Research

My research focuses on self-improving AI agents, particularly how agents can learn, accumulate knowledge, and continuously improve their effectiveness and efficiency over time. Previously, I studied model interpretability.

Papers
  1. 2026

    The Generalization Ridge: Information Flow in Natural Language Generation

    Ruidi Chang, Chunyuan Deng, Hanjie Chen

    COLM 2026

    Models are not monotonically better with depth. InfoRidge traces predictive information across layers: it peaks in the upper-middle layers, a "ridge" where generalization is strongest, then drops in the final layers. Under distribution shift, models down-weight deep layers and rely more on the intermediate ones.

    InfoRidge: predictive information across layers peaks at an intermediate ridge
  2. 2026

    PRISM: A Dual View of LLM Reasoning through Semantic Flow and Latent Computation

    Ruidi Chang, Jiawei Zhou, Hanjie Chen

    arXiv preprint

    Reasoning traces are usually studied either as text across steps or as hidden states across layers. PRISM models both at once: an explicit stage labels each step's semantic role (setup, computation, verification, answer) and models how roles flow, and an implicit stage finds latent computational regimes inside each step, linked across steps by a bridge matrix. Failed trajectories are more likely to get trapped in unproductive verification loops and split into distinct modes such as overthinking and premature commitment.

    PRISM: explicit semantic flow between reasoning steps and implicit computational regimes within each step, linked by a cross-step bridge
  3. 2025

    Steering Information Utility in Key-Value Memory for Language Model Post-Training

    Chunyuan Deng, Ruidi Chang, Hanjie Chen

    NeurIPS 2025

  4. 2025

    Learning Distribution-wise Control in Representation Space for Language Models

    Chunyuan Deng, Ruidi Chang, Hanjie Chen

    ICML 2025

  5. 2025

    SAFR: Neuron Redistribution for Interpretability

    Ruidi Chang, Chunyuan Deng, Hanjie Chen

    NAACL 2025 Findings

    Neurons often mix many features at once (superposition). SAFR redistributes them: monosemantic for important tokens, polysemantic where interactions matter, which makes the model more interpretable.

    Workflow of SAFR neuron redistribution
  6. 2026

    Language Models are Symbolic Learners in Arithmetic

    Chunyuan Deng, Zhiqi Li, Roy Xie, Ruidi Chang, Hanjie Chen

    TMLR 2026

  7. 2024

    Large Language Model Based Multi-Agents: A Survey of Progress and Challenges

    Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V. Chawla, Olaf Wiest, Xiangliang Zhang

    IJCAI 2024

Service
Reviewer
ACL ARR 2025 Feb, IEEE BigData 2025, EMNLP BlackboxNLP Workshop 2025, COLM XLLM-Reason-Plan Workshop 2025, COLING 2025
Volunteer
EMNLP BlackboxNLP Workshop 2024
Mentoring
Nursultan Asilbekov (SURF Program)