Linear probes on frozen LLM activations nearly match fine-tuned ESG classifiers

An arXiv preprint argues that concept content in text can be measured from frozen large language model activations rather than only from surface word statistics or the model's written answers. Existing measures score dictionary word shares, topic proportions, or embedding similarities and therefore read the words a text uses rather than the judgment a reader forms, the authors write. Prior work has shown a gap between what LLMs know internally and what they express in responses. The paper tests whether internal knowledge, read by monitoring activations of out-of-the-box frozen models, can stand in for task-specific fine-tuning, and which extraction method works best. The authors extract measures via the Recursive Feature Machine algorithm and via linear probing, then compare those against an embedding baseline, surface baselines, and the same model's own answer to the question. They demonstrate the approach on financial text using a human-annotated Environmental, Social and Governance dataset. The best linear probe comes within 0.6 percentage points of a fine-tuned domain classifier's accuracy without any task-specific fine-tuning, and outscores the same model's own answer in eleven of twelve comparisons, so activations carry concept content the response does not report. Simple probes consistently beat RFM concept vectors, which instead supply a continuous score of how strongly a concept is present, pending validation on graded labels.