{"id":"91aaf2a7-94f3-4ff1-a02c-74a6d22769c3","arxiv_id":"2508.01728","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GCC discovers multiple concept-specific neuron circuits per query by combining first-order ablation sensitivity with top-k activation overlap.","lead":"Granular Concept Circuits (GCC) automatically finds multiple fine-grained neuron circuits in image classifiers, each meant to capture one visual concept relevant to a query image. The method combines neuron-ablation sensitivity and activation-pattern overlap to trace concept-specific pathways across layers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"First-order ablation plus top-k overlap in Sec. 3.1 may fail to trace concept-preserving circuits, leaving the central claim of distinct concept circuits unvalidated.","rationale":"I agree with the reader that the weakest assumption is the first-order connectivity proxy. The paper's own disclaimers (Sec. 3.1 intractability, Sec. 5 distributed concepts) flag exactly the failure mode where single-neuron ablation is blind. The existing quantitative results are not decisive: Table 2 ablates the entire set of circuits, which cannot establish per-circuit concept identity; Fig. 6 shows S_NS ranking has some functional relevance, but that is a necessary not sufficient condition for concept preservation. The paper has genuine strengths: released code, qualitative demonstrations across models, and a user study; it is a plausible contribution. However, without a controlled test showing that the proxy traces concept-preserving connections, the headline claim of fine-grained concept-specific circuits is under-supported. If the proposed synthetic/counterfactual test passes, the method would be substantially strengthened; if it fails, the method's discovered circuits could be merely sensitivity artifacts. Since the reader already set CONDITIONAL, my read does not move the verdict; it sharpens the specific experiment that would resolve the outstanding risk.","tokens_in":12689,"tokens_out":7327,"duration_ms":88863,"concrete_test":"Train a small CNN on synthetic images where each image contains one known visual concept, and engineer one concept to be encoded by a pair of hidden neurons whose joint activation is required for the downstream target, so that masking either neuron alone has no effect. Apply GCC to a query containing that concept and check whether the recovered circuit includes both neurons and no spurious edges. If GCC misses the pair or includes unrelated neurons, the first-order zero-masking proxy in Eq. 1 does not preserve concept identity, and the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that each automatically discovered GCC encodes a distinct human-interpretable concept. This claim depends on the connectivity scores in Sec. 3.1: the Neuron Sensitivity Score (Eq. 1) uses a first-order zero-masking ablation, and the Semantic Flow Score (Eq. 3) uses top-k sample overlap. The paper explicitly notes that full interaction effects are intractable (Sec. 3.1), and Sec. 5 concedes that 'a single concept may be distributed across multiple circuit pathways.' Those two statements identify the load-bearing weakness: if a concept is encoded by a set of neurons whose joint effect is necessary, zero-masking any single source neuron will produce little or no change, so S_NS will fail to recover the connection and the circuit will be incomplete. In addition, top-k overlap is correlational and cannot distinguish concept-preserving edges from polysemantic neurons active for multiple unrelated concepts. The quantitative faithfulness test (Table 2) ablates all circuit neurons simultaneously, so it validates only the union of circuits, not concept specificity of each circuit; the user study (Fig. 5) is small (n=33) and subjective. Therefore, the first-order proxy is the most load-bearing assumption and is not yet validated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Granular Concept Circuits (GCC), a method for discovering sets of neurons across layers in vision models that are claimed to each encode a distinct, human-interpretable visual concept relevant to a given query image. Connectivity between adjacent-layer neurons is established through two scores: a Neuron Sensitivity Score (zero-masking the source neuron and measuring next-layer activation change, Eq. 1) and a Semantic Flow Score (overlap of top-k highly activated samples, Eq. 3). Starting from root nodes selected by high query activation, the method iteratively grows each circuit and repeats for all root nodes, producing multiple concept-specific circuits. Experiments on CNNs (VGG19, ResNet50/101, MobileNetV3) and Transformers (ViT, Swin, CLIP-ViT) include qualitative visualizations, logit/accuracy ablation comparisons, edge deletion/insertion curves, a 33-participant user study, and two use cases (misclassification audit and common-concept discovery). The paper claims to be the first to identify fine-grained, concept-specific visual circuits.","tokens_in":12923,"tokens_out":2876,"duration_ms":35214,"significance":"If the central claim holds, GCC would be a valuable contribution to vision interpretability by moving from single-neuron or single-circuit analyses to a fine-grained, concept-wise decomposition of a model's internal representations. The paper has several strengths: the code is publicly available; the method is evaluated across a wide range of architectures (CNNs and Transformers); the Peak-over-Threshold procedure for threshold selection is principled and empirically examined; and the user study, while small, is a positive step toward human validation. However, the key claim that each discovered circuit is a single, coherent human-interpretable concept is not yet convincingly demonstrated. The quantitative evaluations validate aggregate influence rather than per-circuit concept specificity, and the faithfulness metric is partly self-confirming because it ablates exactly the neurons that the method selected by their influence. The first-order approximation underlying the Neuron Sensitivity Score is load-bearing for the notion of a 'concept-preserving' circuit and remains untested against alternative interaction-aware connectivity measures.","major_comments":[{"comment":"The first-order zero-masking approximation is load-bearing for the paper's central claim, but the manuscript does not validate that it preserves concept identity. If a concept is encoded by a set of neurons whose joint effect is necessary, zero-masking any single source neuron may produce only a small change in the target, so S_NS would miss the true connection. The paper acknowledges the intractability of full interactions (Sec. 3.1) and later concedes that 'a single concept may be distributed across multiple circuit pathways' (Sec. 5). These two statements together identify a concrete risk: the discovered circuits may be statistically sensitive pathways without coherent semantic meaning. I recommend adding a targeted experiment, for example comparing S_NS against second-order or joint ablations on a small subset of layers, to show that first-order effects are sufficient to recover concept-preserving connections.","section":"Sec. 3.1, Eq. (1)"},{"comment":"The quantitative faithfulness/completeness evaluation ablates all circuit neurons simultaneously, so it validates the union of circuits, not the specificity of each individual circuit. The paper's central claim is that 'multiple circuits, each encoding a distinct concept relevant to the query' (Abstract, Sec. 3.2). Table 2 and Table E report aggregate logit/accuracy drops when all discovered circuits are ablated together; such results would also be observed if the method found one large, unspecific influential pathway. To support the per-circuit claim, the authors should evaluate each circuit separately, for example by ablating a single circuit and showing a selective effect on concept-related outputs or by demonstrating that circuits are not redundant with one another. In addition, the average logit drops in Table 2 and the accuracy drops in Table E are reported as point estimates without error bars or significance tests over the 100 queries; at minimum, standard errors or confidence intervals should be provided.","section":"Sec. 4.2, Table 2 and Appendix B"},{"comment":"The main faithfulness metric is partially self-confirming. Root nodes are selected by high query activation, edges are filtered by high S_NS and S_SF, and then the evaluation ablates these same selected neurons and measures the resulting logit drop. Any criterion that selects high-influence neurons would produce some logit drop, so the large drop in Table 2 does not distinguish concept-specific circuits from a generic set of influential neurons. The random baseline is not matched on activation magnitude or on query relevance, and the complement ablation (OursC) excludes by construction the very neurons the method deems important. A stronger control would ablate the same number of top-activation neurons (without the connectivity constraints) or a set matched on S_NS but not on S_SF, to show that the joint connectivity and semantic alignment criteria are what matter.","section":"Sec. 4.2, Table 2"},{"comment":"The claim that each GCC is a single human-interpretable concept rests primarily on qualitative examples and a 33-participant user study, but the user study does not directly test monosemanticity. The questions in Fig. 5 ask about query relevance, diversity, and prototypicality of the circuit as a whole, not whether each circuit corresponds to one coherent concept that a human can label consistently. The qualitative visualizations in Figs. 3 and 4 are suggestive, but they show only a few examples and the manual textual descriptions could be post-hoc rationalizations. I recommend a more objective concept-coherence metric, such as NetDissect-style IoU against semantic segmentation masks for the concept named by the circuit, or a label-consistency measure across the top highly activated samples of the circuit's root and downstream nodes.","section":"Sec. 4.3 and Sec. 4.1"}],"minor_comments":[{"comment":"The notation f^{l+1}(a^l_c) is ambiguous: it appears to feed the masked or unmasked activation vector of the entire layer, but the subscript c suggests a single channel. Please clarify that a^l_c denotes the full layer activation with only neuron c zero-masked, so that the expression reads as a function of the whole vector.","section":"Sec. 3.1, Eq. (1)"},{"comment":"The paper refers to 'A VCD' in the text near Fig. 6; this appears to be a typo for ADVC (from Ref. [29]). Please correct for consistency.","section":"Sec. 4.2, Fig. 6"},{"comment":"The phrase 'to supporess weak activations' contains a typo ('suppress'). Also, the visualization method description would benefit from stating how the threshold for the binary mask is chosen, since this affects the perceived concept quality in Figs. 3 and 4.","section":"Sec. 4, Visualization Method"},{"comment":"The root-node selection threshold (top 1%) is stated to be tunable, and the appendix mentions that top 10% increases diversity but reduces query relevance, yet no quantitative comparison is given. A small sensitivity analysis, even on one model, would help readers understand the impact of this parameter on the faithfulness/completeness metrics.","section":"Appendix A.2"},{"comment":"Table D in the appendix reports POT threshold comparisons but appears to have a missing entry for the 'Ours' row at POT=70; the table formatting should be checked. Also, the caption labels the table as 'Table D' while the text in Appendix A.3 refers to it; ensure all cross-references are consistent.","section":"Table 3 / Table D"}],"recommendation":"major_revision","confidential_remarks":"This is a competent and interesting paper that addresses a real gap in vision interpretability. The main concern for the editor is that the central scientific claim—each GCC is a single, distinct, human-interpretable concept—is supported more by visual plausibility than by controlled quantitative evidence. The evaluation metric is partly circular, and the first-order approximation is untested for concept preservation. The paper is not fatally flawed; the issues are fixable with additional experiments and analysis within the paper's scope. I recommend major revision rather than rejection, with emphasis on per-circuit evaluation, stronger baselines, and at least one test of the first-order approximation's validity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Kwon et al. propose Granular Concept Circuits (GCC), a new method that discovers multiple per-query concept circuits in vision models by combining a first-order neuron sensitivity ablation (S_NS) with a top-k activation overlap score (S_SF). Compared to ADVC, which produces a single class-conditioned circuit, and VCC, which does not localize to neurons, GCC's granularity is genuinely new. The code is available, thresholds are discussed, and qualitative results across ResNet, VGG, MobileNet, ViT, and CUB show plausible hierarchical concept flows. The misclassification audit use case is a nice demonstration.\n\nThe main soft spot is the gap between the method's claim and its validation. Each circuit is supposed to encode a distinct human-interpretable concept, but the quantitative faithfulness test ablates all circuit neurons together, so it only shows the union is query-relevant, not that each circuit is concept-specific. Root nodes are chosen for high query activation and edges by sensitivity, making the logit drop partly self-confirming. Error bars are missing from Table 2 and the transformer table. The user study (n=33) is subjective, and the paper does not offer per-circuit concept coherence metrics. The first-order approximation is also load-bearing: zero-masking a single source neuron will miss concepts encoded in joint effects, and the paper itself concedes that a concept may be distributed across multiple pathways and that full interactions are intractable. The top-k overlap score is correlational and can be confounded by polysemantic neurons.\n\nThat said, these are known hard problems in circuit discovery, and the paper is honest about its limitations. The method is a plausible heuristic that could be useful for model debugging and hypothesis generation. The quantitative drop is large, and the parameter sensitivity analysis (POT thresholds) is a good practice.\n\nThis paper deserves a serious referee. I would send it to a top vision or interpretability venue, not desk-reject it. The authors should be pushed to add error bars, validate concept specificity per circuit (e.g., using concept probes or targeted interventions), and discuss the first-order limitation more concretely. If they can tighten that, the contribution would be solid. As it stands, it is a conditional accept with major revisions.","headline":"A genuinely new method for finding multiple per-query concept circuits, but the claim that each circuit encodes a single distinct concept is not yet validated by the experiments.","tokens_in":13439,"tokens_out":2711,"would_cite":true,"duration_ms":30390,"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":"The paper claims that a deep vision model's representation of a query image can be decomposed into multiple fine-grained concept circuits.","keywords":["circuit discovery","concept interpretability","neuron ablation","semantic alignment","visual concepts","mechanistic interpretability","deep vision models","fine-grained explanation"],"falsifier":"Build a small network with a planted concept circuit in which the next-layer response depends on the joint activity of two source neurons rather than on either alone, run GCC on it, and check whether both edges are recovered; a failure to connect either neuron would show that the first-order zero-masking proxy misses higher-order concept dependencies.","tokens_in":12484,"feed_emoji":"🧠","tokens_out":8180,"duration_ms":82666,"temperature":0.7,"pith_summary":"The paper tries to establish that a deep image classifier's distributed representation of a query image can be decomposed into multiple granular circuits, each encoding one human-interpretable visual concept relevant to that query. It argues that prior circuit methods either produce a single class-level pathway or miss concept specificity, whereas the proposed method automatically returns several circuits per query. If true, this gives a fine-grained operational account of where and how concepts such as 'sky background', 'flags', or 'clock' live inside a trained model, and supports uses like auditing misclassifications and finding concepts shared across different classes. The central evidence is that ablating neurons inside the discovered circuits degrades model outputs far more than ablating the same number of neurons outside them.","feed_headline":"Ablation tracing splits one image into multiple concept circuits","feed_subtitle":"Each circuit tracks one human-readable concept through the network, showing where sky, clock, or wheel features live.","key_machinery":"The mechanism is a forward, iterative connectivity search through the network. Starting from root nodes, defined as neurons whose activations rank in the top 1 percent for the query, the method scores each candidate edge from a source neuron in layer $l$ to a target neuron in layer $l+1$ with two quantities: the Neuron Sensitivity Score, which zero-masks the source neuron and measures the resulting change in the target's activation, and the Semantic Flow Score, which computes the overlap of the top-$k$ highly activated samples between the two neurons. Edges pass only if both scores clear thresholds, with the Neuron Sensitivity threshold selected automatically by peak-over-threshold extreme-value statistics and the Semantic Flow threshold set to the average score. Surviving edges are added to the circuit, the newly added nodes become new sources, and the search continues until no further connections remain; repeating this for every root node yields the collection of granular concept circuits.","core_discovery":"On its own terms, the paper's central claim is that each discovered Granular Concept Circuit corresponds to a coherent visual concept relevant to the given query, and that the full set of circuits forms a fine-grained decomposition of the model's representation for that query. The claim includes a methodological assertion: connections between adjacent-layer neurons that carry a single concept can be identified by combining a first-order functional-dependence measure, the Neuron Sensitivity Score, with a semantic-alignment measure, the Semantic Flow Score. The paper further claims it is the first circuit discovery method to return multiple concept-specific circuits rather than a single class-level circuit, and it validates this on CNNs and transformer models with ablation experiments and a user study.","pith_inferences":["Beyond the paper's claims, the first-order ablation proxy could be stress-tested on synthetic networks with planted concept circuits: if a concept is encoded by two neurons whose joint effect, but neither alone, drives the next layer, zero-masking one source at a time may fail to find the edge.","An extension not tested in the paper would replace the sample-overlap Semantic Flow Score with a concept-identity metric, such as clustering or textual descriptions of neuron activations, and check whether the discovered circuits remain stable under that change.","The paper itself notes that a single concept may be distributed across multiple circuit pathways under strict connectivity thresholds, so the number of circuits per query should be treated as a tunable granularity rather than a fixed ground-truth decomposition.","Because the forward search does not require class labels, the method could be redirected at attributes, textures, or background objects; one could ask which circuits fire for a specific property and audit the model's reasoning about that property."],"forward_implications":["Ablating the neurons of a discovered circuit produces a substantially larger drop in the model's output than ablating the same number of neurons outside it, which the paper reports as evidence that the circuits are both faithful and complete.","Because each circuit is tied to a specific concept rather than a class, the method separates multiple concepts within one query image and shows how each concept flows through successive layers.","For multiple queries, the method can expose concepts shared across different classes, such as a common 'wheel' pattern among vehicles, by propagating forward from root nodes instead of relying on class logits.","In a misclassification audit, inhibiting or stimulating the neurons of discovered circuits changes the logit for the true class, revealing which concept-level pathways drive the erroneous prediction.","The same circuit-ablation procedure applied to transformer models yields large accuracy drops, indicating that the discovery mechanism transfers beyond convolutional architectures."],"supporting_citations":[{"why":"Supplies the practice of interpreting neurons by their highly activated samples, which the Semantic Flow Score builds on.","marker":"[3]"},{"why":"Supplies the intervention-based zero-masking approach for measuring a source neuron's effect on downstream targets.","marker":"[8]"},{"why":"The main concept-connectivity baseline the paper compares against and improves on in granularity.","marker":"[22]"},{"why":"The single-circuit baseline whose gradient-based forward search is contrasted with GCC's score-based search.","marker":"[29]"},{"why":"Provides the faithfulness and completeness evaluation protocol adopted for the ablation experiments.","marker":"[37]"},{"why":"Provides the circuit-masking evaluation protocol for measuring output corruption after ablating circuit versus complement.","marker":"[41]"},{"why":"Establishes the causal-ablation precedent for zeroing a neuron and measuring downstream change.","marker":"[42]"}],"fun_headline_variants":["Fine-grained concept circuits dissect vision model representations","New method finds multiple concept circuits per visual query","Concept circuits trace where sky, clock, wheel features live in nets","Splitting one image into multiple concept circuits automatically","First method to return concept-specific circuits, not class-level ones"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a first-order ablation of one neuron at a time, combined with overlap of the most strongly activated samples, is enough to identify connections that carry a single coherent concept across layers; if higher-order interactions or polysemantic neurons break this proxy, the discovered circuits may be statistically sensitive pathways without stable semantic meaning.","fun_headline_variants_meta":{"raw":{"variants":["Fine-grained concept circuits dissect vision model representations","New method finds multiple concept circuits per visual query","Concept circuits trace where sky, clock, wheel features live in nets","Splitting one image into multiple concept circuits automatically","First method to return concept-specific circuits, not class-level ones"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000262,"raw_usage":{"total_tokens":1537,"prompt_tokens":825,"completion_tokens":712,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":441,"completion_tokens_details":{"reasoning_tokens":633}},"tokens_in":441,"tokens_out":712,"duration_ms":6866,"temperature":1.0,"reasoning_tokens":633,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:24:51.788690+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a small network with a planted concept circuit in which the next-layer response depends on the joint activity of two source neurons rather than on either alone, run GCC on it, and check whether both edges are recovered; a failure to connect either neuron would show that the first-order zero-masking proxy misses higher-order concept dependencies.","supporting_citations":[{"cited_title":"Network dissection: Quantifying inter- pretability of deep visual representations","cited_arxiv_id":null,"evidence_quote":"Supplies the practice of interpreting neurons by their highly activated samples, which the Semantic Flow Score builds on."},{"cited_title":"Visual concept connectome (vcc): Open world con- cept discovery and their interlayer connections in deep mod- els","cited_arxiv_id":null,"evidence_quote":"The main concept-connectivity baseline the paper compares against and improves on in granularity."}],"review_version":1}