airVen's picture
Update README.md
2fe88f5 verified
metadata
license: cc-by-4.0
language:
  - en
tags:
  - llm
  - interpretability
  - inference-control
  - hidden-states
  - repair-loop
  - code-generation
  - mechanistic-interpretability
  - alignment
pretty_name: The Missing Value Function  Interim Research Report

Between Hidden States and Control: Hidden-State Signals in Iterative LLM Repair

The Missing Value Function — Interim Research Report

Can minimal, non-learned signals derived from hidden states during inference serve as an internal value function to distinguish productive from unproductive revisitation in large language models?

Links

Cite as: Weise, B. (2026). The Missing Value Function: A Preliminary Report on Hidden-State Signals in Iterative LLM Repair. Zenodo. https://doi.org/10.5281/zenodo.18941566


Overview

This repository contains the interim research report and supplementary materials for the project "The Missing Value Function", an independent empirical investigation into whether biological valence signal principles (Damasio's Somatic Marker Hypothesis, Sutskever's emotion-as-value-function framing) can be operationalized as lightweight inference-time control signals in transformer-based LLMs.

Author: Benjamin Weise (Independent Research / Prooftrail) Date: March 10, 2026 Version: 1.0 License: CC BY 4.0


Key Findings

  • Signal Discovery: Hidden-state cosine similarity at Layer 27, Stride 50 detects semantic stagnation that text-based loop detectors (n-gram, codeblock) miss entirely — two reproducible dissociation cases
  • Negative Boundary: Simple prompt-based and sampling-based actuators showed no robust improvement over baseline (Phase 10.3, 10.4)
  • Ambiguity of Coherence: High coherence values mark both productive convergence and unproductive stagnation — coherence alone is insufficient as a standalone actuator
  • Multi-Signal Direction: entropy + margin combination shows modest improvement for regression detection (AUC 0.59)
  • Monotonic Controller: Boundary result — preservation alone does not solve the bottleneck; productive diversity is the missing ingredient
  • Repair Loop Testbed: frontier_02_hard (LRU Cache, 7 test blocks) achieves 37.5% baseline success — the right difficulty corridor for hypothesis testing

Current Status: The evidence supports that hidden-state signals are diagnostically valuable but not yet sufficient as standalone actuators. The research has identified real signal dissociation, established negative boundaries for simple interventions, and motivated a shift toward multi-signal policy design.


Repository Contents

File Description
Between_Hidden_States_and_Control_Interim_Report.pdf Full interim research report (10 phases, all findings)
MVF_Supplementary_Materials.zip Experiment protocol, result files, core scripts

Experimental Setup

Component Value
Model Qwen/Qwen2.5-7B-Instruct (4-bit quantized)
GPU NVIDIA GeForce RTX 5070 (11.9 GB VRAM)
Monitor Layer 27 (96% depth)
Checkpoint Stride 50 tokens
Primary Metric max_prev_similarity (cosine similarity)
Primary Task frontier_02_hard — LRU Cache, 7 test blocks

Citation

Weise, B. (2026). The Missing Value Function: A Preliminary Report on Hidden-State Signals
in Iterative LLM Repair. Zenodo. https://doi.org/10.5281/zenodo.18941566

Related Work

  • Damasio, A. (1996). Somatic Marker Hypothesis
  • Sutskever, I. (2025). Emotions as evolutionarily hardcoded value functions (Dwarkesh Patel interview)
  • Pathak et al. (2017). Curiosity-driven Exploration (ICM)
  • Bengio et al. (2021). Inductive Biases for Deep Learning

This is an interim report. Negative results are documented as completed steps. The project is ongoing.