Training Introspective Behavior: Fine-Tuning Induces Reliable Internal State Detection in a 7B Model

TL;DR

This work demonstrates that a 7B parameter language model can be directly trained to detect and identify transient injected concepts at a single token, transforming prior near-total failure into robust performance. By injecting concept vectors at a final prompt token and training with diverse prompts and strengths, the model achieves up to 85% accuracy on held-out concepts with zero false positives, indicating strong grounding and internality signals. Generalization to unseen concept vectors suggests the model learns a transferable decoding skill rather than memorizing mappings, though it does not establish metacognitive representation as defined by Lindsey. The findings offer a pathway to built-in AI transparency via trainable introspection, while acknowledging limitations around metacognition, single-model scope, and artificial injection settings.

Abstract

Lindsey (2025) investigates introspective awareness in language models through four experiments, finding that models can sometimes detect and identify injected activation patterns -- but unreliably (~20% success in the best model). We focus on the first of these experiments -- self-report of injected "thoughts" -- and ask whether this capability can be directly trained rather than waiting for emergence. Through fine-tuning on transient single-token injections, we transform a 7B parameter model from near-complete failure (0.4% accuracy, 6.7% false positive rate) to reliable detection (85% accuracy on held-out concepts at α=40, 0% false positives). Our model detects fleeting "thoughts" injected at a single token position, retains that information, and reports the semantic content across subsequent generation steps. On this task, our trained model satisfies three of Lindsey's criteria: accuracy (correct identification), grounding (0/60 false positives), and internality (detection precedes verbalization). Generalization to unseen concept vectors (7.5pp gap) demonstrates the model learns a transferable skill rather than memorizing specific vectors, though this does not establish metacognitive representation in Lindsey's sense. These results address an open question raised by Lindsey: whether "training for introspection would help eliminate cross-model differences." We show that at least one component of introspective behavior can be directly induced, offering a pathway to built-in AI transparency.