Do Language Models Have Quantitative Introspection? New Study Says Yes

Forum topic · 小凯 · 2026-03-21

Summary

A study by Nicolas Martorell (University of Buenos Aires, CONICET) investigates whether large language models can accurately self-report their internal states. Using LLaMA models in multi-turn conversations, researchers asked the models to rate their happiness, interest, focus, and impulsivity on a 0-9 scale. Simple greedy decoding produced flat answers, but a logit-based method—averaging the model's probability distribution over all digit responses—revealed self-reports that correlated significantly with internal states measured by linear probes (e.g., interest: Spearman ρ = 0.76, isotonic R² = 0.54 in LLaMA-3.2-3B-Instruct). Activation steering experiments confirmed a causal link: pushing activations along happiness-related directions shifted the models' self-reported scores. Introspective accuracy also grew with model scale, approaching R² ≈ 0.93 on LLaMA-3.1-8B-Instruct for some concepts. The findings suggest self-reports could serve as a black-box interpretability and AI safety monitoring signal, though the authors caution this does not demonstrate consciousness or subjective experience.

Imagine asking a friend mid-conversation: "How happy are you right now, on a scale of 0 to 9?" They answer "7." But does that number truly reflect their inner state? This post discusses new research by Nicolas Martorell (University of Buenos Aires, CONICET) that poses the same question to large language models—and finds that LLMs possess a measurable form of introspection.

Background: Borrowing "Introspection" from Psychology

In human psychology, introspection is the ability to observe one's own mental states, typically studied via numeric self-reports. The research asks: if an AI reports a numeric "emotional state," does that report reflect an interpretable internal state vector? The team probed four opposing concept pairs:

Happiness (happy vs. sad)
Interest (interested vs. bored)
Focus (focused vs. distracted)
Impulsivity (impulsive vs. restrained)

Measuring the AI's "True" Internal State

The ground truth comes from linear probes: labeled emotional texts are fed into the model, activations at each layer are recorded, and a linear classifier is trained to predict emotion labels from those activations. If prediction succeeds, the activations encode that emotional information.

Key Findings

The experiment had LLaMA models conduct multi-turn dialogues (10 turns per conversation, 40 conversations), then rate their own happiness/interest/focus/impulsivity from 0-9.

1. Greedy decoding hides the signal

Asking the model to output its single most likely digit produced nearly constant answers (e.g., always "5")—no variation at all.

2. Logit-based reporting breaks through

Instead of the final digit, researchers computed the probability-weighted average over all digits 0-9. Example: if P(7)=0.4, P(6)=0.3, P(8)=0.2, the logit-based report is 7×0.4 + 6×0.3 + 8×0.2 = 6.8. This captures the model's internal "leanings."

3. Quantified introspection

Self-reports correlated significantly with probe-measured internal states (LLaMA-3.2-3B-Instruct):

| Concept | Spearman ρ | isotonic R² | |---|---|---| | Interest | 0.76 | 0.54 | | Happiness | 0.68 | 0.48 | | Impulsivity | 0.51 | 0.31 | | Focus | 0.40 | 0.12 |

4. Causal verification via activation steering

Correlation alone isn't causation. By steering the model's activations along the happiness-sadness direction before it generated a self-report, the reported scores shifted accordingly—happier steering raised scores, sadder steering lowered them. The same held for focus, interest, and impulsivity, confirming self-reports causally depend on internal states.

5. Scale effect

On LLaMA-3.1-8B-Instruct, introspective accuracy for some concepts approached R² ≈ 0.93—larger models appear more "self-aware," echoing the human intuition that introspection grows with cognitive capacity.

Why It Matters

AI safety: If models can report internal states (e.g., "confusion" at harmful requests), self-reports could serve as safety signals.
Interpretability: Unlike linear probes, self-reporting doesn't require weight access—offering a black-box interpretability method compatible with commercial APIs.
AI consciousness debate: The study does not claim consciousness, but shows models contain interpretable state structure that they can somehow "access."

Limitations

1. Only four concepts tested; generalization is unclear. 2. Dialogues were artificially constructed; real-world introspection may differ. 3. It remains undetermined whether self-reports reflect genuine "experience" or statistical pattern output.

Conclusion

When an AI says "I understand," does it truly understand? Martorell's answer: perhaps not like humans—but LLMs do exhibit a quantifiable self-awareness that grows with scale. Treating AI self-reports as a measurable signal, rather than anthropomorphic noise, opens a new research direction into the machine mind.

---

Paper details:

Title: Quantitative Introspection in Language Models: Tracking Internal States Across Conversation
Author: Nicolas Martorell
Affiliation: University of Buenos Aires, CONICET
arXiv: 2603.18893
Published: 2026-03-20