an input to a distribution over transformations. This is simultaneously trained alongside the base model in a fully end-to-end manner using only the training data. We empirically demonstrate that InstaAug learns meaningful augmentations for a wide range of transformation classes, which in turn provides better performance on supervised and self-supervised tasks compared with augmentations that assume input--transformation independence.

InteL-VAEs: Adding Inductive Biases to Variational Auto-Encoders via Intermediary Latents

We introduce a simple and effective method for learning VAEs with controllable inductive biases by using an intermediary set of latent variables. This allows us to overcome the limitations of the standard Gaussian prior assumption. In particular, it allows us to impose desired properties like sparsity or clustering on learned representations, and incorporate prior information into the learned model. Our approach, which we refer to as the Intermediary Latent Space VAE (InteL-VAE), is based around controlling the stochasticity of the encoding process with the intermediary latent variables, before deterministically mapping them forward to our target latent representation, from which reconstruction is performed. This allows us to maintain all the advantages of the traditional VAE framework, while incorporating desired prior information, inductive biases, and even topological information through the latent mapping. We show that this, in turn, allows InteL-VAEs to learn both better generative models and representations.

Generating Fluent Adversarial Examples for Natural Languages

Efficiently building an adversarial attacker for natural language processing (NLP) tasks is a real challenge. Firstly, as the sentence space is discrete, it is difficult to make small perturbations along the direction of gradients. Secondly, the fluency of the generated examples cannot be guaranteed. In this paper, we propose MHA, which addresses both problems by performing Metropolis-Hastings sampling, whose proposal is designed with the guidance of gradients. Experiments on IMDB and SNLI show that our proposed MHA outperforms the baseline model on attacking capability. Adversarial training with MAH also leads to better robustness and performance.