Gouki Minegishi

Mechanism of Task-oriented Information Removal in In-context Learning

RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs

Emergent Misalignment from Superposition

Topology of Reasoning: Understanding Large Reasoning Models through Reasoning Graph Properties

Beyond Induction Heads: In-Context Meta Learning Induces Multi-Phase Circuit Emergence

Rethinking Evaluation of Sparse Autoencoders through the Representation of Polysemous Words

Bridging Lottery Ticket and Grokking: Understanding Grokking from Inner Structure of Networks