About this repository — Interim Report & Zenodo publication

#1
by airVen - opened

This repository contains the research materials for "Between Hidden States and Control: Hidden-State Signals in Iterative LLM Repair" — an independent empirical investigation into whether minimal, non-learned signals derived from hidden states during inference can serve as an internal value function in LLMs.

Full report (citable, with DOI): https://doi.org/10.5281/zenodo.18941566

What's in this repo:

  • Between_Hidden_States_and_Control_Interim_Report.pdf — full 13-page interim report covering all 10 experimental phases
  • MVF_Supplementary_Materials.zip — experiment protocol, result files, core scripts

Core finding: Hidden-state cosine similarity at Layer 27 / Stride 50 detects semantic stagnation that text-based loop detectors miss. However, coherence alone is not sufficient as a standalone actuator — the problem has shifted from signal discovery to policy design.

If you're working on mechanistic interpretability, inference control, or agentic code systems, feedback and questions are very welcome here.

Sign up or log in to comment