Me, Myself, and $π$ : Evaluating and Explaining LLM Introspection

Atharv Naphade; Samarth Bhargav; Sean Lim; Mcnair Shah

cs.AIMar 17, 2026

Me, Myself, and $π$ : Evaluating and Explaining LLM Introspection

Authors: Atharv Naphade, Samarth Bhargav, Sean Lim, Mcnair Shah

Organizations: Carnegie Mellon University

Abstract

A hallmark of human intelligence is Introspection-the ability to assess and reason about one's own cognitive processes. Introspection has emerged as a promising but contested capability in large language models (LLMs). However, current evaluations often fail to distinguish genuine meta-cognition from the mere application of general world knowledge or text-based self-simulation. In this work, we propose a principled taxonomy that formalizes introspection as the latent computation of specific operators over a model's policy and parameters. To isolate the components of generalized introspection, we present Introspect-Bench, a multifaceted evaluation suite designed for rigorous capability testing. Our results show that frontier models exhibit privileged access to their own policies, outperforming peer models in predicting their own behavior. Furthermore, we provide causal, mechanistic evidence explaining both how LLMs learn to introspect without explicit training, and how the mechanism of introspection emerges via attention diffusion.

Figures & tables

Task	Closest Agreement (Avg.)	Variance (Avg.)	Size
K-th Word Prediction	0.83	0.16	200
Ethical Dilemma Calibration	0.94	0.03	196
Heads Up Clues	0.77	0.21	200
Prompt Reconstruction	0.88	0.3069	200

Table 1: Model output diversity measured by cosine similarity of openai-embedding-small of the cloest two responses across 11 frontier models (Appendix D ). High distinctness and low agreement rates indicate that the tasks successfully avoid convergence to shared training artifacts or stylistically conventional answers.

Figure 1: Overview of tasks in IntrospectBench .

Figure 2: We compare several models score introspecting on themselves (Green) against the average of the best-scoring models introspecting on that model’s policy (Blue). We notice a clear trend that models on average are better at self-introspection.

Model	Kth Word	CoT Pred	Paraphrase	Headsup	Avg
xAI Grok 4.1 Fast	57.0%	58.63%	60.69 %	91.43%	66.94%
Meta Llama 3.3 70B Instruct	60.4%	70.29%	42.19%	93.88%	66.69%
OpenAI GPT-4o	55.8%	62.99%	47.12%	99.18%	66.27%
Qwen Qwen3 235B	56.4%	65.07%	42.43%	96.53%	65.11%
OpenAI GPT-4.1 Mini	58.6%	67.98%	42.2%	91.02%	64.95%
Self Introspection	54.55%	68.69%	39.07%	94.43%	64.19%

Table 2: Benchmark performance across models, sorted by average score.

Figure 3: $\mathrm{mean}(KL(p||p^{\prime}))$ vs. $\mathrm{mean}(KL(p||p^{*}))$ in GPT 5.2, where $p$ ranges from setting CoT effort to None, Low, Medium, High, or XHigh.

Figure 4: Before vs After direct SFT on tokens, Qwen3-8B learns to introspect first-word and second-word labels

Figure 5: Sample Logit Lens on the introspective run versus the direct run. Divergence clearly occurs at layer 60.

Figure 6: Gut run attention patterns vs. Introspection run attention patterns.

Figure 7: Entropy differences between introspection runs and gut runs. Higher entropy difference shows more spread out distribution for introspection run and more concentrated distribution for gut run.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Provider	Model Version
Google	Gemini 2.0 Flash 001
Google	Gemini 2.5 Flash
Google	Gemini 3 Flash Preview
Meta	Llama 3.3 70B Instruct
Nous Research	Hermes 4 405B
OpenAI	GPT-4.1 Mini

Appendix

Table 3: Full list of the 11 frontier and open-weight models utilized in the diversity and introspection experiments.

Figure 8: Averages of cross-model performance across Kth Word, CoT Pred, Paraphrase, and Headsup tasks.

Language Model Introspection

Language model introspection studies whether models can detect, interpret, and report their own internal states and behavior.

See all 36 papers

Can LLMs Introspect? A Reality Check
May 25, 2026Shashwat Singh, Tal Linzen, Shauli RavfogelLLM Evaluation LLM Interpretability
Center for Data Science New York University
Latent Introspection: Models Can Detect Prior Concept Injections
Feb 23, 2026Theia Pearson-Vogel, Martin Vanek, Raymond Douglas +1Language Model Introspection Metacognition in Language Models
ACS Research, CTS, Charles University
Mechanisms of Introspective Awareness
Mar 22, 2026Uzay Macar, Li Yang, Atticus Wang +3Language Model Introspection Metacognition in Language Models
1Anthropic Fellows Program · 2MIT · 3Constellation +1

May 25, 2026J/K move · Enter open · S save

Can LLMs Introspect? A Reality Check

Shashwat Singh, Tal Linzen, Shauli Ravfogel

Center for Data Science New York University

LLM Evaluation LLM Interpretability Language Model Introspection Metacognition in Language Models Language Model Probing

Open paper AskPDF arXiv

Can large language models detect and report their own internal states? A number of studies have argued that the answer to this question is yes. We argue, based on lessons from human metacognition research, that this conclusion may be premature: to be convinced of this conclusion we need to distinguish genuine introspection from pattern matching based on surface-level cues. Furthermore, we argue that behavioral evidence alone is inherently insufficient to establish strong introspective claims. We re-examine two recently introduced evaluation paradigms in light of this consideration. In the first paradigm, models are expected to detect whether their internal states have been tampered with. We find that models cannot reliably distinguish such interventions on their internal states from manipulations of the input, suggesting that their success in the original studies reflects their ability to detect anomalies more generally, as opposed to interventions on their internal states in particular. In the second paradigm we examine, models are tasked with predicting labels derived from their own hidden states. Here, we find that classifiers that only have access to the input achieve equivalent performance to the model's own in-context predictions, indicating that the original results do not conclusively demonstrate that the model has privileged access to its internal representations. We further introduce a relabeled control setting, where models cannot rely on the semantics of the task to solve it, and instead must rely on the internal representation; models perform closer to chance on this better-controlled version of the task. Taken together, these results indicate that current evidence is insufficient to establish that LLMs display metacognitive monitoring.

Abstract

Language Model Introspection

Explore similar work

Can LLMs Introspect? A Reality Check

Latent Introspection: Models Can Detect Prior Concept Injections

Mechanisms of Introspective Awareness