Qikai Chang

Qikai Chang, Zhenrong Zhang, Linbo Chen et al.

Large Language Models (LLMs) have shown promise as educational tutors, yet effective tutoring requires more than solving problems: it must provide progressive Socratic guidance and balance multiple pedagogical objectives across multi-turn interactions. However, training such tutors remains challenging due to limited-fidelity and weakly controllable student simulation, under-specified pedagogical reward modeling, and unstable multi-objective optimization. To overcome these limitations, we propose PEARL, a pedagogically aligned reinforcement learning framework for training Socratic tutoring agents, consisting of three key components. First, we introduce a controllable student simulator that decouples latent cognitive states from response generation to model diverse abilities and misconceptions. Second, we develop a generative reward model that jointly evaluates pedagogical quality and objective correctness for policy optimization. Finally, we propose a stable multi-objective RL scheme that discretizes rewards within each dimension and aggregates normalized advantages across dimensions, preventing high-variance objectives from dominating updates. Experiments on multiple benchmarks show that PEARL achieves the best performance among open-source models and remains competitive with leading proprietary LLMs, despite using only a 30B policy model.

Fei Wu, Zhenrong Zhang, Qikai Chang et al.

Reinforcement Learning with Verifiable Rewards (RLVR) elicits long chain-of-thought reasoning in large language models (LLMs), but outcome-based rewards lead to coarse-grained advantage estimation. While existing approaches improve RLVR via token-level entropy or sequence-level length control, they lack a semantically grounded, step-level measure of reasoning progress. As a result, LLMs fail to distinguish necessary deduction from redundant verification: they may continue checking after reaching a correct solution and, in extreme cases, overturn a correct trajectory into an incorrect final answer. To remedy the lack of process supervision, we introduce a training-free probing mechanism that extracts intermediate confidence and correctness and combines them into a Step Potential signal that explicitly estimates the reasoning state at each step. Building on this signal, we propose Step Potential Advantage Estimation (SPAE), a fine-grained credit assignment method that amplifies potential gains, penalizes potential drops, and applies penalty after potential saturates to encourage timely termination. Experiments across multiple benchmarks show SPAE consistently improves accuracy while substantially reducing response length, outperforming strong RL baselines and recent efficient reasoning and token-level advantage estimation methods. The code is available at https://github.com/cii030/SPAE-RL.

Qikai Chang, Mingjun Chen, Changpeng Pi et al.

The primary objective of Optical Chemical Structure Recognition is to identify chemical structure images into corresponding markup sequences. However, the complex two-dimensional structures of molecules, particularly those with rings and multiple branches, present significant challenges for current end-to-end methods to learn one-dimensional markup directly. To overcome this limitation, we propose a novel Ring-Free Language (RFL), which utilizes a divide-and-conquer strategy to describe chemical structures in a hierarchical form. RFL allows complex molecular structures to be decomposed into multiple parts, ensuring both uniqueness and conciseness while enhancing readability. This approach significantly reduces the learning difficulty for recognition models. Leveraging RFL, we propose a universal Molecular Skeleton Decoder (MSD), which comprises a skeleton generation module that progressively predicts the molecular skeleton and individual rings, along with a branch classification module for predicting branch information. Experimental results demonstrate that the proposed RFL and MSD can be applied to various mainstream methods, achieving superior performance compared to state-of-the-art approaches in both printed and handwritten scenarios. The code is available at https://github.com/JingMog/RFL-MSD.

Qikai Chang, Zhenrong Zhang, Pengfei Hu et al.

Large Language Models (LLMs) have made remarkable progress in mathematical reasoning, but still continue to struggle with high-precision tasks like numerical computation and formal symbolic manipulation. Integrating external tools has emerged as a promising approach to bridge this gap. Despite recent advances, existing methods struggle with three key challenges: constructing tool-integrated reasoning data, performing fine-grained optimization, and enhancing inference. To overcome these limitations, we propose THOR (Tool-Integrated Hierarchical Optimization via RL). First, we introduce TIRGen, a multi-agent actor-critic-based pipeline for constructing high-quality datasets of tool-integrated reasoning paths, aligning with the policy and generalizing well across diverse models. Second, to perform fine-grained hierarchical optimization, we introduce an RL strategy that jointly optimizes for both episode-level problem solving and step-level code generation. This is motivated by our key insight that the success of an intermediate tool call is a strong predictor of the final answer's correctness. Finally, THOR incorporates a self-correction mechanism that leverages immediate tool feedback to dynamically revise erroneous reasoning paths during inference. Our approach demonstrates strong generalization across diverse models, performing effectively in both reasoning and non-reasoning models. It further achieves state-of-the-art performance for models of a similar scale on multiple mathematical benchmarks, while also delivering consistent improvements on code benchmarks. Our code will be publicly available at https://github.com/JingMog/THOR.

Mobai Xue, Jun Du, Zhenrong Zhang et al.

Automatic font generation remains a challenging research issue, primarily due to the vast number of Chinese characters, each with unique and intricate structures. Our investigation of previous studies reveals inherent bias capable of causing structural changes in characters. Specifically, when generating a Chinese character similar to, but different from, those in the training samples, the bias is prone to either correcting or ignoring these subtle variations. To address this concern, we propose a novel Skeleton and Font Generation Network (SFGN) to achieve a more robust Chinese character font generation. Our approach includes a skeleton builder and font generator. The skeleton builder synthesizes content features using low-resource text input, enabling our technique to realize font generation independently of content image inputs. Unlike previous font generation methods that treat font style as a global embedding, we introduce a font generator to align content and style features on the radical level, which is a brand-new perspective for font generation. Except for common characters, we also conduct experiments on misspelled characters, a substantial portion of which slightly differs from the common ones. Our approach visually demonstrates the efficacy of generated images and outperforms current state-of-the-art font generation methods. Moreover, we believe that misspelled character generation have significant pedagogical implications and verify such supposition through experiments. We used generated misspelled characters as data augmentation in Chinese character error correction tasks, simulating the scenario where students learn handwritten Chinese characters with the help of misspelled characters. The significantly improved performance of error correction tasks demonstrates the effectiveness of our proposed approach and the value of misspelled character generation.