{"id":"cc53ac16-8674-4cef-88ce-3212ebfc079b","arxiv_id":"2508.16109","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"GPT-2 small performs syllogisms through truth-copying attention heads and a suppression-plus-MLP pathway that can output a negated truth value.","lead":"This paper reverse-engineers how GPT-2 small solves simple logical syllogisms, finding attention heads that copy truth values and heads plus MLPs that negate them. It maps these components onto previously known IOI circuits and claims a small circuit can recover most of the model's behavior.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Negative Truth Head / MLP mechanism rests on an unspecified path-patching corruption; the claimed suppression-then-promotion pathway could be an artifact of the corruption design, and the reported faithfulness percentages are undefined.","rationale":"The reader's weakest assumption—that path patching's corrupted distribution breaks only task-relevant behavior and is never validated—is indeed load-bearing. I agree that this is the main methodological vulnerability: the central claim about negation depends entirely on which heads and MLPs are selected via patching, and the paper neither specifies pnew nor validates the procedure with synthetic circuits or baselines. I add two concrete sharpening points. First, the natural corruption choice (flipping the premise truth value) can create the appearance of a negation mechanism from a mere copy head, because flipping the premise simultaneously flips the label; the reported SS/OS reversal of negative heads is exactly the kind of sign pattern this corruption would produce. Second, the faithfulness metric is internally inconsistent: the definition says lower is better, while the abstract and introduction report percentages that imply a ratio or recovery interpretation, and the one stated ALD comparison for CSS actually exceeds full-model performance. This makes the sufficiency claim unquantifiable as written. The circuit transfer experiments to other binary pairs are a useful positive sign, and the qualitative mechanism is plausible, but they are evaluated on the same selection task and do not resolve the corruption-design problem. The existing CONDITIONAL verdict is appropriate: the concern is specific and testable, not a demonstrated falsehood. I would keep the verdict unchanged pending the corruption-robustness check.","tokens_in":16340,"tokens_out":6800,"duration_ms":80549,"concrete_test":"Re-run the full component-selection and faithfulness pipeline for SS, OS, and CS under at least three explicit corruptions: (1) flipping the premise truth value, (2) random permutation of statement letters, (3) substituting the truth tokens with an unrelated binary pair (e.g., 'cat/dog'). For each, report the top-k heads/MLPs, the SS-vs-OS sign reversal for head 10.7, and per-prompt accuracy/logit-difference of COS on held-out templates, not just ALD. Also run a null control by applying the same selection+faithfulness pipeline to random head subsets (or a frozen model with head outputs shuffled) and report the distribution of faithfulness scores. If the identified head set, the reversal, or the 85-90% faithfulness number changes materially across corruptions or is within the null distribution, the negation mechanism is not supported. Release code and dataset so pnew can be audited.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Path patching is the sole causal method used to identify the Negative Truth Heads (7.3, 8.8, 8.10, 9.7, 10.7) and the Truth Logit Rescaler MLPs, and the COS faithfulness result is the only sufficiency evidence. The paper never specifies the corrupted distribution pnew for SS/OS/CS, nor does it release code or data. This matters because the natural choice—flip the premise truth value—keeps the same token pair in distribution and makes any head that merely copies the attended truth token appear 'important' on both tasks: its patching sign flips only because the definition of the correct token flips, not because it implements negation. Conversely, corrupting by letter permutation would destroy the logical relation and inflate the role of statement-identifier heads. The claimed SS/OS sign reversal of head 10.7 is therefore not identifiable without the exact counterfactual distribution, and no synthetic-circuit or random-baseline validation is provided. The faithfulness numbers compound this: the metric is defined as |ALD(M)-ALD(C)| (lower=better), yet the paper reports 'over 90%' and '~85% faithfulness'; the one in-text number (CSS ALD 1.9286 vs. full-model 1.8575) gives 103.8% by a ratio interpretation and 0.071 by the defined metric, so the percentages are not defined. Since the same logit-difference metric both selects the components and evaluates faithfulness, the sufficiency claim is partly circular.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a mechanistic interpretability analysis of GPT-2 small on three syllogism tasks: Simple Syllogism (SS), Opposite Syllogism (OS), and Complex Syllogism (CS). The authors use path patching and logit lens to identify circuits they call Truth Heads, Negative Truth Heads, Truth Logit Rescaler MLPs, and Truth Modulation Heads. The central novel claim is that in OS, attention heads suppress the truth token present in the prompt while MLPs rescale the residual stream to promote the opposite truth token, enabling the model to output a token not present in the prompt. The authors report faithfulness percentages for the identified circuits, claim transfer to other binary token pairs (right/wrong, good/bad, positive/negative, correct/incorrect), and provide exploratory results on larger models.","tokens_in":16737,"tokens_out":4531,"duration_ms":50978,"significance":"If the central claims hold, this work would extend circuit analysis beyond IOI and provide evidence that copy-suppression heads can act as binary negators, with a concrete role for MLPs in logical negation. The cross-model and cross-token-pair transfer experiments are useful and give the paper a broader scope than many single-task circuit analyses. However, the current evidence is not sufficient to support the headline quantitative claims: the faithfulness metric is undefined as reported, the path-patching corruption distribution is not specified, and the causal mechanism for negation is only correlational. These issues are load-bearing because the paper's main contributions are the identified circuits and the negation mechanism.","major_comments":[{"comment":"The faithfulness metric is defined as |ALD(M) − ALD(C)| with 'lower value indicates faithful recovery', yet the paper repeatedly reports percentages ('over 90%', 'roughly 85%'). In Section 3.1, the only concrete numbers are CSS ALD 1.9286 vs. full-model 1.8575. The defined metric gives |1.8575 − 1.9286| = 0.0711, not 90%; a ratio interpretation gives 103.8%, which is meaningless for a fidelity measure. The percentages are therefore undefined. This undermines the sufficiency claims in the abstract, introduction, and Section 3.2. Please specify the exact normalization or replace the metric with one that yields interpretable percentages.","section":"Section 3, Faithfulness definition; Abstract; Section 3.2"},{"comment":"The corrupted distribution pnew is never specified for SS, OS, or CS. The paper only says it is 'designed to break task-relevant behavior'. The identification of Negative Truth Heads and the claimed SS/OS sign reversal of head 10.7 depend entirely on what corruption is used. A natural corruption that flips the premise truth value keeps the same token pair in distribution, so a head that merely copies the attended truth token will appear important in both tasks; a letter-permutation corruption would destroy the logical relation and change the patching results. Without pnew, the causal attribution is not reproducible and the central mechanism cannot be assessed. Please provide the exact corruption scheme, code/data, and ideally a synthetic-circuit or random-baseline validation.","section":"Section 2, Path Patching; Sections 3.1–3.3"},{"comment":"The suppression-then-promotion pathway is supported only by QK attention patterns and OV/logit-lens projections (Table 3). Path patching shows that certain MLPs are important, but there is no direct causal intervention demonstrating that the MLP output specifically increases the logit of the opposite token. The claim that 'Truth Logit Rescaler MLPs' promote the negation is therefore correlational. A targeted experiment—for example, patching the MLP output with a value that cancels the suppression, or applying logit lens immediately after the MLP with and without the attention head active—would provide the needed causal evidence.","section":"Section 3.2, 'Mechanistic Interpretation' and Table 3"},{"comment":"The circuit selection and the faithfulness evaluation use the same logit-difference metric on the same tasks: heads are chosen by path-patching importance in LD, and faithfulness is then measured by |ΔLD|. This selection-evaluation overlap inflates apparent sufficiency. The transfer experiments in Appendix B partially address this, but the results are mixed: for Positive/Negative, the COS circuit gives ALD 1.0113 vs. full model 3.0032, a large discrepancy that is not consistent with 'often match or even outperform'. Please add held-out evaluation (different binary pairs, accuracy, or KL divergence) and compare against random circuits or ablations.","section":"Section 3.1–3.2 and Appendix B"}],"minor_comments":[{"comment":"The abstract states that 'a circuit comprising five attention heads achieves over 90% of the original model's performance', while the introduction says three heads reach 90% on SS and five heads plus four MLPs reach about 85% on OS. Please align the claims to avoid contradiction.","section":"Abstract vs. Introduction"},{"comment":"The caption says 'Top 3 highest-scoring query–key token pairs' but does not define how the scores are computed or averaged across prompts. Please clarify the exact quantity reported.","section":"Table 2"},{"comment":"The colorbar indicates 'Logit diff. variation' as percentages, but the computation (normalization by baseline? standard deviation?) is not explained in the text or captions. Please add captions describing how the percentages are derived.","section":"Figures 2 and 3"},{"comment":"The text refers to 'logit contribution along the true direction' while the figure labels say 'Projection along the truth embedding' and 'Dot w Truth Embed'. Please use consistent terminology and clarify whether the projection is onto the unembedding vector or the embedding vector.","section":"Section 3.3 and Figure 4"},{"comment":"The appendix states that 'in all subsequent experiments analyzing MLP effects, we report results with attention paths patched in', but the main text Figures 2a and 3a do not note this. Please add a note in the main text so readers know the MLP effects are measured with attention context restored.","section":"Appendix C"}],"recommendation":"major_revision","confidential_remarks":"The central claim is potentially interesting but the current version lacks the methodological detail needed to evaluate it. The authors should provide the exact corruption distributions, release code/data, and fix the faithfulness metric. The mixed transfer results also need more honest reporting. I recommend major revision rather than rejection because the issues appear fixable within the paper's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper has a genuinely interesting idea—that copy-suppression heads act as binary negators in syllogisms—but the quantitative support is currently too loose to trust, and the key path-patching corruption is never specified. Worth engaging with, but not close to ready as is.\n\nWhat's new: the truth-valued syllogism setup is fresh, and the proposed two-step negation mechanism (attention suppression followed by MLP rescaling) goes beyond the copy-suppression literature. The transfer tests to other binary pairs (good/bad, right/wrong) and the larger-model checks are real positives; they give the qualitative story some independent grounding.\n\nNow the soft spots, in order of severity. First, the faithfulness metric is undefined. It's defined as |ALD(M) - ALD(C)|, lower-is-better, yet reported as percentages. The one in-text number (CSS ALD 1.9286 vs full-model 1.8575) gives 103.8% by a ratio interpretation and 0.071 by the paper's own definition. That's not a rounding issue—it's a missing definition. Second, the abstract says five heads achieve over 90%, while the body says three heads on SS and five heads plus four MLPs on OS at about 85%. Those are inconsistent. Third, no error bars and no code or data release, which makes the numbers uncheckable.\n\nMost load-bearing: the corrupted distribution for path patching is never specified. The stress-test note is on target. If the corruption flips the premise truth value, then any head that copies the attended truth token will look important on both SS and OS, because the definition of the correct token flips. If it permutes letters, you destroy the logical relation and inflate the role of statement-identifier heads. Either way, the claimed sign reversal of head 10.7 is not identifiable without knowing p_new. That gap undermines the central causal claim. The selection-to-evaluation circularity (same tasks and metric used for both) worsens it, though the transfer experiments partially mitigate the circularity.\n\nTo be fair, the qualitative mechanism is plausible and the paper is honest about its exploratory limits; Appendix D appropriately frames the cross-model results as empirical and preliminary. No formal or reproducible artifacts here, so the credibility rests on the causal methodology, which is currently under-specified.\n\nFor the mechanistic interpretability audience, this is a useful hypothesis-generating paper. A serious referee should engage with it rather than desk-reject, but the revision must be substantial: define the corruption distribution, fix the faithfulness metric, add error bars, and release code/data. As it stands, I'd be skeptical of the specific numbers but sympathetic to the qualitative story.","headline":"Plausible negation story in GPT-2 small, but the faithfulness numbers are undefined and the path-patching corruption is left unspecified; worth refereeing, but major revision needed.","tokens_in":17169,"tokens_out":2510,"would_cite":false,"duration_ms":27658,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GPT-2 small solves opposite syllogisms by suppressing the prompt's truth token and using that suppression to boost the opposite token.","keywords":["mechanistic interpretability","circuit discovery","syllogistic reasoning","logical negation","binary truth values","GPT-2","path patching","attention heads"],"falsifier":"Mean-ablating head 10.7 on Opposite Syllogism prompts should eliminate the true→false flips while leaving false→true cases largely intact; if both directions change equally, then suppression of the prompt token is not the specific signal driving negation.","tokens_in":16270,"feed_emoji":"🔁","tokens_out":7888,"duration_ms":82448,"temperature":0.7,"pith_summary":"This paper tries to show that GPT-2 small does not merely copy or retrieve truth values; it can actually negate them, and does so by a specific two-step circuit. On 'opposite syllogism' prompts, where the correct answer never appears in the input, the model's Negative Truth Heads attend to the stated truth value and suppress its logit, while associated MLPs rescale the residual stream to push the opposite token to the top. The authors identify the same style of circuit for simple syllogisms, where three Truth Heads alone recover about 90% of full-model performance, and for complex syllogisms, where redundant inhibition and reinforcement heads implement the same binary operation through multiple pathways. Why this matters: it extends mechanistic interpretability from copying tasks like Indirect Object Identification to logical reasoning, and it gives a concrete, testable account of how a transformer can output a token it never saw in the prompt. A five-head, four-MLP circuit reaches about 85% faithfulness on the opposite task.","feed_headline":"GPT-2 flips truth values by suppressing, then boosting the opposite","feed_subtitle":"A five-head circuit recovers 85% of the model on opposite syllogisms, turning a known copy-suppression head into a negator.","key_machinery":"The load-bearing object is the attention–MLP negation pathway: a Negative Truth Head's QK circuit locates the prompt's truth token, its OV projection suppresses that token's logit, and the paired Truth Logit Rescaler MLP rescales the residual stream so the opposite token's logit rises. Path patching on logit difference isolates each component; the faithfulness metric (absolute difference between full-model and circuit average logit difference) measures how much of the behavior is preserved.","core_discovery":"GPT-2 small performs negation in the Opposite Syllogism task through an attention–MLP pathway: Negative Truth Heads (7.3, 8.10, 9.7, 10.7, 11.10) attend to the prompt's truth token, their OV circuits push that token to the bottom logits, and paired Truth Logit Rescaler MLPs shift the residual stream so the opposite token rises. Suppressing the wrong token is thus the signal for the correct, unseen token. In Simple Syllogisms, Truth Heads (7.2, 9.1, 9.9, 10.1, 10.4) copy the stated value, a three-head circuit reaching about 90% faithfulness. In Complex Syllogisms, inhibition and reinforcement heads operate on the incorrect token redundantly with the Negative Truth Heads (about 85% faithfulnes","pith_inferences":["If suppression-as-negation is a general mechanism, then any binary antonym pair in the vocabulary could be flipped by steering a copy-suppression head's OV direction; a direct test is whether mean-ablating head 10.7 on the opposite syllogism specifically hurts false predictions more than true ones.","The paper's causal accounting leans on one choice of corrupted prompt distribution; testing alternative corruptions (for example, shuffling token order rather than replacing the truth value) would show whether the identified circuit roles are robust.","The true→false asymmetry hints at a corpus-level frequency prior; measuring logit differences on balanced true/false prompts across GPT-2 checkpoints could reveal when the bias emerges during training.","Larger models show weaker syllogism performance and muted negative-head effects, so transplanting the five-head negation circuit into those models via activation patching would test whether the pathway is architecture-specific or transferable."],"forward_implications":["A three-head circuit (7.2, 9.1, 9.9, 10.1, 10.4) recovers roughly 90% of GPT-2 small's behavior on simple syllogisms, showing the task needs almost no MLP support.","On opposite syllogisms, the five Negative Truth Heads plus their MLPs recover about 85% of full-model performance, so the negation pathway is nearly sufficient by itself.","Heads that hurt the simple task—like 10.7—become causally important on the opposite task, so component importance is task-relative, not fixed.","The same circuits transfer to other binary pairs (good/bad, right/wrong, positive/negative, correct/incorrect), often matching or beating the full model, so the mechanism is not tied to the literal tokens 'true' and 'false'.","Negation is asymmetric: true→false is more reliable than false→true, and the truth-token embeddings show 'false' has larger projections, suggesting a learned prior toward false."],"supporting_citations":[{"why":"Supplies the path-patching method and the IOI circuit (duplicate token, S-inhibition, name mover heads) that this paper extends to syllogisms and negation.","marker":"[Wang et al., 2022]"},{"why":"Provides the transformer-circuits notation (QK/OV decomposition, residual stream) used to define heads and analyze attention patterns.","marker":"[Elhage et al., 2021]"},{"why":"Establishes head 10.7 as a copy-suppression/negative head; the paper shows that in binary settings this suppression becomes negation of the attended token.","marker":"[McDougall et al., 2023]"},{"why":"Supplies the logit lens used to inspect what head outputs and MLP outputs contribute to the truth-token logits.","marker":"[Nostalgebraist, 2020]"},{"why":"Provides the hydra-effect account the paper uses to explain why the complex-syllogism circuit has redundant independent pathways.","marker":"[McGrath et al., 2023]"}],"fun_headline_variants":["GPT-2 negates by suppressing input, boosting opposite","Five attention heads flip truth in GPT-2 syllogisms","Inside GPT-2's negation: heads suppress, MLPs boost","How GPT-2 says 'no': a circuit for opposite logic","GPT-2's logical flip: suppress token, boost alternative"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The whole identification of Truth Heads, Negative Truth Heads, and their MLP partners rests on path patching—swapping a component's activations with those from a corrupted prompt and treating the drop in the model's preference for the right answer as that component's causal contribution—and the paper does not separately validate that this corruption-to-drop relationship cleanly isolates causal roles.","fun_headline_variants_meta":{"raw":{"variants":["GPT-2 negates by suppressing input, boosting opposite","Five attention heads flip truth in GPT-2 syllogisms","Inside GPT-2's negation: heads suppress, MLPs boost","How GPT-2 says 'no': a circuit for opposite logic","GPT-2's logical flip: suppress token, boost alternative"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000918,"raw_usage":{"total_tokens":3804,"prompt_tokens":797,"completion_tokens":3007,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":2920}},"tokens_in":541,"tokens_out":3007,"duration_ms":27964,"temperature":1.0,"reasoning_tokens":2920,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:31:08.796965+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Mean-ablating head 10.7 on Opposite Syllogism prompts should eliminate the true→false flips while leaving false→true cases largely intact; if both directions change equally, then suppression of the prompt token is not the specific signal driving negation.","supporting_citations":[{"cited_title":"Interpreting gpt: The logit lens, 2020","cited_arxiv_id":null,"evidence_quote":"Supplies the logit lens used to inspect what head outputs and MLP outputs contribute to the truth-token logits."}],"review_version":1}