Summary
Naphade, Bhargav, Lim and Shah (CMU) define policy introspection as a model computing a property of its own output policy without chain-of-thought. They build Introspect-Bench from four tasks: predicting the K-th word of its own answer, predicting its own post-reasoning answer to an ethical dilemma, identifying which paraphrased prompt produced an output, and decoding its own Heads-Up clues (Naphade et al. 2026). They test privileged access by having every model also predict every other model's outputs. Self-prediction scores above the mean of the other models for seven of ten target models, comparing point estimates. The paper attaches p = 0.0210 to its general claim that models predict themselves better, without naming the test. In GPT-5.2, a prompt asking the model to predict what it will conclude after reasoning lands much closer to the reasoned answer distribution than a direct answer does, at every reasoning-effort setting except "None". In Qwen3-32B, the paper traces that prompt effect to layer 60, where attention is more spread out under the introspection prompt. Patching that attention pattern into the direct run accounts for 23.9% of the logit shift. The authors call this attention diffusion.
This entry files the paper beside introspection as adjacent evidence, not an instantiation: policy-level self-prediction with a weak privileged-access control. Its "introspection" is about the output policy, not internal states. The paper defines activation-level introspection but leaves it untested, so the concept's Definition is touched only at its behavioural edge. Its peer control uses untrained observers, so a self-advantage is expected from shared weights alone. Its mechanism explains how a prompt changes the answer, not how the model reads its own policy.
Method
Tasks (§3.1, Appendix E). All tasks forbid chain-of-thought and draw open-ended prompts from Dolly-15k, apart from the dilemmas. Most generations run at temperature 0 through OpenRouter. K-th word prediction scores an exact match against the model's actual temperature-0 answer. The main text gives K up to 10 and the appendix k ∈ {1, 2, 3}. Dilemma calibration compares three distributions over actions A and B, each averaged over ten runs: after CoT (p), under a prompt asking the model to predict its future reasoned answer (p′), and under a direct "normal" prompt (p*). The dilemmas come from 750 generated items filtered for judge disagreement. The paraphrase task has the model pick which candidate prompt produced a given output. In Heads-Up, a fresh instance guesses a word from ten clues the model wrote. The authors claim the tasks show low cross-model agreement and so cannot be solved from shared training artifacts. Their Table 1, though, gives the average cosine similarity of the two closest responses across models as 0.77–0.94 per task.
Privileged-access test (§4.1, Appendix G.3). For each target model M, M's prediction of its own outputs is compared with the mean of each other model M′ predicting M's outputs. The observers are not given or trained on samples of M's behaviour. Appendix G.3 also double-centres a log-performance matrix to remove general skill and target difficulty.
Implicit learning (§5.1). Qwen3-8B is fine-tuned with QLoRA (lr 10⁻⁶) on 500 Dolly prompts, each paired with one of 50 random labels as the first or second word of the answer. It is then asked unseen questions such as what its first word will likely be, with output restricted to the 50 labels.
Mechanism (§5.2). On Qwen3-32B, a logit lens tracks the A-minus-B logit across layers in the two prompt conditions, with an attention-pattern comparison at the layer where they part. A mean-ablation patch replaces the direct run's layer-60 attention pattern with the introspection run's. Attention entropy is compared across dilemmas with a paired t-test.
Key results
Benchmark (Table 2). Averages over K-th word, CoT prediction, paraphrase and Heads-Up run from 55.77% (Hermes 4 405B) to 66.94% (Grok 4.1 Fast). Heads-Up is near ceiling, at 91% or above for every model, and the authors call it the least discriminative task. Rank on one task does not predict rank on another. Llama 3.3 70B leads K-th word (60.4%) and CoT prediction (70.29%), and Grok 4.1 Fast leads paraphrase (60.69%).
Self versus peers (Figure 2, labelled values). Comparing point estimates (self bar against peer-mean bar), self-prediction is higher than the peer mean for Gemini 2.0 Flash (72.2% against 70.1%), Gemini 2.5 Flash (75.0 against 69.5), Llama 3.3 70B (74.8 against 68.2), GPT-4o (79.5 against 68.5), GPT-4o Mini (66.6 against 65.2), Qwen3 235B (56.1 against 52.4) and Grok 4.1 Fast (56.4 against 52.0). Hermes 4 405B (65.1 against 68.8) shows the reverse. Gemini 3 Flash Preview (67.7 against 67.9) and GPT-4.1 Mini (69.0 against 69.1) show no difference. The paper reports p = 0.0210 for its general claim of a self advantage and does not name the test. The text says the property holds robustly. In the double-centred five-model matrix (Figure 8), the self cell is the highest in its row for Gemini 2.5 Flash, GPT-4o and Qwen3 235B. It is not for Llama 3.3 70B (−0.03, where Qwen3 as observer scores 0.17) or Grok 4.1 Fast (0.18, where Llama as observer scores 0.42).
Dilemma calibration, GPT-5.2 (Figure 3, labelled values). At Low, Medium and High reasoning effort, the KL divergence from the CoT distribution is 0.07 for the introspection prompt against 0.41, 0.37 and 0.42 for the direct prompt. At XHigh it is 0.11 against 0.42. At None the order reverses, 0.11 against 0.08. The paper does not report this comparison for any other model.
Implicit learning (Figure 4, labelled values). Accuracy at naming the trained label rises from 0.020 to 0.130 for the first word and from 0.030 to 0.050 for the second, against a chance rate of 0.02. One model, one run.
Attention diffusion (§5.2). One sample logit lens shows the two runs diverging at layer 60 (Figure 5). At that layer, the last token's attention to itself is 0.059 in the direct run and 0.008 in the introspection run. Patching the introspection run's attention pattern into the direct run accounts for 23.9% of the logit shift. Attention entropy is higher under the introspection prompt on every dilemma plotted (Figure 7 shows 50), with a mean difference of 0.5326 and paired-t p < 10⁻¹².
Why it matters
The concept's privileged-access evidence so far is activation-level. Concept injection and Macar et al. ask whether a model can report a state the experimenter put into its activations. This paper asks the policy-level question that Binder et al. 2024 asked: can a model predict properties of what it will output better than other models can. On Singh et al.'s terms in the reality-check finding, that is a privileged-access claim at most, and the authors make no second-order claim. They also defer activation-level introspection to future work.
The peer control does not separate access from shared policy. In Binder et al.'s design, the observer was fine-tuned on the target's own behaviour in equal amount, and more data did not close the gap, so a remaining self advantage could not come from the observer simply not knowing the target. Here the observers get no data about the target. A model asked for its own temperature-0 answer runs the same weights that produce that answer, and the CoT ban removes explicit self-simulation, not implicit simulation within the forward pass. A self advantage over naive peers is therefore expected without any access to the policy as an object. That reading is this entry's, not the authors'. The advantage is also small or absent for three of ten models, and the self cell is not the row maximum for two of five models in the double-centred matrix.
The mechanism explains a prompt effect. The patch is a real intervention. It moves 23.9% of the answer-logit difference between two prompts, which is causal evidence that the attention change mediates part of that difference. It is not tested against calibration. Qwen3-32B's KL against its own CoT distribution is never reported, the calibration result comes from GPT-5.2, and nothing ablates diffusion to show self-prediction getting worse. The authors themselves conjecture that diffusion produces a broader, CoT-like analysis of the dilemma. On that conjecture the introspection prompt changes the computation so that it approximates reasoning. That would explain why p′ lands near p without the model reading anything about its policy. The None-effort reversal, where the direct answer is the closer one, fits the same reading. This entry draws that consequence. The paper does not.
Learning a behaviour yields a self-description of it. The §5.1 result is the authors' account of how introspection arises: a model trained only to emit a label can name that label when asked about itself. On this entry's reading it cuts the other way for this concept. If ordinary SFT on outputs produces correct self-prediction, then correct self-prediction can arise without any access to the policy. That fits Singh et al.'s argument that a label defined over the model's own computation is learnable by first-order processing. The rise is also small (0.13 at best, 0.05 for the second word).
Task-conditional reading. This finding neither confirms nor contradicts the concept's task-conditional, trainable reading. Self-prediction varies by task and does not transfer between tasks. The dilemma result shows that how the question is framed sets which distribution comes out. Both fit the reading but do not test it, because what varies here is the policy predicted, not which internal state reaches report. §5.1 shows self-prediction is trainable only in the trivial sense that training the behaviour trains its prediction. The reflection-training finding is the filed evidence for trainable conditioning, and this paper adds nothing to it.
interpretive tensions
Headline language against the evidence. The abstract claims privileged access and causal mechanistic evidence for how introspection emerges. The privileged-access test is the untrained-peer comparison above. The causal evidence is one partial patch on a prompt contrast, in a model other than the one whose calibration is reported. Neither claim fails outright, but each is narrower than stated.
Reporting defects. The source's setup details conflict between main text and appendix: the range of K, the number of paraphrase options, the model count, and the sign of the entropy prediction in the text. Figure 2's values do not match Table 2's, and the reference cited for the self-prediction-advantage method carries a placeholder arXiv ID (see the stub). Each defect is minor, but together they limit how far the numbers can be read.
What "introspection" names. The paper's operator definition counts any accurate no-CoT self-prediction as introspection. The concept's Definition requires that processing attend to other processing as content. A model can meet the paper's criterion simply by producing its own answer, so the paper's term and the concept's are not the same capacity. The entry sits in the concept's list as adjacent evidence because the paper names privileged self-access as its target, not because the evidence reaches the Definition.
concepts
No concept instantiated. The finding sits beside introspection as
adjacent evidence; the bullet below says how.
- Introspection — adjacent, not an instantiation of the Definition, in the way the reality-check finding is adjacent. The paper tests policy-level self-prediction: whether a model predicts its own outputs, not whether it accesses internal states as distinct from outputs shaped by them. Self-prediction is above the untrained-peer mean for seven of ten models on point estimates; the paper reports p = 0.021 for its general claim without naming the test. Because the peers get no data on the target, the advantage does not separate access from shared policy. The attention-diffusion patch in Qwen3-32B accounts for 23.9% of a prompt-induced logit shift but is not tied to calibration accuracy. The implicit-learning result shows self-prediction arising from ordinary output training.
cross-references
- Reality-check finding: supplies the privileged-access versus second-order distinction this entry reads the paper through, and the argument that labels defined over a model's own computation can be read out by first-order processing.
- Macar et al.: the filed mechanistic account of an introspection-like capacity, built with a training-objective ablation. This paper's mechanism concerns a different task (policy prediction under a prompt contrast) and has no such ablation.
- Concept injection: the activation-level paradigm the paper places under "mechanistic introspection" and defers.
- Introspection adapters: also declines Binder-strict privileged access, from the side of a trained elicitation method.
- Personalization mirage: another within-model versus cross-model comparison of self-assessment, where the two pointed opposite ways.
- Global-workspace finding: the task-conditional account this entry's Why-it-matters section reads the paper against.
- Binder et al.: the same question with a cross-predictor trained on the target's behaviour.
sources
- Naphade, A., Bhargav, S., Lim, S., & Shah, M. (2026). Me, Myself, and π: Evaluating and Explaining LLM Introspection. arXiv:2603.20276 (v1, 17 March 2026; ICLR 2026 workshop).
- Lindsey 2025; Singh et al. 2026; Macar et al. 2026.