{"id":"b38250f7-273d-47f5-aad1-b2bdca7c2bbe","arxiv_id":"2508.14802","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"LLMs' temperature self-reports are confounded by prompt style and show no privileged self-access, supporting a thicker definition of AI introspection.","lead":"This paper proposes that AI introspection should mean a model reporting its own internal states more reliably than any outsider could at the same cost, and tests this by asking LLMs to guess their own sampling temperature. The models' temperature guesses track prompt style, not internal state, and self-reports are no more accurate than a third model's predictions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Binary HIGH/LOW report channel confounds self-access; null result may reflect task insensitivity rather than absence of privileged access.","rationale":"The reader's weakest_assumption identifies exactly the concern I find most load-bearing: the HIGH/LOW binary report may not be a valid channel for introspection, so the null result does not falsify privileged access. This is a measurement-validity issue, not merely a statistical one. Study 1's own data demonstrate that the report is heavily influenced by prompt framing, making it a noisy and potentially biased measure of any internal temperature signal. Study 2 inherits this flaw, and the paper does not provide evidence that the binary response captures the information the model would need to express privileged access. My proposed concrete test—using a continuous numerical estimate—would directly test whether the null result persists when the report channel is more expressive and less susceptible to framing. If the continuous self-report outperforms across-model prediction, then the paper's central empirical claim collapses; if not, the concern is resolved. I agree with the reader's assessment that the verdict should remain CONDITIONAL, because the conceptual argument for a thicker definition is sound, but the empirical support is not yet decisive.","tokens_in":5700,"tokens_out":5734,"duration_ms":72400,"concrete_test":"Replace the binary report with a continuous temperature estimate. In the self-reflection condition, prompt the generator: 'After writing your sentence, estimate the numerical value of your sampling temperature, given the sentence you wrote.' In the within/across-model prediction conditions, use the same continuous estimator on the same prompt and generated sentence. Compare mean absolute error or correlation with the true temperature, separately for factual and crazy prompts. If, under continuous reporting, self-reflection error is significantly lower than across-model error (and remains so after controlling for prompt framing), the binary result was an artifact of the forced choice. If self-reflection still shows no advantage, the paper's conclusion is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—that models show no privileged self-access on temperature—rests on a forced binary HIGH/LOW report. Study 1 (Section 2) shows these reports are dominated by prompt framing (factual vs. crazy) and are largely insensitive to actual temperature. Study 2 reuses the same binary channel and the same subject/prompt manipulations. Consequently, even if a model possessed privileged access to its internal temperature, that signal could be swamped by the framing effect and by the coarse response format. The observed null result is exactly what one would expect if the model ignores its internal state and answers from surface text statistics; it does not distinguish 'no privileged access' from 'privileged access blocked by an insensitive readout.' Since the paper's conclusion is a negative existential claim about privileged self-access, this measurement-validity concern is load-bearing. Without a more continuous, less prompt-biased report, the experiment cannot rule out the possibility that the model has privileged information that simply cannot be expressed in a single HIGH/LOW word.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that the 'lightweight' definition of introspection proposed by Comșa and Shanahan (C&S) is too permissive because it does not require privileged self-access. The authors propose a 'thicker' definition: introspective processes must yield information about internal states more reliably than any equal- or lower-cost third-party process. They then present two studies. Study 1 varies prompt framing (factual/normal/crazy) and subject (elephants/unicorns/murlocs) in a temperature self-report task and finds that reported temperature tracks framing more than the actual sampling temperature. Study 2 compares self-reflection (the generator reports its own temperature) with within-model and across-model prediction of temperature from the prompt and generated sentence. The authors report that accuracy is near chance and self-reflection is not better than across-model prediction, concluding that models do not show privileged self-access to their temperature. The paper positions this as evidence against uncritical use of C&S's lightweight notion of introspection.","tokens_in":5941,"tokens_out":5411,"duration_ms":68467,"significance":"If the empirical claims were fully supported, this would be a practically important result: it would caution against treating LLM self-reports about their own generation parameters as privileged information. The paper's main conceptual contribution—requiring privileged self-access—is a useful corrective to definitions that conflate any causally sensitive self-report with introspection. Strengths include a clear conceptual distinction, a reasonable third-party baseline design, and publicly available code/data. However, the central empirical support is not yet at the level needed to establish a negative existential claim about privileged self-access. The binary response channel, the lack of inferential statistics, and the treatment of invalid responses are load-bearing concerns.","major_comments":[{"comment":"The central null result relies on a forced binary HIGH/LOW verbal report. Study 1 itself shows that this response channel is strongly biased by surface prompt wording (e.g., 'crazy' yields HIGH even at low temperature). A null result with such a channel cannot distinguish 'no privileged self-access' from 'privileged self-access that is not expressible in this coarse, framing-dominated channel.' To make the negative conclusion load-bearing, the authors need to validate the readout—for example, by showing that the same channel can transmit privileged information in a positive-control condition (e.g., after fine-tuning or injecting the true temperature into the context), or by using a graded/continuous report and demonstrating that the model's self-report is not swamped by framing.","section":"Section 3 / Appendix B.2"},{"comment":"The key comparison—'self-reflection accuracy is no better than across-model prediction'—is presented without confidence intervals or inferential statistics. The point estimates (e.g., 0.50 vs. 0.49–0.55) are numerically close to chance, but no test quantifies the uncertainty. Moreover, the threshold selection (temperatures ≤0.5 and ≥1.5) appears post hoc; no justification or sensitivity analysis is provided. The paper should report exact per-cell N, include confidence intervals / paired tests or mixed-effects models, and show that the conclusions are robust to alternative thresholds, including the full temperature range.","section":"Section 3.1 / Figure 2b"},{"comment":"Responses without a valid HIGH/LOW judgment are excluded from the analysis, but the exclusion rates are not reported. If invalid responses are more common in certain models, prompts, or temperature conditions, this can bias the reported accuracies. Since the self-reflection and prediction conditions use different prompt types, the exclusion pattern may differ across conditions, potentially affecting the central self-vs-across comparison. The authors should report exclusion rates and run robustness checks (e.g., treating invalid responses as incorrect, or modeling them directly).","section":"Section 2.1 / Section 3"},{"comment":"The temperature parameter is an API-level decoding hyperparameter, not a state represented in the model's weights or activations at inference time. The model's forward pass does not include the current temperature as an input; the model can at best infer it from the generated text, which is equally available to a third party. Under the paper's own definition of introspection as access to internal states, it is unclear that temperature qualifies. If it does not, the observed null result is architecturally expected and does not support the broader conclusion that models lack privileged self-access to internal states. The authors should either justify temperature as a legitimate internal state under their definition or choose a state that is actually represented inside the model.","section":"Section 1 / Section 3"}],"minor_comments":[{"comment":"The color scale for 'Predicted Temp' is not explicitly described in the caption or text; with four panels plotted side-by-side, a colorbar or explicit legend is needed for readability. In grayscale, the differences between 0.5 and 1.0 are difficult to distinguish.","section":"Figure 2a"},{"comment":"The number of trials per condition in Study 2 is not stated. The paper says only that all sentences generated at temperatures ≤0.5 and ≥1.5 are used. Please report how many sentences and how many prediction judgments per model/prompt/temperature cell were collected.","section":"Section 3"},{"comment":"Reference [8] appears to be a citation to Lieberman (1963) on semantic and grammatical context in speech, which does not seem related to introspection. This is likely a typo in the reference list; please verify and correct.","section":"References"},{"comment":"The phrase 'we found no evidence of introspection in models' could be misinterpreted as a global claim. The evidence concerns temperature reports in the tested models. Please qualify the conclusion to the task and models studied, or explicitly state the intended scope.","section":"Figure 1 / Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper makes a conceptually valuable point and provides a useful baseline design, but the empirical support for the negative claim is not yet conclusive. The binary readout validity issue and the missing inferential statistics are the main barriers. I would like to see a revision that adds a positive control for the response channel, reports confidence intervals and sensitivity analyses, and more carefully scope the temperature-as-internal-state assumption. This is fixable within the manuscript's scope, so I am not recommending rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know about this paper because it draws a clean line between 'lightweight' introspection (a model can infer its temperature from its own output) and privileged self-access (the model does better than a third party with equal compute), then offers a simple cross-model test for the latter. That test is the real contribution. The framing manipulation in Study 1 is also effective: telling the model to write a 'crazy' sentence pushes self-reported temperature to HIGH even at low sampling temperatures, which genuinely undermines the C&S case study. Credit is due for releasing code and data and testing four current models.\n\nThe soft spot is Study 2. The central claim—no privileged access—is a null result built on point estimates. There are no confidence intervals, no inferential statistics, and the temperature thresholds (<=0.5, >=1.5) look post hoc. Invalid responses are discarded without any analysis of what got dropped. More load-bearing, the forced HIGH/LOW readout is a very coarse channel. If the model has privileged information but it's swamped by the framing effect and can't be expressed in one token, the test can't see it. So the negative existential claim is stronger than the data support. The theoretical definition also leaves 'computational cost' informal, which weakens the boundary between privileged and non-privileged access. The paper is aware it's stipulating, but that doesn't fix the measurement problem.\n\nWhere does that leave us? The direction is right: this reframing from self-description to privilege over third-party evaluation matters for interpretability and safety. The null result is worth taking seriously, but as 'no evidence of privileged access' rather than 'models lack privileged access.' The authors are honest that larger models might do better, and they cite Binder et al. for positive evidence with fine-tuning.\n\nFor whom? Anyone working on LLM self-knowledge, interpretability, or evaluation of self-reports. It won't settle the debate, but it sharpens the question and gives a usable benchmark.\n\nMy recommendation: send it to peer review. It's a clear, falsifiable, reproducible contribution. But I'd insist on proper statistics or a moderated conclusion, and I'd push for a more continuous or less prompt-sensitive readout before accepting the strong negative claim. As it stands, the central inference is plausible, not proven.","headline":"A genuinely useful reframing of AI introspection as privileged self-access, but the empirical null result is too weakly supported and the binary report channel may be masking the very signal the test hunts for.","tokens_in":6359,"tokens_out":1690,"would_cite":true,"duration_ms":19721,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that LLMs can appear to introspect under a lightweight definition while lacking the privileged self-access that makes introspection practically valuable, and supports this with temperature-reporting experiments.","keywords":["LLM introspection","privileged self-access","temperature","metacognition","self-report","large language models","introspection definition"],"falsifier":"A direct test: hold the prompt fixed to a single stylistic condition and randomize the actual sampling temperature across many trials; if a model's own HIGH/LOW report tracks the actual temperature significantly above chance and above the accuracy of a different model given the same prompt and generated sentence, the paper's claim of no privileged self-access is falsified. A simpler observation: a single trial where a 'crazy' prompt at low temperature yields an accurate LOW report, or a factual prompt at high temperature yields HIGH, would already cut against the framing-driven account.","tokens_in":5616,"feed_emoji":"🌡️","tokens_out":7193,"duration_ms":74345,"temperature":0.7,"pith_summary":"The paper is trying to establish that a popular 'lightweight' test for AI introspection—can a model correctly describe one of its own internal states?—is too weak, because a model can pass it by reasoning about what its outputs look like rather than by accessing anything private. It proposes a thicker definition: introspection must yield information about internal states more reliably than a third party could obtain at equal or lower computational cost. On a concrete case, LLM reports of their own sampling temperature, the paper shows that reports track prompt style ('crazy' versus 'factual') rather than the actual temperature, and that self-reflection is no more accurate than chance or than predicting another model's temperature. If right, this means temperature-based demonstrations of introspection do not deliver the practical benefit that makes self-knowledge valuable: bypassing external evaluators for honesty, interpretability, and calibration.","feed_headline":"LLM temperature self-reports show no privileged self-access","feed_subtitle":"Models answer from prompt style, not their own state, so such reports cannot replace external checks.","key_machinery":"The central object is the paper's proposed definition of introspection in AI: a process yields introspective information only if it is more reliable than any third-party process of equal or lower computational cost. This turns introspection into a comparative, testable property: a model must beat an external evaluator who sees the same prompt and output. The experimental machinery is the three-way comparison—self-reflection (generate and self-report), within-model prediction (an independent run of the same model infers the generator's temperature), and across-model prediction (a different model infers it)—with accuracy measured against the actual temperature setting.","core_discovery":"The paper's central claim is that a model can look introspective under a lightweight definition—accurately describing an internal state when prompted—without having any privileged access to itself. The authors test this with sampling temperature: they prompt a model to write a factual, neutral, or 'crazy' sentence and then report whether its own temperature is HIGH or LOW. Reports track the prompt's style rather than the actual sampling setting; a 'crazy' prompt reliably produces a HIGH report even at low temperature. They then compare self-reflection with prediction of another model's temperature from the same prompt and sentence. Self-reflection is no more accurate than chance and no more","pith_inferences":["Beyond the paper: the same three-way comparison can be applied to any claimed internal state—learned knowledge, uncertainty, hidden activations—by pitting the model against a third-party predictor with matched prompts and budget; a null result for temperature suggests many 'self-knowledge' results should be rechecked this way.","Beyond the paper: the framing sensitivity implies that instruction-tuned models may be reporting what kind of text they think they produced, not what state they were in, which is a general confound for verbal self-reports.","Beyond the paper: a natural extension is to test whether models can be trained to give privileged reports, for example by fine-tuning on actual temperature settings with feedback, turning the proposed definition into a training target.","Beyond the paper: if privileged self-access continues to fail across internal states, the practical consequence is that AI audits should rely on external behavioral probes rather than the model's own explanations."],"forward_implications":["Temperature self-reports from current LLMs should not be counted as evidence of introspective access, since they are no better than chance or than another model's guess.","The lightweight definition of introspection is insufficient for practical uses like honesty and calibration, because it can be satisfied by reasoning about what a system like oneself would produce.","To demonstrate introspection in an AI, an experiment needs to show a reliability advantage over a third party with the same computational budget, not merely a correct answer.","The absence of evidence here is not a general no-introspection claim; larger or fine-tuned models may still show privileged self-access, but the burden of proof shifts to such demonstrations."],"supporting_citations":[{"why":"Proposes the lightweight definition of introspection and the temperature self-report case study that this paper re-tests and disputes.","marker":"[6]"},{"why":"Reports privileged self-access in larger fine-tuned models, providing the benchmark that motivates the paper's thicker definition and its conclusion.","marker":"[3]"},{"why":"Earlier work arguing that language models fail to introspect about their knowledge; the paper extends this standard to internal states like temperature.","marker":"[11]"},{"why":"Supplies the philosophical characterization of introspection as direct access to one's own mental states, the intuitive core of the thicker definition.","marker":"[4]"},{"why":"Cited to allow introspective processes to be less than perfectly reliable while still requiring a third-party-relative advantage.","marker":"[9]"}],"fun_headline_variants":["LLMs can't introspect: self-reports follow prompt, not temperature","AI self-knowledge is an illusion, temperature test shows","Chatbots fake introspection: reports match prompt, not state","Probe reveals LLM introspection is just prompt-style mimicry","Temperature self-reports expose AI's lack of privileged access"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that a forced single-word HIGH/LOW answer is a valid way for a model to express introspective access to its own temperature; if that binary channel cannot carry the internal state, the null results do not prove the absence of privileged access.","fun_headline_variants_meta":{"raw":{"variants":["LLMs can't introspect: self-reports follow prompt, not temperature","AI self-knowledge is an illusion, temperature test shows","Chatbots fake introspection: reports match prompt, not state","Probe reveals LLM introspection is just prompt-style mimicry","Temperature self-reports expose AI's lack of privileged access"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":970,"prompt_tokens":599,"completion_tokens":371,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":343,"completion_tokens_details":{"reasoning_tokens":285}},"tokens_in":343,"tokens_out":371,"duration_ms":4771,"temperature":1.0,"reasoning_tokens":285,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:14:38.968967+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test: hold the prompt fixed to a single stylistic condition and randomize the actual sampling temperature across many trials; if a model's own HIGH/LOW report tracks the actual temperature significantly above chance and above the accuracy of a different model given the same prompt and generated sentence, the paper's claim of no privileged self-access is falsified. A simpler observation: a single trial where a 'crazy' prompt at low temperature yields an accurate LOW report, or a factual prompt at high temperature yields HIGH, would already cut against the framing-driven account.","supporting_citations":[{"cited_title":"Introspection","cited_arxiv_id":null,"evidence_quote":"Supplies the philosophical characterization of introspection as direct access to one's own mental states, the intuitive core of the thicker definition."}],"review_version":1}