Evaluating Introspection of Self-reflective R-Lens Feedback Training
Evan Coats, Heather Broome
We present an evaluation of model introspection that scores each self-report as two numbers rather than one: how well a model tells the prompts it answers consistently from those where it varies, and how readily it claims consistency at all. A single stated-minus-actual gap cannot separate a model that cannot tell from one that can tell but misreports, and because over-claims and under-claims cancel when averaged, a gap can sit near zero while most individual reports are wrong. The decomposition, with a polarity-flipped rewording, a cross-model control, a serving check and a comparison against an internal readout, lets particular non-introspective reports be located rather than summarized away.
We applied it to a novel self-reflective reinforcement learning technique that rewards a Qwen3.5-4B model for verbalizing its own R-lens readouts, testing six checkpoints and five off-the-shelf models on 48 prompts with resampled ground truth. The evaluation identified three kinds of cases where the trained model's report was not tracking its own behavior.
This is a really cool idea, and I would have liked to see a lot more detail in the write-up to properly appreciate it. For the R-lens training, it would have been nice to see more of the effects on the model generally. For the evaluation of introspection, more detail and particular examples would have made this much easier to follow. Concrete examples, full dataset documentation, and a sketch of the architecture would all help.
Binary probes of consistency may also be too low-information and too dependent on the specific input — they might not carry enough signal to be informative about introspection in particular. Whether or not an output is likely to be consistent across multiple samples strongly correlates with the distribution the input is sampled from. Given this, models might learn to pick up on that signal rather than actually introspecting on the output.
Consider being asked to name a random color five times versus naming the capital of Australia five times. It doesn't take introspection to know you'd be a lot more consistent on the second one—this confound would need to be addressed.
This project attempts to create a new method of measuring introspective capacity in LLMs. However, the project does not engage with the extensive existing literature on how introspective capacity or metacognition is quantified in any non-LLM system (like humans). Therefore, it claims to create a new metric when the metric is instead derivative of or even identical to existing and mature approaches. I suggest the authors read work by Fleming, Lau, Peters, Maniscalco, Rahnev, Mamassian, Desender, Steyvers, and many others who have been working on these questions -- in humans and increasingly now in artificial systems -- for a very long time.
Cite this work
@misc {
title={
(HckPrj) Evaluating Introspection of Self-reflective R-Lens Feedback Training
},
author={
Evan Coats, Heather Broome
},
date={
},
organization={Apart Research},
note={Research submission to the research sprint hosted by Apart.},
howpublished={https://apartresearch.com}
}
