{"id":"87985827-f6ed-4fd7-9b93-1aef407d0ef6","arxiv_id":"2606.14022","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Putting a degree-scalar after LayerNorm preserves topology magnitude that pre-LN multiplication erases, and a zero-parameter post-LN inverse-degree scale outperforms the LayerNorm baseline on influence maximization, dismantling, and maximum independent set.","lead":"A graph-learning paper shows that a positive per-node signal like degree survives LayerNorm only when multiplied after the normalization step, not before it. The resulting parameter-free operator, PostDeg, improves three node-selection tasks by 3.5, 2.5, and 5.6 percent relative to the LayerNorm backbone.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Confound in the same-slot capacity falsifier: learned variants were trained with weight spectral normalization on the GAT backbone, while PostDeg was not; this weakens the 'placement, not parameterization' claim.","rationale":"The algebraic identity in Eq. (1) is standard and the empirical absorption envelope is well-documented. The paper's positive results on three tasks are solid, and the four falsifiers are a good design. However, the same-slot capacity falsifier is confounded by the spectral-normalization asymmetry, which undermines the strongest claim that parameterization is irrelevant. This is a concrete, testable issue rather than a disagreement with consensus. If the proposed ablation is run and equivalence persists, the claim would be supported; until then, the verdict should remain conditional.","tokens_in":58301,"tokens_out":6687,"duration_ms":73095,"concrete_test":"Retrain PostDeg-L-FG and PostDeg-L-Adaptive with use_weight_sn=False (same GAT backbone as PostDeg), keeping all hyperparameters, data, and seeds fixed; optionally also train PostDeg with use_weight_sn=True as a control. Compare paired 10-seed results against PostDeg on InfluMax, Dismantle, and MIS. If either no-spectral-norm learned variant exceeds PostDeg beyond the TOST ±1% margin, the 'placement, not parameterization' claim is weakened; if they remain equivalent, the confound is cleared.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that placement, not parameterization, carries the gain rests on the TOST equivalence between the parameter-free PostDeg and the learned same-slot variants PostDeg-L-FG/Adaptive (Section 4.3, Table A13). However, Appendix C's setup states that 'Learned-variant runs also use weight spectral normalization on the GAT projections through use_weight_sn; PostDeg and the alternative-normalization baselines do not.' Thus the learned variants differ from PostDeg in two ways: (1) they add learnable scale/gate parameters, and (2) they change the GAT backbone via spectral normalization. The same-slot capacity falsifier is therefore not clean: the additional parameters could be beneficial while the spectral norm could be detrimental, and the two effects could cancel within the ±1% margin. The paper never reports PostDeg with spectral normalization or learned variants without it, so the net-zero result cannot be attributed specifically to the irrelevance of parameterization. The fact that learned parameters stay near initialization (Table A25) is suggestive but does not resolve the confound, since spectral norm can also suppress movement in the scale parameters. This gap directly affects the headline conclusion and should be closed before the claim is accepted as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"PostDeg claims that in LayerNorm-based GNNs, a positive per-node scalar multiplier is absorbed when inserted before LayerNorm (Eq. 1, Prop. 16), while the same scalar inserted after LayerNorm survives to the score head as representation magnitude. The paper instantiates this post-LN slot with a parameter-free inverse-degree scale (PostDeg), reports gains over the LN backbone on three node-selection tasks (+3.5%/+2.5%/+5.6%, with 10/10 paired-seed wins per task), and tests four pre-registered falsifiers (graphwise scalar, extra LayerNorm, same-slot learned capacity, and PNA backbone-agnostic source). The central claim is that placement, not parameterization, drives the improvement.","tokens_in":58641,"tokens_out":3767,"duration_ms":41217,"significance":"If the result holds, the paper gives a simple, explainable mechanism for why degree signals are erased in normalized GNNs and a parameter-free fix. The absorption identity is proven and empirically bounded (stabilizer residual ≤2.44×10−5 across tasks, Table A10), the experimental grid is carefully controlled with paired seeds, TOST equivalence tests and Wilcoxon tests are reported, and the code/data layout is unusually detailed. The four falsifier design is a strong example of pre-registered reasoning, and the public prediction sheet for unrun benchmarks is a welcome addition. The main weakness is a confound in the same-slot capacity falsifier that currently prevents the headline 'placement, not parameterization' conclusion from being fully supported.","major_comments":[{"comment":"The same-slot capacity falsifier is confounded by weight spectral normalization. Appendix C states: 'Learned-variant runs also use weight spectral normalization on the GAT projections through use_weight_sn; PostDeg and the alternative-normalization baselines do not.' Thus the TOST equivalence between PostDeg and PostDeg-L-FG/Adaptive (Table A13) shows only that the combined change (extra parameters + spectral norm) is neutral within ±1%, not that parameterization is irrelevant. The additional parameters could be beneficial and the spectral norm detrimental, with cancellation inside the equivalence margin. This gap directly affects the headline claim and should be closed by reporting at least PostDeg with spectral normalization and learned variants without it (or a 2×2 design).","section":"§4.3 and Appendix C setup"},{"comment":"The 'Epidemic = task-objective null' label is a post hoc explanation for a near-zero result. Epidemic has degree skew 0.45, nearly identical to Dismantle (0.46), yet Dismantle gains +2.5% while Epidemic gains +0.1%. The paper attributes this to reward misalignment without giving a formal, a priori definition of 'aligned with degree-conditioned magnitude'. To make the boundary case a genuine test of the placement rule, provide a measurable definition of reward alignment or run an auxiliary experiment that manipulates the reward to flip alignment while holding graph statistics fixed.","section":"§4.2, Table 4"},{"comment":"The theory is proven for a single LayerNorm block (Eq. 1, Prop. 16, Thm. 3), but the experiments use a 3-layer GAT stack. The paper measures per-layer variance at convergence (Table A10) but does not show that the absorption identity remains sharp after residual interactions across layers. The claim that a post-LN scalar 'reaches the score head as representation magnitude' in a 3-layer stack needs one of: (i) a multi-layer analysis, (ii) a control applying PostDeg only at the final layer versus all layers, or (iii) explicit per-layer absorption measurements after the full stack, to bridge this gap.","section":"Appendix C.10 / §3.1"}],"minor_comments":[{"comment":"The Epidemic column is described as a 'task-objective null' but the win rate against LN backbone is 9/10 (Table A17). The near-zero relative improvement with a high win rate should be commented on; the text currently says 'within seed noise' while the win rate suggests a small but consistent difference.","section":"Table 4 / Table A17"},{"comment":"The BA-vs-SBM MIS comparison uses different evaluation sizes (n=200 vs n=300). The caption acknowledges this, but the text calls it a 'dose-response' and reports ratios 1.57×–1.88×. Please present a matched-size comparison or clearly label the result as a directional finding only, as is done in Table A30.","section":"Figure 3(e) and §4.4"},{"comment":"Corollary 5 predicts a log-log slope of −γ for the scalar factor, while Table A20 reports slopes around −0.83 (InfluMax) and −0.63 (MIS). The paper explains this via Remark 2, but it would help to tabulate the predicted slope alongside the empirical slope for each task, making the discrepancy and its explanation transparent.","section":"Equation (3) and Table A20"},{"comment":"The abstract says the placement rule 'generalizes to any positive topology scalar' but only degree is tested; PageRank, k-core, and centrality are explicitly out of scope in the limitations. Suggest softening to 'is expected to generalize' or adding a proof-of-concept for one additional scalar.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The spectral normalization confound is the central issue; the missing control cells should be required before the 'placement, not parameterization' claim can be accepted as stated. The Epidemic boundary explanation should also be formalized or relabeled as post hoc. The paper is otherwise well-executed, with strong reproducibility and a clever falsifier framework, so I see no reason for rejection if these loads-bearing points are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the LayerNorm absorption identity is standard, but locating where in the residual block the topology signal dies, and turning that into a placement diagnostic with four pre-registered falsifiers, is genuinely new. The algebra is correct, the controlled grid is careful, and the parameter-free PostDeg gains are modest but consistent on three tasks. The paper deserves a serious referee.\n\nWhat it does well: the empirical envelope (2.44e-5) confirms the theory, the falsification set is thoughtfully designed, and the limitations section is candid about the single-layer theory and the PNA boundary. The treatment of DD collapse in the variance-zeroing baselines is a nice touch.\n\nWhere I have real doubt: the same-slot capacity falsifier. The learned variants were trained with weight spectral normalization on the GAT projections; PostDeg was not. That means the TOST equivalence between PostDeg and PostDeg-L at ±1% is consistent with the extra parameters helping and the spectral norm hurting by a similar amount. The paper's observation that the learned parameters sit near initialization is suggestive but does not resolve the confound, since spectral norm can also suppress movement. This directly weakens the 'placement, not parameterization' headline. I'd want to see PostDeg with spectral norm and learned variants without it, four cells, before believing that specific claim.\n\nTwo smaller concerns. The Epidemic null is explained post hoc as a 'task-objective null' with no a priori definition of reward alignment, which matters because the same degree skew appears in Dismantle and Epidemic yet the gains diverge. And the theory is proven for a single block but applied to a three-layer stack; the paper concedes this, but it means the headline effect is empirically observed, not explained, at depth.\n\nThe code and data are described in detail but not accessible from the text, so the 10/10 win rates and TOST numbers can't be independently audited. That's fixable.\n\nOverall: the placement rule is likely right, the central algebraic observation is sound, and the empirical work is serious. Fix the SN confound and tighten the Epidemic boundary, and this is a useful contribution. I'd send it to review now rather than desk-reject, but I'd make the SN fix a condition of acceptance.\n\nFor a reading group, this is a good example of a small, well-scoped intervention with honest falsification practice; I'd bring it, but with the SN caveat in hand. I'd cite the placement diagnostic if I were working on degree-sensitive GNN policies. Yes to peer review.","headline":"Solid placement diagnostic, but the 'parameterization is irrelevant' claim rests on a spectral-norm confound that should be fixed before acceptance.","tokens_in":59104,"tokens_out":2172,"would_cite":true,"duration_ms":25332,"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":"A per-node scalar placed after LayerNorm survives to the score head; placed before, it is absorbed (residual at most 2.44e-5). PostDeg, a parameter-free post-LN inverse-degree scale, turns that surviving slot into +2.5% to +5.6% gains on no","keywords":["LayerNorm","graph neural networks","node selection","degree centrality","placement rule","normalization absorption","influence maximization","maximum independent set"],"falsifier":"Raise the LayerNorm stabilizer ε_LN from 10⁻⁵ to a value comparable to feature variance (σ² ≈ 0.5). The absorption identity predicts pre-LN degree multipliers should then survive and PostDeg should lose its edge; if pre-LN multipliers stay erased under a large stabilizer, the mechanism is wrong. Complementarily, on a reward provably aligned with degree magnitude (e.g., sum of selected degrees) versus an unaligned reward at matched degree skew, the rule predicts a gain only on the aligned side.","tokens_in":58167,"feed_emoji":"📍","tokens_out":13053,"duration_ms":121373,"temperature":0.7,"pith_summary":"The paper tries to establish a placement rule for topology signals in LayerNorm-based graph neural networks: a positive per-node scalar (degree, centrality, or any such multiplier) inserted before LayerNorm is numerically erased, while the same scalar inserted after LayerNorm survives as representation magnitude that the score head can read. The rule is grounded in the LayerNorm absorption identity — LN(a*z) approx LN(z) up to a stabilizer term bounded by 2.44e-5 empirically — and is instantiated by PostDeg, a parameter-free inverse-degree scale placed in the surviving post-LayerNorm slot. PostDeg improves over the LayerNorm backbone by +3.5% on influence maximization, +2.5% on network dismantling, and +5.6% on maximum independent set, with 10/10 paired-seed wins per task, while learned same-slot variants are statistically equivalent. Four pre-registered falsifiers — a graphwise spectral scalar, extra normalization, expressive same-slot capacity, and a degree-injecting backbone — all fail to reject the rule, supporting the conclusion that placement, not parameterization, carries the gain.","feed_headline":"Place degree scaling after LayerNorm, not before: 2.5–5.6% gains","feed_subtitle":"A zero-parameter inverse-degree scale in the surviving slot beats learned variants on three node-selection tasks","key_machinery":"The central object is the LayerNorm absorption identity for positive scalars: LN(a_i z_i) ≈ LN(z_i) up to the stabilizer term ε_LN/(a_i²σ²), empirically at most 2.44e-5. The paper turns it into a placement diagnostic: pre-LN positive multipliers are absorbed; post-LN positive multipliers survive as representation magnitude. The companion degree-separation ratio R_ij = ((c_i+ε)/(c_j+ε))^γ > 1 predicts that a post-LN inverse-degree scale s_i = (c_i+ε)^(-1/2) gives low-degree nodes larger magnitude with a negative log-log slope. PostDeg, the parameter-free operator occupying the surviving slot, carries the argument: learned same-slot variants are TOST-equivalent to it, showing the mechanism is","core_discovery":"On its own terms, the paper's discovery is that LayerNorm erases topology signals at a specific, locatable place in the residual block. For positive inputs, LayerNorm is almost exactly scale-invariant, so a per-node multiplier placed before it survives only through the stabilizer term (residual at most 2.44e-5 across nodes, layers, seeds, and tasks), while the same multiplier placed after it multiplies the representation and reaches the score head as magnitude contrast. A parameter-free post-LN inverse-degree scale, PostDeg, exploits that surviving slot for +3.5%/+2.5%/+5.6% on three degree-sensitive node-selection tasks with 10/10 paired-seed wins; learned same-slot variants are TOST-equiva","pith_inferences":["My inference: the placement diagnostic should bind for other positive per-node scalars — PageRank, k-core, eigenvector centrality, learned positive attention — and the paper's own prediction sheet for standard benchmarks is a direct, runnable test of that generalization.","My inference: because the proof covers one LayerNorm block while the experiments stack three, whether the absorption residual compounds across depth is open; a multi-layer mean-field analysis would either harden the rule for deep GNNs or reveal a depth correction.","My inference: the pre/post distinction is a general probe of normalizer erasure — the same move (multiply before versus after LayerNorm) in any normalized residual network, graph or not, should reproduce the absorption pattern, giving architects a cheap test of whether their normalizer hides per-item signals."],"forward_implications":["Pre-LN positive scalar multipliers — degree, PageRank, k-core, learned positive attention scores — are numerically absorbed by LayerNorm in GNN residual blocks, so a policy that must rank by topology needs the signal on the post-normalization side.","The post-LayerNorm slot is a general surviving channel: the same placement rule should hold for any positive per-node topology scalar in any normalized residual stack, not just inverse degree.","A zero-parameter operator in that slot suffices: learned same-slot variants are TOST-equivalent at ±1%, so practitioners gain without adding capacity.","The gains (+3.5% InfluMax, +2.5% Dismantle, +5.6% MIS, 10/10 paired-seed wins per task) persist under 2–3× size transfer, making the fix usable at larger evaluation graphs.","The boundary predictions are part of the rule: low degree heterogeneity (DD) or a reward not aligned with degree-conditioned magnitude (Epidemic) yields a null result, and a backbone that already injects degree inside aggregation (PNA) leaves no headroom."],"fun_headline_variants":["LayerNorm erases degree signals before it, not after: PostDeg gains","Post-LN degree scaling wins: placement beats parameterization","Where LayerNorm kills topology: pre-LN, and PostDeg fixes it","Put degree scale after LayerNorm for up to 5.6% gains"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the absorption identity proven for a single LayerNorm block carries over to the 3-layer stack, and that the Epidemic task shows no gain because its reward is not aligned with degree-conditioned magnitude — an alignment judgment the paper makes post hoc, without a formal a priori definition.","fun_headline_variants_meta":{"raw":{"variants":["LayerNorm erases degree signals before it, not after: PostDeg gains","Post-LN degree scaling wins: placement beats parameterization","Where LayerNorm kills topology: pre-LN, and PostDeg fixes it","Put degree scale after LayerNorm for up to 5.6% gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000645,"raw_usage":{"total_tokens":2813,"prompt_tokens":769,"completion_tokens":2044,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":1965}},"tokens_in":513,"tokens_out":2044,"duration_ms":13446,"temperature":1.0,"reasoning_tokens":1965,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T11:30:32.896474+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Raise the LayerNorm stabilizer ε_LN from 10⁻⁵ to a value comparable to feature variance (σ² ≈ 0.5). The absorption identity predicts pre-LN degree multipliers should then survive and PostDeg should lose its edge; if pre-LN multipliers stay erased under a large stabilizer, the mechanism is wrong. Complementarily, on a reward provably aligned with degree magnitude (e.g., sum of selected degrees) versus an unaligned reward at matched degree skew, the rule predicts a gain only on the aligned side.","supporting_citations":[],"review_version":1}