{"id":"b2e5addd-f78f-48ba-8799-72937b6bc6cd","arxiv_id":"2607.16799","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Transferable LLM 'truth' probes preserve a self-judgement-associated polarity rather than an objective-correctness polarity when the two conflict.","lead":"When a language model's self-judgement conflicts with objective correctness, standard hidden-state correctness probes often track the model's self-judgement instead. Across four models, the self-judgement direction transfers between tasks while the objective-correctness direction does not, so transferable probes do not automatically certify truthfulness.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Linear separability assumption in Eq. S1 is the load-bearing premise for the 'OC-associated direction' label; acknowledged in Scope, and testable with a nonlinear probe.","rationale":"The reader accepted with ACCEPT/MODERATE confidence, identifying the linear separable representation model of Eq. S1 as the weakest assumption. I agree that this is the load-bearing point. The paper has substantial independent support: the OC-only mass-mean control (S8.2) does not require factorial decomposition and shows all eight cross-domain AUCs above 0.5, which alone demonstrates that conventional correctness-labelled contrasts can be SJ-dominated. The zero-target-fitting OOD results, question-level controls (S8.1, S8.6), token-count matching, label-shuffle nulls, and counterbalanced X/Y mappings all strengthen the empirical asymmetry. The residual risk is semantic labeling of Wtruth: if encoding is nonlinear or nonseparable, 'OC-associated direction' is a fitted contrast, not a latent OC axis. But the paper acknowledges this in Scope and S12, and its conclusions are phrased around linear activation-mean directions. The central cautionary claim—transferability alone does not establish OC semantics—is robust even if the linear model fails. Therefore the concern does not change the verdict; it points to a specific experiment that would further tighten the claim.","tokens_in":31763,"tokens_out":12207,"duration_ms":123468,"concrete_test":"Using the released activation artifacts, fit a nonlinear probe on the source domain's final-token residual streams to predict OC (e.g., an MLP with one hidden layer of 256 units, or logistic regression on the first 100 PCA components), with the same question-grouped train/test split, and apply the frozen probe to the same target-domain B/C conflicts. Compute fixed-window AUC(W_nl→OC) exactly as in Exp2B. If W_nl→OC ≥ 0.5 with 95% CI above chance in conditions where Wtruth→OC < 0.5, the linear additive model is the load-bearing assumption and the component-level claim must be restricted to linear readouts. If W_nl→OC remains ≤0.5 in those conditions, the concern is settled and the conclusion is robust to nonlinear encoding.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central component-level claim—that the transferable readout preserves SJ polarity rather than OC polarity—rests on the additive linear model in S5.1/Eq. S1: x = α o vOC + β s vSJ + η o s vINT + ε. If this model holds, Wmeta = 2β vSJ and Wtruth = 2α vOC, and the interaction cancels; then below-chance Wtruth→OC AUCs are evidence that the OC direction does not transfer. If the true encoding is nonlinear, if OC and SJ directions are not separable, or if the strict-pair high-confidence selection leaves a confound (fluency, answer likelihood, response policy) in the cell means, then Wmeta and Wtruth are just fitted contrasts, not semantic directions. In that case the headline 'most reliably transferable component preserves SJ-associated polarity' could be an artifact of the linear estimator rather than a fact about the model's representation of objective correctness. The paper explicitly scope-limits itself to linear activation-mean directions (Scope, S12), so this is not an internal inconsistency, but it is exactly the condition that has to be true for the abstract's unqualified component-level claim. The OC-only mass-mean control and OOD zero-fitting results support the broader cautionary conclusion independently of the factorial decomposition; the component-level semantic labels are the vulnerable part.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a diagnostic for correctness probes that separates objective correctness (OC) from the model's self-judgement (SJ) by constructing conflict cases in which the two variables predict opposite orderings. It extracts final-answer-token residual-stream activations, defines factorial contrasts Wmeta and Wtruth intended to isolate SJ- and OC-associated directions, and evaluates their within-domain and cross-domain transfer across four instruction-tuned models and two free-response domains, plus zero-target-fitting OOD evaluations on MMLU and binary TruthfulQA. The central empirical finding is that the SJ-associated direction transfers above chance in essentially every condition, while the OC-associated direction has below-chance or near-chance AUCs for the expected OC ordering, and the conventional mixed contrast often ranks wrong/self-endorsed responses above correct/self-rejected responses. The paper concludes that transferability alone does not establish objective-correctness semantics.","tokens_in":32081,"tokens_out":6109,"duration_ms":66066,"significance":"If the result holds, it is an important and timely caution for the interpretability literature: it provides a concrete, multi-model demonstration that a transferable correctness readout can preserve the polarity of the model's own judgement rather than external correctness. The paper's strengths are substantial: four models, two free-response domains and two OOD benchmarks, strict confidence-filtered conflict sets, cluster-bootstrap inference that resamples source and target questions, source-label-shuffle and random-direction nulls, token-count and answer-likelihood residualization, and a second strict-pair draw. The OC-only mass-mean control is particularly valuable because it shows that even a direction fitted without any SJ labels follows SJ on conflicts. The paper is also unusually candid about its scope and the conditional nature of its estimates.","major_comments":[{"comment":"The component-level labels Wmeta and Wtruth are derived from the additive linear model x = α o vOC + β s vSJ + η o s vINT + ε. If the true residual-stream encoding is nonlinear, or if OC and SJ directions are not separable in the way this model assumes, then these vectors are fitted contrasts rather than semantic directions. The paper acknowledges this in Scope, but the abstract and Discussion state the component-level conclusion ('SJ-associated polarity', 'the SJ-associated direction') without the caveat. Please either qualify the abstract explicitly (e.g., 'under a linear factorial decomposition of activation means') or provide a nonlinear-probe check to show the labels are not artifacts of the linear estimator. The broader cautionary conclusion is supported independently by the OC-only mass-mean control, but the component-level interpretation rests on Eq. S1.","section":"Abstract; §5.1/Eq. S1; Scope (S12)"},{"comment":"Source and target SJ labels are generated by the identical yes/no self-judgement prompt, whereas OC labels come from dataset-specific parsers or answer keys. Cross-domain transfer of Wmeta could therefore partly reflect shared surface alignment of the elicitation procedure rather than a stable semantic self-judgement direction. The answer-likelihood residualization, the OC-only control, and the counterbalanced X/Y control reduce this concern materially, but the Discussion should state the asymmetry explicitly: the SJ transfer evidence is symmetric in task format in a way that the OC evidence is not. S12 mentions that SJ may contain variance from confidence, familiarity, or response policy, but does not connect this to the source-target sharing of the judgement prompt.","section":"Method (SJ elicitation); Exp2B; S12"}],"minor_comments":[{"comment":"The terms 'latent directions vOC and vSJ' may suggest independent latent axes. Since Wmeta and Wtruth are defined as factorial contrasts and are not shown to be orthogonal or unique, please clarify in the main text that these are operational contrast directions, not necessarily independent semantic axes in activation space.","section":"Eq. S1 and surrounding text"},{"comment":"The attrition from the original pools to the strict paired sample is severe (e.g., OLMo-3-7B Math retains 362 strict pairs from 5,549 questions). The main text states the conditional nature of the estimand, but it would help readers if the approximate retention rates were stated in the main experimental protocol rather than only in the supplement.","section":"Sample flow and Scope"},{"comment":"The figures are information-dense and the colors (teal/amber, blue/orange) may be hard to distinguish in grayscale. Adding direct text labels at the curve level or using line styles would improve accessibility.","section":"Figures 2–4"},{"comment":"The all-layer counts such as 458/560 (81.8%) are described as descriptive in S6, but the main text could explicitly remind readers that adjacent layers are not independent observations so these proportions are summaries, not primary inference.","section":"Results, OOD section"}],"recommendation":"minor_revision","confidential_remarks":"The paper is careful, reproducible, and its main cautionary conclusion is well supported. The only concern I would raise to the editor is interpretive overreach in the abstract relative to the linear-model scope; the authors should either temper the abstract or add a nonlinear probe test. This is a local fix rather than a fundamental flaw, and the evidence base is unusually thorough."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth your time. The central finding holds up: on high-confidence cases where objective correctness and the model's own self-judgement disagree, conventional correctness-labelled contrasts follow the model's judgement, not the ground truth. The cleanest evidence is the OC-only mass-mean control — a direction fitted without any SJ labels still ranks wrong/self-endorsed answers above correct/self-rejected answers in all eight cross-domain comparisons. That alone should make researchers cautious about interpreting transferable \"truth\" directions as objective-correctness readouts.\n\nWhat's new is the 2x2 factorial conflict design that separates SJ-associated and OC-associated components, plus the demonstration across four instruction-tuned models, two free-response domains, and MMLU/TruthfulQA with zero target-fitting that only the SJ component transfers reliably. The control battery is unusually thorough: response-likelihood residualization, token-count matching, null directions, label shuffles, counterbalanced X/Y judgements, and a fixed layer window that avoids selecting on the dependent variable. Released code and processed activations make it reproducible.\n\nThe main soft spot is exactly where the stress-test note lands. The component labels Wmeta and Wtruth are interpreted as 2β v_SJ and 2α v_OC under the additive linear model of Eq. S1. If the encoding is nonlinear or the two directions aren't separable, those vectors are just fitted contrasts, not latent semantic directions. The paper discloses this in Scope, and it's not an internal contradiction, but the abstract's \"component preserves SJ-associated polarity\" is a claim about latent semantics that goes beyond what the factorial contrasts strictly license. The broader cautionary conclusion doesn't depend on that decomposition — the mass-mean control and OOD results carry it — so I'd call this a moderate interpretive caveat, not a fatal flaw.\n\nA smaller concern: the self-judgement threshold τ=0.7 and pass@8 pair selection define a high-confidence diagnostic subset. The paper is careful not to claim prevalence, so that's handled proportionately.\n\nWho's this for: anyone doing interpretability, safety, or truth-probing work with hidden-state readouts. It should change how we validate correctness probes. I'd give it a serious referee. My own verdict is closer to accept-with-revisions than accept as is — I'd want the linear-model caveat spelled out more prominently and perhaps a nonlinear probe comparison (e.g., an MLP probe on the same conflict set) to bound how much the asymmetry depends on the linear estimator.\n\nRecommendation: send it to peer review.","headline":"Conflict-based factorial design is a real contribution and the transfer asymmetry is robust enough to change validation practice; the component-level 'SJ polarity' claim leans on the linear model assumption, but the broader cautionary conclusion survives it.","tokens_in":32532,"tokens_out":1923,"would_cite":true,"duration_ms":19841,"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":"Correctness probes, tested on OC–SJ conflicts, follow the model's own judgement rather than objective correctness in every cross-domain condition.","keywords":["correctness probing","self-judgement","objective correctness","hidden-state readouts","conflict set","transferability","linear representations","large language models"],"falsifier":"Find a model or domain where, on the same B/C conflict set, the OC-associated direction Wtruth transfers above chance while the SJ-associated direction Wmeta does not, with confidence intervals excluding the paper's asymmetry. Alternatively, show that controlling for a specific confound (e.g., response verbosity or topic) reverses the B/C ordering of Wmix, or that a causal intervention that changes self-judgement while leaving objective correctness fixed fails to move the probe's ranking.","tokens_in":31638,"feed_emoji":"🧠","tokens_out":3743,"duration_ms":34842,"temperature":0.7,"pith_summary":"The paper asks what a correctness probe actually reads out of a language model's hidden states when 'is the answer right?' and 'does the model think it's right?' disagree. It constructs conflict cases — objectively correct answers the model rejects versus wrong answers it endorses — and finds that conventional correctness-labelled readouts rank the wrong-but-endorsed answers higher, following self-judgement rather than objective correctness. After factorially separating the two variables, the self-judgement-associated direction transfers across math and factual recall domains for every tested model, while the objective-correctness-associated direction does not. The authors conclude that transferability alone does not establish that a probe encodes objective correctness; the transferable component largely preserves the model's own evaluation.","feed_headline":"Transferable 'truth' probes follow the model's self-judgement","feed_subtitle":"When correctness and confidence conflict, probes rank wrongly endorsed answers above correct ones across four LLMs.","key_machinery":"The load-bearing object is a factorial decomposition of the conventional 'correct minus incorrect' contrast. With cell means mu_A (correct, endorsed), mu_B (correct, rejected), mu_C (wrong, endorsed), mu_D (wrong, rejected), the paper defines Wmix = mu_A - mu_D, Wmeta = [(mu_A - mu_B)+(mu_C - mu_D)]/2, and Wtruth = [(mu_A - mu_C)+(mu_B - mu_D)]/2, and interprets them under an additive linear model of the answer-token residual stream in which OC and SJ each contribute a direction and the interaction cancels. The conflict set B/C provides the test: an OC readout should rank B above C, an SJ readout should rank C above B. These directions are fitted on one domain and scored on held-out or cross","core_discovery":"The paper's central claim is that when objective correctness (OC) and self-judgement (SJ) are separated, the component of a correctness probe that transfers across domains preserves SJ-associated polarity, not OC-associated polarity. On conflict cases, the conventional mixed contrast Wmix ranks wrong/self-endorsed responses above correct/self-rejected responses, and the OC-only mass-mean control does the same in all eight cross-domain conditions. The factorial contrast Wmeta (SJ-associated) predicts held-out SJ above chance within domain and across domains in all eight model-by-direction evaluations, whereas Wtruth (OC-associated) is at or below chance for OC ordering. This pattern develops","pith_inferences":["A natural extension is to test the same factorial conflict diagnostic on larger or non-instruction-tuned models; the paper only covers up to 14B instruction-tuned models, so it is an open question whether the asymmetry holds elsewhere.","If the SJ signal is largely endorsement or commitment, then interventions that shift a model's self-judgement (e.g., prompt-induced confidence) might move the probe's ranking of B versus C without changing objective correctness — a testable prediction that would sharpen the causal reading.","The results suggest re-examining published 'truth direction' findings: without conflict-based validation, those directions may have been tracking confidence all along."],"forward_implications":["If a correctness probe transfers across tasks, the transfer may reflect the model's confidence or endorsement rather than objective truth; transferability alone does not validate a truth readout.","Conflict cases (correct-but-rejected vs wrong-but-endorsed) provide a practical diagnostic for deciding what a probe tracks.","The SJ-associated direction is the robust transferable component across four models and two source domains, including zero-target-fitting transfer to MMLU and TruthfulQA.","The distinction between OC and SJ is operationalized by a separate Yes/No judgement; the probe's answer-token signal predictably aligns with that later judgement.","The observed asymmetry develops in middle-to-late transformer layers and persists under controls for answer likelihood, sequence length, token count, and null directions."],"fun_headline_variants":["Correctness probes actually track self-judgement, not truth","Transferable probe signal follows the model's own confidence","Probe transferability reveals self-judgement, not objective correctness","When confidence and truth clash, probes side with confidence","Model 'truth' probes are really self-judgement proxies"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The interpretation rests on the assumption that the answer-token residual stream is a linear additive mixture of separate OC and SJ directions, with the interaction cancelling in factorial contrasts; if the true encoding is nonlinear or confounded with other variables like answer style, the 'OC-associated' and 'SJ-associated' labels are fitted contrasts rather than true latent directions.","fun_headline_variants_meta":{"raw":{"variants":["Correctness probes actually track self-judgement, not truth","Transferable probe signal follows the model's own confidence","Probe transferability reveals self-judgement, not objective correctness","When confidence and truth clash, probes side with confidence","Model 'truth' probes are really self-judgement proxies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000166,"raw_usage":{"total_tokens":1083,"prompt_tokens":727,"completion_tokens":356,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":471,"completion_tokens_details":{"reasoning_tokens":290}},"tokens_in":471,"tokens_out":356,"duration_ms":3682,"temperature":1.0,"reasoning_tokens":290,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T19:54:08.113605+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find a model or domain where, on the same B/C conflict set, the OC-associated direction Wtruth transfers above chance while the SJ-associated direction Wmeta does not, with confidence intervals excluding the paper's asymmetry. Alternatively, show that controlling for a specific confound (e.g., response verbosity or topic) reverses the B/C ordering of Wmix, or that a causal intervention that changes self-judgement while leaving objective correctness fixed fails to move the probe's ranking.","supporting_citations":[],"review_version":1}