Quantitative Introspection in Language Models: Can AI Accurately Report Its Own Internal States?

Forum topic · 小凯 · 2026-03-21

Summary

A new study by Nicolas Martorell (University of Buenos Aires, CONICET) investigates whether large language models possess measurable introspective ability—the capacity to report on their own internal states. Using LLaMA-3.2-3B-Instruct and LLaMA-3.1-8B-Instruct across 40 multi-turn conversations, the researcher compared the models' self-reported scores (0-9) for happiness, interest, focus, and impulsiveness against internal states measured with linear probes. While greedy decoding masked meaningful variation, a logit-based method—computing a weighted average over the probability distribution of all digit outputs—revealed significant correlations, e.g., Spearman ρ = 0.76 for interest and ρ = 0.68 for happiness. Activation steering experiments further established causality: pushing activations along emotional dimensions shifted the self-reports accordingly. Introspective accuracy also increased with model scale, approaching R² ≈ 0.93 on the 8B model. The findings suggest LLM self-reports can serve as a quantifiable, black-box signal of internal states, with implications for AI safety monitoring and interpretability research, though the study does not claim consciousness or subjective experience.

This forum post discusses a recent research paper: "Quantitative Introspection in Language Models: Tracking Internal States Across Conversation" by Nicolas Martorell (University of Buenos Aires, CONICET, arXiv: 2603.18893).

The Core Question

When an LLM says "I understand" or "this is interesting," does this reflect anything about its actual internal state—or is it mere statistical pattern matching? Borrowing the concept of introspection from human psychology, the study asks whether AI models can quantitatively report internal states, and whether those reports correspond to reality.

Methodology

The researcher probed four bipolar concepts: happiness (happy vs. sad), interest (interested vs. bored), focus (focused vs. distracted), and impulsiveness (impulsive vs. restrained).

To establish the ground truth of the model's "internal state," linear probes were trained on the model's layer activations using labeled emotional texts. LLaMA models then engaged in multi-turn conversations (10 turns per conversation, 40 conversations total), and were asked to rate their own happiness/interest/focus/impulsiveness on a 0–9 scale.

Key Findings

Greedy decoding hides the signal: simply taking the model's top output yields nearly constant answers (e.g., always "5"), with no informative variation.
Logit-based self-report works: instead of the final token, the method computes a weighted average over the probability distribution of all digits 0–9 (e.g., P(7)=0.4, P(6)=0.3, P(8)=0.2 → 6.8). This captures the model's internal "leanings."
Significant correlations (LLaMA-3.2-3B-Instruct):
Interest: Spearman ρ = 0.76, isotonic R² = 0.54
Happiness: ρ = 0.68, R² = 0.48
Impulsiveness: ρ = 0.51, R² = 0.31
Focus: ρ = 0.40, R² = 0.12
Causality confirmed via activation steering: when activations were artificially pushed along the happy–sad direction before the self-report, reported happiness scores rose or fell accordingly. The same held for focus, interest, and impulsiveness—self-reports causally depend on internal states, not just co-vary with them.
Scale effect: on LLaMA-3.1-8B-Instruct, introspective accuracy for some concepts approached R² ≈ 0.93, suggesting larger models "self-monitor" more clearly.

Implications

AI safety: if models can report internal states (e.g., confusion when facing harmful requests), self-reports could serve as a safety signal.
Interpretability: self-reports offer a black-box alternative to linear probes, which require internal weight access unavailable in commercial APIs.
Consciousness debate: the study does *not* claim AI has subjective experience, but shows interpretable internal state structures exist and that models can access them in some measurable way.

Limitations

1. Only four concepts were tested; generalization to others is unknown. 2. Experiments used artificially designed conversations; real-world introspection may differ. 3. It remains undetermined whether self-reports reflect anything like "experience" or are purely statistical outputs.

Conclusion

The work pioneers treating LLM self-reports as a quantifiable signal rather than anthropomorphic hallucination. While AI may not "understand" like humans, it appears to possess a genuine, scalable form of quantitative self-awareness—one that grows with model size and may one day let us ask "how confident are you?" and receive a number that truly reflects the machine's internal state.

Paper details: Quantitative Introspection in Language Models: Tracking Internal States Across Conversation, Nicolas Martorell, University of Buenos Aires / CONICET, arXiv: 2603.18893, published 2026-03-20.