{"id":"26f229d2-78db-4e39-9202-605547faee96","arxiv_id":"2601.16366","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An activation-aware graph-curvature score ranks neural connections for post-training pruning, preserving accuracy as well as or better than magnitude, SNIP, or SynFlow on MNIST/CIFAR models in one-shot removal tests.","lead":"This paper ranks a trained neural network's connections using a graph-curvature score built from the network's structure plus how strongly neurons respond to real inputs. Removing the lowest-ranked connections first preserves accuracy as well as or better than standard magnitude, SNIP, or SynFlow pruning on small image-classification models in these tests.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pruning gains may stem from ReLU-dead edges and inverse-weight magnitude rather than from the negative/positive curvature signal; no ablation isolates this confound.","rationale":"The reader's weakest assumption concerns the hand-built graph model, specifically the unvalidated choice of edge cost (inverse weight magnitude) and the activation-normalization steps. My concern is a sharper version of this: the construction may encode dead-ReLU status and inverse weight magnitude so directly that the ORC computation is not the operative mechanism. This is partially in agreement with the reader because both point to the arbitrariness of the graph encoding, but I emphasize a specific confound that could be tested with a simple control baseline. The reader's CONDITIONAL verdict is appropriate: the paper's central claim is plausible but not yet established because the necessary control is missing and the artifacts (code/data) are not available. My concern does not change the verdict; it reinforces the conditionality.","tokens_in":19265,"tokens_out":9655,"duration_ms":192609,"concrete_test":"Run the same edge-removal protocol with a control score: s(p)=2 if the parameter has a ReLU-dead input or output (activation <= 0 on all calibration examples), else s(p)=|W_p|/max_l|W| (or 1/|W_p| with per-layer normalization). Sort descending and compare accuracy-vs-pruned-fraction curves to neural curvature on all models in Fig. 3. If the control is within 1% accuracy at every pruning level, the ORC ranking is confounded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing risk is that the reported pruning advantage is not caused by the sign/geometry of neural curvature but by two simpler signals embedded in the construction. In Def. 5, graph edge cost is 1/|W|; in Def. 7, ReLU edges incident to a dead neuron receive infinite cost, and Prop. 1 maps those to curvature 2 (the highest value). Sorting descending thus removes all dead-ReLU edges first, then likely proceeds through small-|W| edges after a cross-layer normalization induced by shortest-path distances. The paper's own §7.3.2 concedes that within a layer the ordering behaves like magnitude pruning, and only claims cross-layer normalization as the added value. The ablation in §7.3.3 compares against static Lin et al. Ricci, not against a control that uses activation-deadness and magnitude alone. If such a control reproduces the edge-removal curves, then the abstract's claim that negative-ORC edges are bottlenecks and that curvature provides a data-flow map is unsupported. No code or data are released, so this confound cannot currently be checked.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a graph-curvature-based method for ranking connections in trained feedforward NNs. It constructs a directed weighted neural graph (Def. 5) with edge costs equal to inverse weight magnitudes, defines activation-dependent neighbor distributions (Def. 6), modifies the edge cost to account for activation saturation/deadness (Def. 7), and introduces a \"neural curvature\" via the limiting α-Ricci curvature (Def. 8). Edges are ranked by the minimum curvature over a small calibration set, and the ranking is evaluated in edge-removal experiments on MNIST, CIFAR-10, and CIFAR-100 against magnitude pruning, SNIP, and SynFlow. The central claim is that negative-curvature edges are bottlenecks and carry the main data flow, while positive-curvature edges are unimportant; the method is claimed to identify more unimportant edges than prior pruning methods.","tokens_in":19483,"tokens_out":5644,"duration_ms":58140,"significance":"If the central claim is robust, this would be a useful contribution: it provides a geometrically motivated, data-dependent importance score that can be computed on layered DAGs with a dynamic-programming shortest-path implementation, and it shows promising results across several architectures, activations, and training schemes. The paper is honest about several limitations, noting comparable performance to magnitude pruning on MNIST and for weight-decayed models, and it includes an ablation of calibration-set size. The method's data efficiency (one example per label is often sufficient) is also a strength. However, the significance is currently conditional: the experimental evidence does not yet separate the proposed curvature signal from simpler signals embedded in the construction, and the notation/definition of the neural distributions is under-specified in a way that affects reproducibility.","major_comments":[{"comment":"The central claim that negative-neural-curvature edges are bottlenecks is not isolated from simpler confounds. Definition 7 assigns infinite cost to edges incident to ReLU-dead neurons, and Definition 8/Proposition 1 map those edges to curvature 1–2, i.e., the highest values. Since ranking is descending, all ReLU-dead edges are removed first, then edges with small |W| (Def. 5). The ablation in §7.3.3 only compares against static Lin et al. Ricci; it does not include a control that removes ReLU-dead edges first and then prunes by inverse magnitude, nor a control using per-layer normalized magnitude. Section 7.3.2 concedes that within a layer the ordering behaves like magnitude pruning, so the remaining claimed advantage is cross-layer normalization. Without a control that uses activation-deadness and inverse magnitude alone, the abstract's statement that negative-ORC edges are bottlenecks","section":"§7.3.3 and Defs. 5/7/8, Prop. 1"},{"comment":"The definition of the α-neural distribution is ambiguous. Equation (3) defines m^α_x using graph weights w_n, while Definition 6 defines an activation-based probability vector. The text says m^α_{v_i,l}(x) and m^α_{v_j,l+1}(x) are 'calculated according to Definition 6 and equation 3', but it does not specify how the two are combined. If Equation (3) is used literally with graph weights, the activation data enter only through the edge cost dσ, which would weaken the claimed data dependence. If the Definition 6 vector replaces the weight-proportional distribution in Eq. (3), this must be stated explicitly. This ambiguity affects the implementation of every curvature value and must be resolved before the method can be reproduced.","section":"Def. 8 and Eq. (7)"},{"comment":"The per-layer normalization (n_i - n_min)/(n_max - n_min) is undefined whenever n_max = n_min, which occurs not only when all activations are zero (the case discussed in the footnote) but also when all |n_i| in a layer are equal to the same positive constant. The footnote only addresses the all-zero case. If the implementation adds a small constant to the denominator in all cases, that should be stated; otherwise a well-defined fallback (e.g., a uniform distribution) is needed.","section":"Def. 6, Step 1"},{"comment":"The paper claims that the method's advantage over magnitude pruning comes from 'effective normalization across layers'. This is a load-bearing claim, because within-layer behavior is admitted to match magnitude pruning. However, no experiment directly tests this normalization hypothesis. A natural control is per-layer normalized magnitude pruning (e.g., dividing weights by the maximum |W| in each layer before global ranking). If that control matches the neural-curvature curve, the cross-layer signal reduces to a standard normalization and the geometric interpretation is weakened. Please include this control and, if possible, report the curvature distribution over layers to substantiate the claim.","section":"§7.3.2 / Fig. 5"}],"minor_comments":[{"comment":"The arXiv title is 'Post-Training Neural Network Pruning using Graph Curvature' while the body title is 'Analyzing Neural Network Information Flow Using Differential Geometry'. Please align the title and the abstract with the paper's actual scope.","section":"Title/Abstract"},{"comment":"In Eq. (1), d(u',v') is described as 'the cost from u' to v', e.g., d(u,v)=w(u,v)'. For nodes that are not directly connected by an edge, this cost must be a shortest-path distance in the weighted graph. Please state this explicitly, since the implementation section relies on shortest-path distances.","section":"Def. 1"},{"comment":"The choice α=0.9 is justified only by 'empirical observations'. Please provide a sensitivity analysis or at least a statement of how curvature rankings change for α in {0.5, 0.7, 0.9, 0.99}. This is especially important because the theoretical definition uses the limit α→1.","section":"Sec. 6"},{"comment":"The text says results are 'equally effective across all considered combinations', but CIFAR-100 is evaluated with only one model (ReLU). Please temper the generalization claim or add additional CIFAR-100 configurations.","section":"Sec. 7.1 / Model description"},{"comment":"No code or data are provided. Given the number of ad-hoc choices in the pipeline (normalization, edge costs, α, min aggregation), releasing code would substantially increase confidence in the reported results.","section":"Reproducibility"},{"comment":"The vector for m_n5 appears to have a typo: '0.2  0.2  0.2  0.2  0  0 .2  0  0  0  0  0' contains a stray '.2'. Please reformat and verify the entries sum to 1.","section":"Example 1"}],"recommendation":"major_revision","confidential_remarks":"The central idea is interesting and the paper is readable, but the evidence for the geometric interpretation is not yet compelling because the construction contains simpler signals (ReLU deadness and inverse magnitude) that are not controlled for. The requested ablation is straightforward and should be feasible with the existing codebase. I would also strongly encourage the authors to release code/data, because the method has several hand-designed normalization steps that are hard to reproduce from the text alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe core idea here is real: turn a trained NN into a weighted directed graph, compute an activation-dependent version of Ollivier-Ricci curvature, and use the sign/ordering of the curvature to rank edges for one-shot removal. The neural neighbor distribution (Def 6) and the beta-scaled edge cost (Def 7) are new as far as I can tell, and the paper does the honest work of reporting where magnitude pruning wins (WD-trained models) and admitting the within-layer behavior matches magnitude (Sec 7.3.2). The ablation in Fig 6 shows both the neural distributions and the edge cost each beat static Lin et al. Ricci on the same graph, which is useful evidence that the data-dependent part matters.\n\nThe soft spot is real and structural. The construction stacks two strong priors on top of the curvature: edge cost is 1/|W| (Def 5), and any ReLU edge with a dead source or target gets infinite cost (Def 7). By Prop 1, those infinite-cost hidden edges get curvature exactly 2—the maximum—so descending-sort pruning removes all dead-ReLU edges first, then continues through small-|W| edges after a distance-based normalization. The paper never runs the control that would separate the curvature signal from those two ingredients: a \"deadness + inverse magnitude\" ranking that skips the transport/curvature computation. Without that control, the abstract's claim that negative-ORC edges are bottlenecks and that curvature gives a data-flow map is not supported. The within-layer result actually makes the worry sharper—if curvature only adds cross-layer normalization to a magnitude-like ranking, then the bottleneck semantics are decorative.\n\nOther issues are smaller: Def 8 is under-specified about how the alpha-mass blends with the activation-based distributions; no code, models, or seeds are released, so the figure-based evidence can't be checked; and the abstract overclaims relative to the body's hedged language. None of these are fatal, and the transport decomposition in Prop 1 checks out for layered DAGs.\n\nI'd send this to reviewers—the idea is novel enough and the evaluation broad enough to deserve referee time—but I would not cite it yet. The confound is load-bearing and needs either a control ablation or a formulation that proves the curvature sign is doing the work.\n\nRecommendation: engage seriously, request a control and artifacts before publication.","headline":"A genuinely novel activation-aware curvature score for pruning, but the headline result may be dead-ReLU detection plus inverse-magnitude ranking in disguise; worth refereeing, not yet worth citing.","tokens_in":20035,"tokens_out":1992,"would_cite":false,"duration_ms":18554,"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":"This paper claims that a graph-curvature score computed from a trained network and a small calibration set ranks connections by functional importance, so curvature can drive pruning.","keywords":["neural curvature","Ollivier-Ricci curvature","post-training pruning","connection importance ranking","activation-aware curvature","graph-based data flow analysis","cross-layer pruning","calibration data efficiency"],"falsifier":"Run the same pruning experiment on the same trained models with two edge-cost choices—inverse weight magnitude versus absolute weight magnitude, or a sign-aware cost—and check whether the set of negative-curvature edges and the pruning curves remain approximately the same. If the curvature ranking flips under this substitution, the claim that curvature identifies information-flow bottlenecks is not robust to the encoding; if it does not, the geometric signal is stable. A complementary check is to measure edge-level data transmission directly and test whether inverse-weight edges actually carry","tokens_in":19096,"feed_emoji":"🧠","tokens_out":6444,"duration_ms":96936,"temperature":0.7,"pith_summary":"The paper tries to establish that the bottlenecks and redundancies of a trained neural network can be read off from a geometric quantity: the Ollivier-Ricci curvature of a graph whose nodes are neurons and whose edge weights encode both the learned weight magnitudes and the activation values observed on a handful of input examples. Negative-curvature edges, it argues, are the network's critical data-flow bridges—removing them first collapses accuracy—while positive-curvature edges can be pruned in large numbers with little effect. This gives a single cross-layer importance score for every connection, one that avoids the layer-collapse failure of magnitude-based pruning and identifies more removable parameters than existing gradient- or conservation-based pruning baselines. The practical payoff would be a cheap, data-efficient way to build a data-flow map of a trained model, which the paper points toward robustness analysis, model repair, and focusing retraining on underused connections.","feed_headline":"Curvature finds which neural connections are safe to prune","feed_subtitle":"A graph-geometry score built from activations beats magnitude pruning at spotting unimportant edges across layers.","key_machinery":"Neural curvature: the limit, as the alpha-parameter goes to 1, of a modified alpha-Ollivier-Ricci curvature computed on the neural graph, where alpha leaves probability mass on the source and target nodes. The graph's edge costs are inverse weight magnitudes, neighbor distributions are built from min-max normalized, reciprocated, then exponentially smoothed absolute activations, and the edge cost is inflated (to infinity for ReLU dead neurons) according to the saturation of the activation function. This mechanism turns a static structural graph into a per-example data-flow geometry, and taking the minimum over calibration examples gives a global cross-layer importance ranking.","core_discovery":"On its own terms, the paper's central discovery is that Ollivier-Ricci curvature, adapted to be activation-dependent, separates neural connections into two functionally distinct classes: edges whose curvature is negative act as bottlenecks whose removal quickly degrades test accuracy, and edges whose curvature is positive can be removed without immediate harm. The adaptation works by building a 'neural graph' from the network (one node per neuron, one edge per connection, edge cost equal to the inverse weight magnitude), replacing the graph's static neighbor distributions with distributions derived from neuron activations for each calibration example, and scaling edge costs by how much the a","pith_inferences":["A testable extension the paper leaves implicit: if the edge cost is changed from inverse weight magnitude to, say, absolute weight or a sign-aware cost, the curvature ordering may shift; comparing pruning curves under these weightings would separate the geometric signal from the modelling convention.","The negative-edges-as-bottlenecks view suggests an attack-oriented use: adversarial perturbations might be most effective when targeted at edges with the most negative curvature, and defending those edges could be a cheap robustness intervention—this is not tested in the paper.","The method's cross-layer normalization is the likely source of its advantage over magnitude pruning; the paper's per-layer ablation shows the two agree within a layer, implying curvature's main added value is not local importance but a common scale across layers.","The same curvature machinery could be carried to other architectures, such as transformers or recurrent networks, by defining appropriate neighbor distributions and edge costs; whether the bottleneck semantics survive attention's softmax coupling is an open question."],"forward_implications":["If the central claim holds, pruning a trained network by removing the most positive-curvature edges first should preserve accuracy far longer than removing negative-curvature edges first, as demonstrated across MNIST, CIFAR-10, and CIFAR-100 models with both ReLU and Tanh activations.","The curvature score provides a global ranking across layers, so pruning decisions are not distorted by differing weight scales between layers, which is the mechanism the paper gives for avoiding layer collapse.","A very small calibration set, on the order of one example per label, appears sufficient to separate important from unimportant edges, making the method practical when data are scarce.","The same ranking can serve as a data-flow map for the network: negative-curvature edges identify the paths the model actually relies on, which the paper proposes as a starting point for robustness analysis and model repair."],"fun_headline_variants":["Negative curvature flags can't-prune edges; positive flags safe cuts","Activation-driven curvature ranks which neural connections to keep or cut","Ollivier-Ricci curvature finds critical vs expendable neural edges","Curvature separates essential neural links from disposable ones","Neural curvature: negative edges = bottlenecks, positive = prunable"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The ranking inherits everything from the chosen graph encoding: edge cost is the inverse absolute weight and neighbor distributions are a hand-built transform of activations, and the paper does not validate that this encoding measures actual information transmission; if a plausible alternative encoding changes the curvature ordering, the bottleneck semantics rest on convention.","fun_headline_variants_meta":{"raw":{"variants":["Negative curvature flags can't-prune edges; positive flags safe cuts","Activation-driven curvature ranks which neural connections to keep or cut","Ollivier-Ricci curvature finds critical vs expendable neural edges","Curvature separates essential neural links from disposable ones","Neural curvature: negative edges = bottlenecks, positive = prunable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000812,"raw_usage":{"total_tokens":3407,"prompt_tokens":761,"completion_tokens":2646,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":2566}},"tokens_in":505,"tokens_out":2646,"duration_ms":19462,"temperature":1.0,"reasoning_tokens":2566,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T08:37:30.916754+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same pruning experiment on the same trained models with two edge-cost choices—inverse weight magnitude versus absolute weight magnitude, or a sign-aware cost—and check whether the set of negative-curvature edges and the pruning curves remain approximately the same. If the curvature ranking flips under this substitution, the claim that curvature identifies information-flow bottlenecks is not robust to the encoding; if it does not, the geometric signal is stable. A complementary check is to measure edge-level data transmission directly and test whether inverse-weight edges actually carry","supporting_citations":[],"review_version":1}