{"id":"7a0c8418-bdd7-4bba-ab45-b3718a3595fa","arxiv_id":"2504.13151","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"MIB, a multi-task benchmark for mechanistic interpretability, reports that attribution and mask optimization methods excel at circuit localization, and DAS outperforms sparse autoencoder features at causal variable localization.","lead":"MIB is a new benchmark that tests how well mechanistic interpretability tools locate the internal components that drive language model behavior on a task. It compares these tools across four tasks and five models, finding that attribution and mask optimization methods win at circuit localization and DAS beats sparse autoencoders at causal variable localization.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Causal variable track scores presuppose that the high-level causal models are true abstractions; App. B admits these graphs may not exist in the evaluated models, so DAS-vs-SAE rankings may measure fit to a hypothesized model rather than localization of real mechanisms.","rationale":"I agree with the reader's weakest assumption: the causal variable localization track's validity rests on the high-level causal models being faithful abstractions of the networks, and the paper itself flags this in App. B. The concern is load-bearing because the headline findings about DAS and SAEs are drawn from this track; if the high-level models do not exist in the evaluated models, those findings describe how well methods can match a hypothesized counterfactual labeling scheme, not how well they localize real mechanisms. The paper deserves credit for including explicit limitations, a ground-truth InterpBench model for the circuit track, two carefully defined circuit metrics, and for reporting private test results for circuit localization. Those elements support the circuit-track claims. For the causal variable track, however, the absence of private test results and the admitted confound in the arithmetic XCarry analysis (App. F.5.2) leave the abstraction assumption unvalidated. The proposed check would settle this: if DAS generalizes to held-out multi-variable interventions and outperforms a permuted-model control, the abstraction assumption is supported; if not, the causal-variable findings should be treated as unverified. Because the reader already recommended conditional acceptance with these caveats, I keep the verdict unchanged rather than moving it.","tokens_in":40543,"tokens_out":9350,"duration_ms":96212,"concrete_test":"For MCQA or ARC, run the causal variable track with multi-variable interchange interventions (e.g., simultaneously intervening on XOrder and OAnswer) on a held-out set of counterfactual types not used to fit alignments, and compare DAS IIA against a null control using randomly permuted high-level variable mechanisms or random subspaces. If DAS IIA on held-out multi-variable interventions is near chance, or if the permuted-model control achieves comparable IIA, then the reported high scores are an artifact of fitting the counterfactual distribution and the track does not validate the hypothesized causal abstraction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that MIB enables meaningful comparisons of causal variable localization methods depends on the high-level causal models (HMCQA's XOrder/OAnswer, H+'s XCarry, HIOI, RA VEL) being faithful abstractions of the networks. If a variable is not actually implemented, IIA scores cannot be interpreted as evidence about real mechanisms; a sufficiently expressive featurizer such as DAS can still achieve high IIA by steering outputs to the counterfactual labels on the training distribution. The authors acknowledge this in App. B: 'this paradigm presumes the existence of the high-level model in the computation graph... graphs may not always exist in the models we evaluate in the exact forms shown here.' This is not merely hypothetical: App. F.5.2 reports that the initial XCarry alignment in Llama was 'likely a representation of the tens-digit output, which perfectly correlates with the ones-carry variable in these counterfactuals,' and that with random counterfactuals the method fails completely. That internal check shows the metric can be confounded. Because causal variable track results for MCQA/ARC/RAVEL are reported only on the counterfactual types used to fit alignments, and private test results are not reported, the DAS/SAE comparison may reflect overfitting to the specific counterfactual distribution rather than the presence of the hypothesized variables. The load-bearing condition is therefore the existence and validity of the high-level models; the benchmark does not independently establish this for any task.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MIB, a two-track benchmark for mechanistic interpretability. The circuit localization track introduces integrated circuit performance ratio (CPR) and circuit-model distance (CMD) metrics that marginalize over circuit size, and compares random, activation patching, attribution patching (EAP/NAP with and without integrated gradients), information flow routes, and uniform gradient sampling on IOI, arithmetic, MCQA, and ARC across Llama-3.1 8B, Gemma-2 2B, Qwen-2.5 0.5B, and GPT-2 Small, plus an InterpBench ground-truth circuit model. The causal variable localization track uses interchange intervention accuracy (IIA) to evaluate alignments of high-level causal variables (XCarry, XOrder, OAnswer, STok, SPos, and RAVEL attributes) to hidden-vector features, comparing DAS, DBM on raw dimensions, DBM on PCA and SAE features, and full-vector baselines. The paper's central claims are that attribution and mask optimization methods perform best on circuit localization and that supervised DAS performs best on causal variable localization, while SAE features are not better than raw neurons. The paper also describes public and private leaderboards for future submissions.","tokens_in":40858,"tokens_out":6490,"duration_ms":59793,"significance":"Conditional on the validity of the high-level causal models, MIB would be a valuable community resource: it spans four tasks and five model classes, provides standardized counterfactual datasets with public/private splits, introduces CPR/CMD metrics that avoid a circuit-size threshold, includes an InterpBench AUROC control for ground-truth circuits, and proposes two public leaderboards for cross-method comparison. The comparative baseline sweep is substantial, and the headline findings are falsifiable claims of the kind the field needs. However, the causal variable track's construct validity is the load-bearing issue: if the hypothesized variables are not implemented by the evaluated models, IIA-based rankings measure fit to a speculative abstraction rather than localization of real mechanisms. The lack of uncertainty quantification in the circuit localization tables and the absence of private-test results for the causal variable track further weaken the comparative claims as currently stated.","major_comments":[{"comment":"The causal variable track's interpretation presupposes that the high-level causal models H are faithful abstractions of the evaluated networks, and App. B itself states that 'these graphs may not always exist in the models we evaluate in the exact forms shown here.' The arithmetic results show that this is not a merely hypothetical concern: App. F.5.2 reports that the initial apparent XCarry alignment in Llama was 'likely a representation of the tens-digit output, which perfectly correlates with the ones-carry variable in these counterfactuals,' and that with random counterfactuals the method fails completely. Because IIA is computed only on counterfactual types correlated with the target variable, the DAS-versus-SAE ranking in Table 3(b), and by analogy the other causal variable tables, may measure fit to a hypothesized variable rather than localization of a mechanism the model actually implements. I ask the authors to add negative-control counterfactual sets for every task-variable pair (e.g., random counterfactuals that decorrelate the target variable from output-relevant features) and to report IIA on those controls, or to explicitly restrict the causal variable ranking claims to task-variable pairs that pass such controls.","section":"§4.1 and App. F.5.2"},{"comment":"The IOI causal model HIOI is fitted to the same model being evaluated: the paper states that after replicating Wang et al.'s experiments, the authors 'fit our own linear model' to GPT-2's logit differences, yielding the coefficients 0.048 + 2.005*PositionSignal + 0.768*TokenSignal. Evaluation of STok and SPos alignments against this fitted target is therefore partly circular: the 'high-level' target is a compressed description of the exact model and dataset used in Table 3(d), not an independent causal hypothesis. At minimum, the authors should specify whether the linear coefficients were fixed before benchmarking, and ideally evaluate on a held-out model or a pre-registered causal model. Without this, the MSE values in Table 3(d) primarily measure regression fit rather than the faithfulness of an independently specified abstraction.","section":"§4.6 and Figure 14"},{"comment":"The headline circuit-localization claims ('EAP-IG-inputs achieves the highest performance on average') are made without error bars, significance tests, or sensitivity analysis, despite differences as small as 0.00–0.07 between methods in several columns (e.g., Table 2, ARC (E) Gemma: EAP-IG-inputs CMD 0.14 versus EAP-IG-activations 0.13; IOI Qwen CMD 0.02 versus 0.01). Many entries appear to be single runs; only the random baseline is averaged over three seeds. This does not establish that the best method is distinguishable from close competitors. I ask the authors to add bootstrap confidence intervals over examples or seeds, report variance for all methods, and, where possible, run pairwise significance tests or at least report effect sizes with intervals.","section":"§3.3, Tables 2 and 14–16"},{"comment":"The benchmark filters out all examples where the model is incorrect on the base input or on any counterfactual input before computing IIA. This is a post-hoc selection that can inflate IIA and change method rankings, particularly in the low-data MCQA and ARC settings. App. F.5.3 explicitly states 'We filter out failure cases' for RAVEL after reporting 66.5–70.5% model accuracy. The paper should report results on the unfiltered set alongside the filtered set, or provide evidence that the filtering does not drive the DAS advantage over unsupervised featurizers. As written, the scores may reflect selection of easy examples rather than the quality of the localization methods.","section":"§4.1 and App. F.5.3"},{"comment":"No private-test results are reported for the causal variable localization track. The circuit localization track includes private-test CMD and CPR tables (App. D, Tables 15–16), but the causal variable track reports only Table 3, which is computed on the same public counterfactual distributions used to train DBM and DAS. The causal variable rankings may therefore reflect overfitting to the particular counterfactual types, and the paper's claim that MIB enables leaderboard evaluation on a private test set is not demonstrated for this track. Either add private-test causal variable results, or explicitly define held-out counterfactual types as the evaluation protocol and report those results.","section":"§2.4 and §4"}],"minor_comments":[{"comment":"In the paragraph introducing the metrics, there is a typo: 'we propose to to quantify CPR' should read 'we propose to quantify CPR'.","section":"§3.1"},{"comment":"The caption contains a typo: 'Webold and underline' should read 'We bold and underline'.","section":"Table 2 caption"},{"comment":"The paper refers to 'this HuggingFace URL', 'this GitHub URL', and 'this HuggingFace URL' as placeholders; the actual URLs should be included in the final version so that the promised datasets, code, and leaderboards are accessible.","section":"§2 and App. C.6"},{"comment":"The sentence 'The dataset contains many cities, so considering that there are many countries and languages to choose from, this is good performance' is awkwardly phrased and should be rewritten for clarity.","section":"App. F.5.3"},{"comment":"The notation table defines 'c' as 'An counterfactual input'; the article 'An' should be 'A'.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"I am sympathetic to the goals of this benchmark and to the candid limitations discussion in App. B. The main blocker is the causal variable track's construct validity: the authors' own arithmetic control experiment shows that IIA can be confounded by correlated output features. This is fixable by adding negative-control counterfactual sets and by reporting private-test or held-out-counterfactual results, so I am not recommending rejection. The lack of uncertainty quantification in the circuit localization tables is also a required revision. If the authors add these controls and re-verify the DAS-versus-SAE ranking, the benchmark would be a strong contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things up front. First, MIB is the most serious attempt yet to give mechanistic interpretability a shared evaluation standard, and it has a real chance of getting adopted the way GLUE was for NLP. Second, the headline empirical findings are real but softer than the abstract suggests, because the causal variable track assumes the high-level causal models are faithful abstractions and never independently validates that assumption.\n\nWhat is actually new: CPR and CMD, threshold-free faithfulness metrics defined as area under the faithfulness curve and area between the curve and 1; the two-track structure spanning four tasks and five models; public/private test splits with live leaderboards; and InterpBench as a known-ground-truth circuit for AUROC. This is the first cross-method evaluation putting attribution patching, IG variants, IFR, UGS, DAS, DBM, PCA, and SAEs on the same footing. Code and datasets ship, and the appendix is unusually candid.\n\nThe circuit track is well designed. The metrics are principled, the InterpBench anchor is smart, and the finding that EAP-IG with counterfactual ablations beats exact activation patching is worth taking seriously. The SAE result—DBM on SAE features does not beat DBM on raw neurons—is consistent with AxBench and useful, though the authors concede the hyperparameter search was not exhaustive.\n\nThe soft spots, in proportion. The causal variable track's load-bearing assumption is that the high-level models exist in the networks. App. B admits the graphs \"may not always exist in the models we evaluate in the exact forms shown here,\" and App. F.5.2 shows the danger concretely: the apparent XCarry alignment in Llama is \"likely a representation of the tens-digit output,\" and with random counterfactuals the method fails completely. The confound is disclosed, and the low XCarry scores expose it rather than hide it, but it means the DAS-vs-SAE rankings on MCQA/RAVEL/ARC should be read as \"aligns to a hypothesized variable under matched counterfactual distributions,\" not proof that the variable exists. The IOI causal model fits its linear coefficients to the same LM's logit differences on the same counterfactual types used in evaluation—a mild circularity, though it replicates Wang et al.'s earlier structure. Comparison tables lack error bars; some top-method differences are small. Private test results are reported for the circuit track but not the causal variable track, which matters for a paper that sells the leaderboard. MCQA is a roughly 50-example test set. None of these are fatal; the benchmark is the contribution, and the empirical claims are secondary.\n\nWho this is for: anyone working on circuit discovery, SAEs, or causal abstraction. It deserves a serious referee and will be cited widely. My recommendation: accept with revision—add variance estimates, report causal-variable private-test results, soften the SAE claim. The benchmark itself should ship.","headline":"A credible, field-shaping benchmark whose circuit track is solid; the causal-variable findings are real but rest on high-level causal models the paper never independently validates.","tokens_in":41547,"tokens_out":5083,"would_cite":true,"duration_ms":45803,"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":"The paper introduces MIB, a two-track benchmark for mechanistic interpretability, and reports that attribution and mask optimization methods lead circuit localization while supervised DAS leads causal variable localization.","keywords":["mechanistic interpretability","benchmark","circuit localization","causal variable localization","activation patching","sparse autoencoders","distributed alignment search","interchange interventions"],"falsifier":"Train a transformer with a hand-verified ground-truth circuit on a second task and check whether the method ranking by CPR and CMD matches the ranking by edge-level precision and recall; divergence would show the benchmark metrics do not measure true circuit recovery.","tokens_in":40303,"feed_emoji":"🧩","tokens_out":5700,"duration_ms":52439,"temperature":0.7,"pith_summary":"MIB is a benchmark that tries to settle a recurring question in mechanistic interpretability: when a new method claims to find the circuit or the concept a language model uses, how do we know it is actually better? It does this by defining two tasks—locating the model components and connections that drive a behavior, and locating the hidden-vector features that implement a specific causal variable—across four tasks and five models, with fixed counterfactual inputs, standardized metrics, and private test sets. If the benchmark is right, the field gains a shared yardstick: a method can be judged by how precisely and concisely it recovers causal pathways or causal variables, rather than by ad hoc comparisons. The paper's headline findings are that attribution and mask-optimization methods perform best on circuit localization, and that supervised distributed alignment search (DAS) outperforms unsupervised featurizers, with sparse autoencoder features no better than raw neuron dimensions for localizing known causal variables.","feed_headline":"Attribution methods top new benchmark for finding LM circuits","feed_subtitle":"Standardized tracks and private test sets now rank circuit discovery and causal-variable localization across models.","key_machinery":"The load-bearing mechanism is the counterfactual interchange intervention, applied in two ways. For circuits, activation patching—replacing a component's activation with the value it would take on a counterfactual input—produces a faithfulness curve across circuit sizes, summarized by two new metrics: circuit performance ratio (CPR, area under the faithfulness curve) and circuit-model distance (CMD, area between the faithfulness curve and perfect faithfulness). For causal variables, the same intervention is applied to aligned features of a hidden vector, and the metric is interchange intervention accuracy (IIA): the fraction of cases where patching the features has the same output effect as patching the variable in a high-level causal model. The featurizers compared—identity (full vector), PCA, sparse autoencoders, and DAS—are all evaluated through this shared intervention lens.","core_discovery":"The central claim is that MIB is a valid, useful benchmark for comparing mechanistic interpretability methods across method families, and that using it reveals a clear differentiation between methods. On the circuit localization track, edge attribution patching with integrated gradients over inputs (EAP-IG-inputs) with counterfactual ablations generally achieves the lowest circuit-model distance and highest circuit performance ratio; on the causal variable localization track, DAS, the supervised featurizer, consistently gives the most faithful alignments, while masks learned on PCA or SAE features do not beat masks learned on ordinary hidden-unit dimensions. The paper also claims these findings increase confidence that real progress has been made in the field.","pith_inferences":["If the causal-variable results generalize, they put pressure on the common assumption that sparse autoencoders are the right unit of analysis for mechanistic interpretability; a similar pattern already appears in steering benchmarks, where simple directions beat SAE features.","A natural extension beyond the paper is to test whether EAP-IG's advantage persists when circuits are built greedily rather than by top-k edge selection, since the paper uses top-k for most methods.","The near-chance carry-the-one results suggest the benchmark could be extended to non-linear feature spaces, such as onion representations, as a next testbed; the paper notes this possibility implicitly.","One could use the same two-track design on non-language modalities, since the paper limits itself to language models but the metrics are modality-agnostic."],"forward_implications":["Circuit discovery methods can now be ranked by CPR and CMD instead of single-circuit faithfulness, and methods that optimize for sparsity plus faithfulness, like EAP-IG, set a strong default baseline.","SAE features should not be assumed to be better units than neurons for causal localization; any new featurizer must beat DBM-on-neurons to claim progress.","DAS provides a skyline for supervised alignment, and unsupervised methods are measured against it.","The benchmark's private test sets and public leaderboards allow future methods to be compared on unseen data, reducing hill-climbing on known datasets.","The carry-the-one causal variable in two-digit addition remains an open target, since current baselines are near chance; beating them would constitute genuine progress."],"supporting_citations":[{"why":"Supplies the IOI circuit and S-inhibition head analysis that the causal variable track builds on, along with the logit-difference evaluation.","marker":"Wang et al., 2023"},{"why":"Defines distributed alignment search, the supervised featurizer that achieves the best causal-variable localization results.","marker":"Geiger et al., 2024b"},{"why":"Provides the semi-synthetic transformer with a known ground-truth circuit, used for the AUROC completeness metric.","marker":"Gupta et al., 2024"},{"why":"Provides the faithfulness ratio formulation and the EAP-IG-activations attribution method evaluated in the circuit track.","marker":"Marks et al., 2025"},{"why":"Contributes uniform gradient sampling (UGS) and optimal ablations, the mask-based circuit localization baselines.","marker":"Li & Janson, 2024"},{"why":"Introduces attribution patching, the approximation underlying EAP and NAP methods.","marker":"Syed et al., 2024"},{"why":"Supplies the GemmaScope sparse autoencoders used as unsupervised featurizers in the causal variable track.","marker":"Lieberum et al., 2024"}],"fun_headline_variants":["Attribution patching bests other circuit-finding methods in MIB","MIB benchmark: DAS wins causal variable search over SAEs","New interpretability benchmark ranks circuit and causal methods","MIB shows SAEs not better than plain neurons for causal features","Benchmark for interpretability: attribution and DAS come out on top"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The causal variable track presumes that the high-level mental model it tests (like 'carry the one' in addition or 'find the answer position' in multiple choice) is actually the algorithm the network implements; if that variable is not really computed by the model, the alignment scores cannot tell us about real mechanisms.","fun_headline_variants_meta":{"raw":{"variants":["Attribution patching bests other circuit-finding methods in MIB","MIB benchmark: DAS wins causal variable search over SAEs","New interpretability benchmark ranks circuit and causal methods","MIB shows SAEs not better than plain neurons for causal features","Benchmark for interpretability: attribution and DAS come out on top"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000267,"raw_usage":{"total_tokens":1574,"prompt_tokens":864,"completion_tokens":710,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":621}},"tokens_in":480,"tokens_out":710,"duration_ms":6890,"temperature":1.0,"reasoning_tokens":621,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:13:29.728711+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a transformer with a hand-verified ground-truth circuit on a second task and check whether the method ranking by CPR and CMD matches the ranking by edge-level precision and recall; divergence would show the benchmark metrics do not measure true circuit recovery.","supporting_citations":[{"cited_title":"R., Variengien, A., Conmy, A., Shlegeris, B., and Steinhardt, J","cited_arxiv_id":null,"evidence_quote":"Supplies the IOI circuit and S-inhibition head analysis that the causal variable track builds on, along with the logit-difference evaluation."},{"cited_title":"Interp B ench: Semi-synthetic transformers for evaluating mechanistic interpretability techniques","cited_arxiv_id":null,"evidence_quote":"Provides the semi-synthetic transformer with a known ground-truth circuit, used for the AUROC completeness metric."}],"review_version":1}