Reproduce “Internalizing LLM Reasoning via Discovery and Replay of Latent Actions” on OpenResearch

About this paper

Zhenning Shi, Yijia Zhu, Junhan Shi, Xun Zhang, Lei Wang, Congcong Miao

The STIR framework, developed by researchers at Tsinghua University and the National University of Singapore, empowers Large Language Models to internalize complex reasoning by dynamically controlling latent representations. It achieves average accuracy improvements of 1.9% to 7.5% while reducing token consumption by up to 35% through the discovery and replay of self-distilled "latent actions."

Key insights

• Reasoning failures within LLMs are not random but organize into identifiable, coherent clusters in the latent manifold.
• Steering impulses derived from contrasting successful and failed reasoning trajectories act as precise translational vectors to guide latent states from erroneous to optimal regions.
• Dynamically modulated interventions, incorporating both a lightweight forward pass for 'lookahead' probing and an anchor-based gating mechanism, are crucial for effective and efficient latent state correction, preventing over-intervention.

Summary by alphaXiv.