{"id":"d9f08ab1-117a-4625-b974-8281574de293","arxiv_id":"2608.13337","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"In ablation-based SAE evaluation, the measurement token is selected by the dictionary under test, and holding that token fixed collapses most of the variance that is usually attributed to differences between dictionaries.","lead":"Ablation-based tests of sparse autoencoder features in language models usually measure the effect at the token where the feature fires hardest, but that token is chosen by the dictionary being tested, not by the experimenter. This paper shows that this choice accounts for most of the apparent disagreement between dictionaries, and gives a one-line fix for evaluation code.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The controlled repair asserts, without measuring it, that shared initialisation keeps latent i identical across six arms that differ in sparsity, learning rate, and decoder freezing; if decoder directions drift, the variance attributed to position is partly identity drift.","rationale":"The reader's weakest-assumption call is the right one, and it is load-bearing rather than a routine disclaimer. The central quantitative claim is not that dictionaries select different positions, which is established on released Gemma Scope pairs and is robust, but that holding position fixed removes almost all latent×arm variance, which is what turns the artifact into a correction. That inference requires latent i to denote the same feature in all six arms. Shared initialisation makes this plausible but does not demonstrate it, and the arms are deliberately chosen to differ in ways that can move features around. The paper is unusually transparent: it pre-registers decision rules, reports the non-selective union-rule result (Appendix C), and flags the structural constraint in Section 7. Those disclosures mean the concern is not that the authors hide the assumption; it is that the headline strength of the claim currently exceeds the evidence. Appendix C reinforces the concern: on Gemma-3-1B the collapse under the non-selective union rule is only to 6.3%, and the paired gain is not distinguishable from zero, so even within the paper's own numbers the magnitude of the repair is rule-dependent on one base model. A cheap alignment check on the existing artifacts would settle whether the identity assumption holds; until then, acceptance should be conditional rather than unconditional. I would not reject, because the premise, the audit, and the qualitative direction of the repair are solid and independently supported.","tokens_in":1928,"tokens_out":3924,"duration_ms":145246,"concrete_test":"Using the released evaluation code, compute for each of the 240 shared latents the six pairwise cosine similarities between trained decoder columns (or Spearman correlations of the arms' activation vectors over the 384-sequence corpus), and report the median and the lower decile. Then recompute Table 1's per-arm and shared variance components restricted to latents with minimum pairwise decoder cosine above 0.8, or another pre-specified threshold. If the 7.6 to 0.0 and 11.9 to 2.4 collapses are unchanged on the high-alignment subset, the identity assumption is not the driver; if the per-arm component shrinks or the shared component rises, latent drift is confounded with the position effect and the headline numbers need revision. This check runs in an afternoon on the authors' own artifacts and would settle the attribution.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The six-arm design's central attribution (§4, Table 1) depends on the claim that because all arms share seed 0, latent i denotes the same thing in every arm (§3). Shared initialisation fixes the starting direction, but nothing freezes the trained decoder column: the arms deliberately vary sparsity k=41 vs 82, a 10× lower learning rate, and soft-frozen decoders, and any of these can rotate a column or split or absorb a feature (Leask et al., 2025; Chanin et al., 2024, both cited). The paper cites Paulo and Belrose (2025) for the result that same-data different-seed dictionaries share only about 30% of features, but reports no post-training overlap, decoder-cosine, or activation-correlation check across its own six arms. If latent i in the k=41 arm is not the same feature as latent i in the k=82 arm, then the per-arm latent×arm component mixes identity drift with position choice, and the collapse from 7.6%/11.9% to 0.0%/2.4% is not attributable to position alone. Section 7 calls this structural, and the lower-bound claim that shared-init arms are the most similar dictionaries a practitioner would compare (§3) is likewise untested. This is the load-bearing joint of the headline result; the released-dictionary premise does not rescue it, because it is the controlled repair that turns disagreement about position into a claim about what carries the variance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies a previously unexamined degree of freedom in ablation-based sparse autoencoder evaluation: the token position at which a latent's causal effect is measured. Using released Gemma Scope dictionaries, it shows that matched latents often disagree about the maximally-activating token, and that even near-duplicate decoders agree only about 60% of the time. In a controlled six-arm design trained from a shared initialization, the latent×arm variance component falls from 7.6%/11.9% to 0.0%/2.4% when all arms are measured at a common position rather than each arm's own top position. The paper further shows that more evaluation data does not reduce the disagreement, documents two other reporting conventions that can flip sign, and proposes a protocol for reporting ablation-based causal numbers.","tokens_in":22306,"tokens_out":6448,"duration_ms":66692,"significance":"The strength of the paper is its two-pronged design: the premise that production dictionaries select different positions is tested directly on released dictionaries, and the controlled repair is compared under paired bootstrap CIs with an alternative shared-position rule, a pre-registered decision rule, and an explicit discussion of filter bias and clipping. The authors are also admirably explicit that the Eρ² gain is an arithmetic consequence of the variance-component collapse, not independent evidence. If the controlled attribution to position holds, this identifies a convention that is invisible in current practice yet moves the answer more than the dictionary choice, which would make the paper a useful methodological contribution to SAE evaluation and to interpretability more generally.","major_comments":[{"comment":"The central attribution of the collapse to position rests on the assertion that shared initialisation makes latent i denote the same thing in every arm, but the paper never measures this after training. The six arms differ in sparsity (k=41 vs 82), learning rate (10× lower), and decoder freezing, and the paper itself cites Leask et al. (2025) and Chanin et al. (2024) showing that such fitting choices split and absorb features. Shared initialisation fixes only the starting decoder direction; it does not prevent trained decoder columns from rotating or drifting. Without a post-training identity check (for example, median decoder cosine between matched columns, activation overlap at shared positions, or matched-feature agreement), the latent×arm component in Table 1 mixes identity drift with position choice, and the 7.6%→0.0% / 11.9%→2.4% collapse cannot be attributed to position alone. Section 7 calls this structural, but the manuscript needs a direct empirical check before the headline attribution is supported.","section":"§3–§4, Table 1"},{"comment":"The claim that the six shared-initialisation arms are 'the most similar dictionaries a practitioner would ever compare' and hence that the measured instability is a lower bound is not tested. It presupposes that fitting choices (sparsity, learning rate, decoder freezing) induce less feature drift than seed variation, which is exactly what the paper's own cited literature (Paulo and Belrose, 2025) leaves open. This claim should either be verified with a cross-seed comparison or removed.","section":"§3"}],"minor_comments":[{"comment":"In panel (b), the y-axis is labelled '% same token', but the panel plots the latent×arm variance component; the label appears to be a copy-paste error from panel (a) and should be corrected.","section":"Figure 2"},{"comment":"The caption says that '—' means the sections reviewed do not state the convention, but the symbols in the table body are not rendered consistently; please define the exact mark used in the table itself.","section":"Table 2"},{"comment":"The reproducibility discrepancy with Cho et al. (0.41–0.44× their prevalence at four times the apparent corpus size) is reported without the detector configuration or the comparison protocol; if it is not moved to an appendix with full details, it reads as an unverifiable aside.","section":"§2"},{"comment":"The statement that 'the problem grows with scale' should carry the scope qualifier that the growth is measured on two Gemma models at evaluation corpora of 96–1536 sequences, not established for arbitrary scales.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper is unusually transparent and well-executed, and I would be inclined to accept once the identity check is added. The main unresolved point is empirical rather than conceptual: shared initialisation is a necessary condition for the controlled design, but the manuscript does not establish that it is sufficient. I do not think this rises to rejection, because the released-dictionary premise is direct and the repair can be tested with the authors' existing pipeline, but the load-bearing attribution in Table 1 needs the additional evidence before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is one of the few papers in SAE evaluation that treats a reporting convention as an experimental variable and actually measures its effect. Second, the headline result is believable: on both released Gemma Scope dictionaries and on six self-trained arms, fixing the measurement position collapses most of the latent-by-arm variance that would otherwise be read as dictionary disagreement.\n\nThe genuinely new piece is the observation that the measurement position is not a neutral experimental choice: it is selected by the dictionary under evaluation, because the top-activating token is computed from the dictionary's own gate. That makes it a sampled facet in the generalizability-theory sense, and the paper demonstrates quantitatively that it dominates the comparison. The released-dictionary evidence (30–48% agreement across pairs, only 60% even for near-duplicates) is strong and independent of their own training. The controlled six-arm design is careful: shared seed, crossed latent×arm structure, paired bootstrap intervals, two different shared-position rules, and a filter whose direction of bias is measured and goes against the claim. The reproducibility statement is unusually honest, including pipeline defects and the retracted earlier conclusion. The literature audit is scoped properly and does not overclaim what the five papers did.\n\nThe main soft spot is exactly the one flagged in the stress-test note: shared initialisation is asserted, not verified, to make latent i the same thing across arms. The arms deliberately differ in sparsity, learning rate, and decoder freezing, all of which can rotate, split, or absorb features. The paper says the shared-init arms are the most similar dictionaries a practitioner would compare, but it does not report post-training decoder cosine or activation correlation across the six arms. If the index correspondence drifts, the per-arm latent×arm term mixes identity drift with position choice, and the collapse to ~0 may be partly rule-induced rather than purely positional. This is a real gap, though not a fatal one: the released-dictionary premise does not depend on it, and the alternative union rule keeps the collapse on Gemma-2 and roughly halves it on Gemma-3, so the conclusion is not a single fragile estimate. I would like to see a direct identity check (decoder cosine matrix across arms, or overlap of top activating sets) added; it may be that the shared-init arms are indeed close, but 'likely' is not the same as 'shown'.\n\nMinor points: only two Gemma models, 12M-token training, and the retained-subpopulation caveat are all acknowledged and argued to bias conservatively. That is fair. The unexplained 67.4% within-cell variance is large, but that is a strength of the analysis, not a weakness.\n\nWho is this for? Anyone reporting or consuming single-latent ablation numbers in SAE interpretability. This should go to peer review; it deserves a serious referee and likely acceptance after the identity check is addressed.","headline":"A careful, reproducible demonstration that the top-activating-token convention is a hidden, dictionary-chosen measurement facet; the central claim is solid, but the controlled repair would be stronger with a direct check that shared initialisation preserves latent identity post-training.","tokens_in":22854,"tokens_out":3189,"would_cite":true,"duration_ms":33361,"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":"An unreported choice of which token to measure, not the dictionaries themselves, drives most apparent disagreement in sparse-autoencoder ablation comparisons.","keywords":["sparse autoencoders","interpretability","ablation-based evaluation","measurement position","variance decomposition","generalizability theory","causal effect","evaluation protocol"],"falsifier":"Run the paper's own recipe on any two sparse autoencoders for the same base model: record each dictionary's top-activating token for matched latents. If the two pick the same token for nearly every latent, the claimed position artifact is absent. Then run the six-arm shared-initialisation design at a production-scale token budget; if the latent-by-arm variance does not fall when all arms are measured at a common position, the repair fails. Both can be checked in an afternoon on a single GPU.","tokens_in":21834,"feed_emoji":"📏","tokens_out":7464,"duration_ms":71342,"temperature":0.7,"pith_summary":"Sparse autoencoders are the current tool for naming what a language model computes, and the standard way to test whether a named latent matters is to zero it out and measure how the next-token output changes. The paper shows that this causal number depends on which token the latent is measured at, and that the choice of token is made by the dictionary under evaluation, not by the experimenter. Comparing two dictionaries under the usual top-activation convention therefore compares them at different tokens. In a controlled six-arm experiment holding the latent's identity fixed, the latent-by-dictionary variance that a comparison would attribute to the dictionaries fell from 7.6% and 11.9% to near zero once every dictionary was measured at a shared token. The paper concludes that an ablation-based causal number reported without its position describes the measurement token as much as it describes the latent, and that this problem grows with more evaluation data.","feed_headline":"Fix one token and AI-dictionary disagreements nearly vanish","feed_subtitle":"Ablation scores change with the token chosen for measurement; the usual convention lets the dictionary pick it.","key_machinery":"The central object is the measurement position: the token at which a latent's causal effect is read when it is zero-ablated, conventionally the token where the latent fires hardest. Because that token is computed from the dictionary's own activations, two dictionaries can disagree about a latent even when they encode it nearly identically. The paper's second piece of machinery is a crossed latent-by-arm experimental design with a shared initialisation seed: six autoencoders differ only in fitting choices, so any variance between arms is attributable rather than confounded, and a generalizability-theory decomposition separates variance into latent, arm, latent-by-arm interaction, and position-within-cell components, with the generalizability coefficient $E\\rho^2 = v_a/(v_a + v_{ab} + v_e/n)$ tracking how well a latent ranking would survive a different dictionary. Keeping the latent sample fixed and switching from each arm's own positions to a shared position set is the single flag that isolates the convention.","core_discovery":"The paper's central claim is that the effect of zero-ablating a sparse-autoencoder latent is not a property of the latent alone; it is a function of the token at which the intervention is applied, and that token is chosen by the dictionary's own activations. Across released dictionaries for the same base model, matched latents that are nearly identical still pick different top-activating tokens for a large share of pairs, and even the closest near-duplicate pairs agree only about 60% of the time. To separate the convention from the dictionaries, the paper trains six autoencoders from one shared initialisation so that a given latent denotes the same direction in every arm. Measured at each arm's own favourite position, the latent $\\times$ arm interaction is 7.6% and 11.9% of variance; measured at one common position, it collapses to 0.0% and 2.4%. The paper also reports that the position a latent is measured at accounts for 67.4% of the total variance, more than latent, arm, and their interaction combined, and that larger evaluation corpora increase disagreement about where to measure rather than reducing it.","pith_inferences":["The paper only tests two models from one family; if the position artifact persists across architectures, many existing single-latent ablation comparisons in the interpretability literature would need re-reading, not just the five audited papers.","A natural extension the paper gestures at is to report a distribution of effects over a latent's firing positions, or an activation-weighted expectation, instead of a scalar; which choice is most stable across dictionaries can be decided empirically on the same crossed design.","The audit's finding that none of five papers report position suggests a cheap standardization: requiring the position, and the number of positions per latent, in a protocol would let future causal numbers be pooled across papers.","The same controlled design could be applied to rate-based readouts to check whether the position artifact is specific to magnitude readouts or general across evaluation metrics."],"forward_implications":["A causal number reported without its measurement position cannot be compared across papers; the position is an uncontrolled confound.","More evaluation data does not stabilise the convention; across a sixteenfold corpus range, position agreement between arms falls monotonically.","Reported differences between dictionaries that disagree about a latent may be dominated by differences in where they chose to measure, not by what the latent does.","A one-line evaluation change, measuring every dictionary at a shared token, collapses most of the apparent latent-by-dictionary variance.","Within a latent, the choice of firing position explains 67.4% of variance, so scalar causal numbers for a latent are poorly defined; the effect is spread over positions."],"supporting_citations":[{"why":"Shows that seed alone changes which features are learned, motivating the shared-initialisation design that makes latent identity comparable across arms.","marker":"Paulo and Belrose (2025)"},{"why":"Supplies the released production dictionaries on which the position-agreement premise is checked.","marker":"Lieberum et al. (2024)"},{"why":"Defines the generalizability-theory variance decomposition and coefficient used to separate latent, arm, and position contributions.","marker":"Cronbach et al. (1972)"},{"why":"Names the treated-as-fixed sampled-facet fallacy that the measurement position instantiates.","marker":"Clark (1973)"},{"why":"Closest published methodology; its single-token detector and depth-correlation claims are compared against this paper's protocol.","marker":"Cho et al. (2026)"},{"why":"Documents the zero-ablation magnitude readout that the audited convention standardises.","marker":"Gao et al. (2024)"}],"fun_headline_variants":["The token you measure at decides your ablation score","SAE ablation scores depend on where you measure","Position choice, not the latent, drives SAE disagreement","One code line fixes SAE evaluation variance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that six autoencoders trained from one shared initialisation on 12M tokens make latent i denote the same thing in every arm, so that the controlled comparison can attribute all disagreement to measurement position, and that these 12M-token arms are representative enough of production-scale dictionaries; if shared initialisation does not fix functional identity, the attribution collapses.","fun_headline_variants_meta":{"raw":{"variants":["The token you measure at decides your ablation score","SAE ablation scores depend on where you measure","Position choice, not the latent, drives SAE disagreement","One code line fixes SAE evaluation variance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1461,"prompt_tokens":1114,"completion_tokens":347,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":730,"completion_tokens_details":{"reasoning_tokens":287}},"tokens_in":730,"tokens_out":347,"duration_ms":3813,"temperature":1.0,"reasoning_tokens":287,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:22:53.866515+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the paper's own recipe on any two sparse autoencoders for the same base model: record each dictionary's top-activating token for matched latents. If the two pick the same token for nearly every latent, the claimed position artifact is absent. Then run the six-arm shared-initialisation design at a production-scale token budget; if the latent-by-arm variance does not fall when all arms are measured at a common position, the repair fails. Both can be checked in an afternoon on a single GPU.","supporting_citations":[{"cited_title":"Are Single-Token Sparse Autoencoder Features Causally Necessary? Layer-Depth and SAE-Family Effects","cited_arxiv_id":"2607.20596","evidence_quote":"Closest published methodology; its single-token detector and depth-correlation claims are compared against this paper's protocol."}],"review_version":1}