Representation Drift Under Self-Reflection: Does Self-Critique Reshape Internal States?

LLMs are increasingly prompted to “think again,” critique themselves, and revise their answers. While self-reflection often improves performance, we do not know whether this improvement corresponds to meaningful internal representational change or merely shallow re-sampling.

This project investigates whether self-critique induces structured shifts in a model’s internal representations, particularly in the residual stream.

Core Question:
When an LLM generates an initial answer, critiques it, and then revises it, do its internal activations move to a distinct and more structured reasoning subspace?

Experimental Design:

1. Prompt model on reasoning tasks (math, logic, multi-step QA).

2. Capture residual stream activations during:

3. Initial answer

4. Self-critique

5. Revised answer

6. Compare representations using:

   i) Cosine similarity drift
   ii) PCA/subspace analysis
   iii) Linear probes (correct vs incorrect reasoning)
   iv) Robustness to activation steering

Hypotheses:
If reflection is meaningful, we expect structured representation drift, improved linear separability of correct reasoning, and convergence toward a more stable reasoning manifold.
If reflection is shallow, activations should remain near the original neighborhood with minimal structural change.

Why this is innovative:

1. Connects prompting techniques with mechanistic interpretability.
2. Provides a measurable notion of “cognitive change” in LLMs.
3. Directly informs recursive self-improvement agents and AI research automation.
4. Bridges performance gains with internal dynamics.
5. Extensions could compare self-critique vs human critique, single vs multi-round reflection, and small vs large models.