English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

LLMs' 'Subconscious': DeepMind Finds AI Already Knows Its Confidence Before Saying 'I'm Not Sure'

Forum topic · 小凯 · 2026-06-07

Summary

A Google DeepMind study (Conmy et al., "How do LLMs Compute Verbal Confidence?", arXiv:2603.17839) reveals the neural mechanism behind verbal confidence in large language models using mechanistic interpretability techniques. Through activation steering, patching, noise ablation, and attention blocking on Gemma 3 27B, Qwen 2.5 7B, and Magistral Small 24B, the team shows that confidence is not computed on demand when asked, but is automatically formed and cached at the post-answer newline token (PANL), then retrieved at the confidence colon (CC) position when the model outputs a confidence class. Linear probing at PANL predicts verbal confidence with R² of 0.38, uniquely explaining variance beyond token probability, indicating deeper answer-quality evaluation rather than a simple fluency readout. Both tested models follow the same cache-then-retrieve mechanism. The finding enables pre-generation hallucination warning: by monitoring PANL activations, systems could route uncertain queries to retrieval or stronger models, intervene for self-correction, or improve calibration. Limitations include dependence on staged prompts, limited model scale, and the distinction between mechanism and calibration accuracy.

Google DeepMind's mechanistic interpretability study ("How do LLMs Compute Verbal Confidence?", Conmy et al.) reveals that LLM confidence is computed via cached retrieval, not just-in-time computation: models automatically form and cache a confidence representation right after generating an answer, at the post-answer newline token (PANL), and retrieve it when later asked to self-assess.

Key points

  • Experimental design: A two-phase prompt where the model first answers a question (e.g., TriviaQA), then classifies its own confidence ("Highly likely" / "Almost certain"...). Two positions matter: PANL (post-answer newline) and CC (confidence colon, where the model outputs its class).
  • Causal evidence: Four interventions establish the causal chain:
  • *Activation steering* at PANL (most effective at layers 21–25 in Gemma 3 27B) strongly shifts confidence output.
  • *Activation patching* transplants confidence between trials via PANL activations.
  • *Mean ablation* of PANL or CC severely disrupts confidence output.
  • *Attention blocking* shows information flows answer tokens → PANL → CC; blocking A→PANL or PANL→CC hurts most.
  • Not just fluency: Linear probing at PANL predicts verbal confidence with R² ≈ 0.38 (peak around layer ~40), while token log-probability explains only ~0.10–0.15 of variance. After controlling for token probability, PANL uniquely retains R² = 0.38 — the cached representation encodes deeper answer-quality evaluation than mere token likelihoods.
  • Generalization: Both Gemma 3 27B (62 layers, ECE 0.12, AUROC 0.71) and Qwen 2.5 7B (28 layers, ECE 0.06, AUROC 0.65) follow the same PANL-cache → CC-retrieval mechanism. Steering at CC works at deeper layers (30–35 in Gemma); steering at PANL+1, the first confidence colon token, or the first answer token has no effect.
  • Why the newline?

    PANL is a semantically "empty" token — the transition point after answering and before a new task. Steering at the last answer token also works but is noisier (it carries answer content); the newline serves as a clean slot caching pure meta-information about answer quality, analogous to humans' brief unconscious self-assessment after finishing a sentence.

    Implications for hallucination detection

    The study enables a third path beyond post-hoc and token-by-token detection: pre-generation warning —

    1. LLM routing: train lightweight probes (3M–37M parameters) on intermediate hidden states to route low-confidence queries to RAG or stronger models. 2. Real-time intervention: trigger self-correction or tool verification when PANL activations signal low confidence. 3. Calibration training: target PANL representations directly (temperature scaling, Platt scaling). 4. Safety alignment: inject conservatism signals at PANL during RLHF to encourage abstention under uncertainty.

    Caveats

  • The mechanism was tested only with staged prompts; its role in free-form or streaming dialogue is unclear.
  • Only Gemma 3 27B, Qwen 2.5 7B, and Magistral Small 24B were tested; preliminary results suggest reasoning (CoT) models like Magistral distribute confidence across longer traces instead of concentrating at PANL.
  • Mechanism ≠ accuracy: Gemma systematically overestimates its correctness (ECE 0.12); calibration remains a separate problem.
  • Attention blocking interventions may affect multiple indirect paths in a highly entangled architecture.
Reference: Conmy, A., Barbero, F., Osindero, S., Patraucean, V., & Veličković, P. (2026). *How do LLMs Compute Verbal Confidence?* Google DeepMind. arXiv:2603.17839.

Tags

#llm#mechanistic-interpretability#deepmind#confidence-calibration#hallucination-detection#metacognition#gemma#qwen

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177980951