{"id":"43fc841e-446e-44b7-9754-fd3c3e70de85","arxiv_id":"2607.16254","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Scaling a model's own per-layer feed-forward writes shows signed, task-dependent retrieval roles, and the sign of the write–gradient dot product identifies suppressors from amplifiers.","lead":"The paper scales each layer's own feed-forward network write, one layer at a time, and finds that retrieval is helped by silencing some layers and hurt by silencing others—a signed, task-dependent map. It offers a gradient-alignment score for identifying which layers to attenuate, with a transfer test on long-context QA.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The s_l sign test is not an independent confirmation; it is guaranteed by the first-order formula if response surfaces are linear, and the paper never tests curvature over [0,2].","rationale":"The reader's weakest-assumption identification matches my own reading. The paper's headline claim about s_l separating suppressor/amplifier roles is the most load-bearing part of the mechanism story, and it is indeed potentially circular given the first-order formula. The direct causal intervention results (signed response surfaces, role switches) are well-supported and not threatened by this concern; what is threatened is the claim that the sign of the local directional derivative is an independent confirmation of the taxonomy. The paper provides R² values for the attenuation side, which cover only part of the interval, and does not report curvature or lack-of-fit checks for the full [0,2] range. The proposed test—examining the sign separation on high-curvature layers—would directly settle whether the mechanism holds where the first-order approximation breaks down. If the separation persists there, the concern is resolved; if not, the mechanism claim should be weakened to a linearity-dependent description. Since the reader already assigned CONDITIONAL and this test would strengthen or condition the claim further, the verdict remains UNCHANGED: the concern is real but does not invalidate the paper's empirical core, and my proposed check is a concrete path to resolution.","tokens_in":17643,"tokens_out":6508,"duration_ms":59326,"concrete_test":"For every layer/suite in Table 4, fit D_l(α) to a quadratic over the full grid α∈{0,...,2} and compute the curvature coefficient (or the difference between the secant slope (D(2)−D(0))/2 and the local derivative −s_l). Then recompute the suppressor/amplifier sign-separation statistics on the subset of layers with curvature above a threshold (e.g., where the quadratic term is significant or where the linear fit R² < 0.5). If the 34/35 and 18/18 counts remain near-perfect in this high-curvature subset, the mechanism is robust; if they degrade, the sign test is a linearity artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism claim (Q3) rests on the cross-tab between endpoint-derived roles and the local directional derivative s_l = ΔF_l^T ∇M. Suppressors are defined by D(0)<0 and D(2)>0; amplifiers by D(0)>0 and D(2)<0. Under the paper's own first-order expansion (§3.3), D_l(α) ≈ (1−α)s_l, so for a suppressor the endpoints imply s_l<0 (D(0)≈s_l, D(2)≈−s_l), and for an amplifier s_l>0. Thus the 34/35 and 18/18 separations are logically entailed by role definitions whenever the response surface is approximately linear over α∈[0,2]. The paper reports R² for the attenuation side (α∈{0.25,0.5,0.75}), but the endpoint D(2) lies at α=2, outside that range; no curvature or lack-of-fit test for the full [0,2] interval is provided. The low semantic R² values (e.g., 0.074 for Qwen2.5) show the predictor can fail in amplitude, and the paper then appeals to sign/rank signal—but the sign signal could be inherited from the endpoint definitions via monotonicity, not from a genuine local-direction mechanism. Without a direct check of whether s_l predicts endpoint signs in layers where the response is strongly nonlinear, the 'mechanism anchor' is potentially a restatement of the role definitions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies whether native FFN residual writes in Transformer LMs have signed, task-conditioned causal effects on long-context retrieval. The intervention scales each layer's own FFN write by a gain α (0.25, ..., 2) without editing weights, measures retrieval-margin damage D_l(α), and classifies layers into calibrated, suppressor, amplifier, and weak/mixed roles from endpoint signs at α=0 and α=2. A local directional derivative s_l = ΔF_l^T ∇M is then used as a mechanism anchor: the paper reports that suppressors have s_l<0 in 34/35 cases and amplifiers have s_l>0 in 18/18 cases, that 60% of layers switch roles between literal and semantic retrieval, and that s_l predicts attenuation damage with raw R²≈0.79 on a safety-filtered LongBench retrieval-QA probe. The paper also reports threshold sweeps, leave-one-seed-out stability, norm/depth controls, and a held-out suppressor-attenuation policy check.","tokens_in":18016,"tokens_out":10229,"duration_ms":96036,"significance":"If the findings hold, they would sharpen the standard FFN-as-memory view by showing that the direction of the residual write is causally relevant and task-dependent in retrieval settings, and they would offer a compact post-hoc diagnostic for FFN replacement or decoupling studies. The study has notable strengths: direct native-scaling interventions, complete layer-gain grids with auditable prediction files, explicit robustness analyses (threshold sweep, seed LOO, norm controls), a zero-parameter first-order predictor, and a cross-protocol model (SmolLM2). However, the central mechanism test in §5.3 is not fully independent of the role definitions, and the transfer claim uses different metric emphases across protocols. These issues are fixable but require additional analysis.","major_comments":[{"comment":"The 34/35 and 18/18 sign separation is not an independent confirmation of the s_l mechanism. Roles are defined by endpoint signs at α=0 and α=2 (§3.2). Under the paper's own first-order relation D_l(α)≈(1−α)s_l, a suppressor (D(0)<0, D(2)>0) entails s_l<0 and an amplifier (D(0)>0, D(2)<0) entails s_l>0, provided the response is approximately linear over [0,2]. The paper validates the first-order predictor only for α∈{0.25,0.5,0.75} (§5.3, Table 5) and does not test curvature over the full interval containing α=2. The 61 calibrated layer-suite pairs in Table 4 show the response is not globally linear. Please add a direct curvature/lack-of-fit test (e.g., quadratic fit per layer, or a sign test restricted to layers with small curvature) or redefine roles from attenuation-only data and treat endpoints as a prediction.","section":"§5.3 and §3.3"},{"comment":"The headline transfer R²=0.796/0.791 is computed on a post-hoc selected LongBench subset and is emphasized even though the same diagnostic has raw R²=0.074 (Qwen2.5 semantic, Table 5) and 0.139 (Pythia semantic). The paper explains the controlled-semantic gap by shifting to Pearson/MCC, but then uses raw R² on LongBench as evidence of quantitative transfer. To be comparable, report the same diagnostic set (raw R², Pearson, sign accuracy, MCC) for both protocols and specify whether the subset selection and R² metric were fixed before seeing the results. Otherwise the transfer claim is not apples-to-apples.","section":"§5.4, Table 6"},{"comment":"The held-out policy sanity check selects top suppressor layers using the same single-layer margin-damage screen that defines the suppressor role. Attenuating those layers on a disjoint split is therefore a consistency check of the taxonomy, not an independent test of the s_l diagnostic. To support the diagnostic claim, select layers using s_l alone (or using attenuation data that excludes the endpoint definitions) and show that the held-out policy still outperforms controls.","section":"§5.4, Appendix D"}],"minor_comments":[{"comment":"The phrase 'final FFN is a suppressor in 7 of 8 model-suite cases' is correct only when counting SmolLM2 semantic as calibrated; this is stated later, but the abstract could be clearer that the exception is a semantic-suite calibrated role.","section":"Abstract"},{"comment":"The retrieval-margin definition should specify token-level aggregation: is the margin the mean over gold tokens of per-token logit margins, or the margin computed from mean logits? This affects the interpretation and reproducibility of s_l.","section":"§4"},{"comment":"Table 5 leaves literal Pearson/MCC entries as dashes. Please fill them or explicitly state why they are omitted, so readers can compare all diagnostics across protocols.","section":"§5.3, Table 5"},{"comment":"The 'both-help' category appears in the LongBench transfer table but is not in the main controlled taxonomy Table 3. The relationship between 'both-help' and 'weak/mixed' should be stated explicitly.","section":"§5.4"},{"comment":"The 'Qwen3.5-9B-Base' model needs a public identifier or a detailed configuration (tokenizer, training data, availability) for reproducibility; as written it is ambiguous.","section":"§7"}],"recommendation":"major_revision","confidential_remarks":"The stress-test note's concern about the s_l sign test is valid: the near-perfect separation in Table 4 is largely entailed by the endpoint role definitions unless the response surfaces are nonlinear over [0,2]. I would ask the authors to add a curvature test or attenuation-only role definition before acceptance. The rest of the robustness work is solid, and the paper is publishable in principle after this issue is addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First take: this is a careful empirical paper on a real question. The claim that a native FFN write can have a signed, task-conditioned causal effect on retrieval margin is well supported by direct interventions, threshold sweeps, seed LOO, and norm/rotation controls. The LongBench transfer result (zero-parameter D_pred=(1-alpha)s_l giving R^2 ~ 0.8) is a genuinely useful diagnostic for people who edit, prune, or decouple FFN memory. You should read it if you work in that area.\n\nWhat's new: the signed taxonomy (calibrated/amplifier/suppressor) and the observation that a layer can flip role between literal and semantic retrieval. The opposite-sign examples are convincing, and the final-layer suppressor pattern in 7/8 model-suite pairs is a clean result.\n\nNow the soft spots, in order. The biggest issue is the mechanism anchor in §5.3. Roles are defined from endpoint damage D(0) and D(2). The paper's own first-order expansion says D(alpha) ≈ (1-alpha) s_l. For a suppressor, D(0)<0 and D(2)>0 implies s_l≈D(0)<0; for an amplifier, D(0)>0 and D(2)<0 implies s_l>0. So the 34/35 and 18/18 separation is largely entailed by the role definitions plus a linearity assumption over [0,2]. The paper fits R^2 only on attenuation gains (0.25–0.75), all on one side of the native point, and never tests curvature over the full interval. The claim that s_l 'explains' the roles therefore overstates what the data show. It's a good consistency check, not an independent confirmation. The paper would be stronger if it either said this plainly or examined layers with nonlinear endpoint responses.\n\nSecond, the role-switch rate: 60% with 95% CI [50%, 69%]. That CI includes 50%, so 'most layers switch' is not statistically established, though the specific opposite-sign layers are real.\n\nThird, reproducibility: Qwen3.5-9B is an unreleased checkpoint and the code isn't linked. The paper mentions artifacts, but no URL. That will matter to referees.\n\nNone of this breaks the empirical taxonomy; it just narrows the mechanism claim. Who should read it: people in mechanistic interpretability, model editing, and memory-modular architectures. It's a solid within-subfield contribution.\n\nRecommendation: send to peer review. The taxonomy and transfer diagnostic deserve careful refereeing, with the mechanism claim reframed as a consistency check or tested with a curvature analysis.","headline":"Solid empirical taxonomy of signed FFN writes in retrieval; the s_l sign test is a consistency check rather than an independent mechanism test.","tokens_in":18485,"tokens_out":3475,"would_cite":true,"duration_ms":32274,"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":"A transformer feed-forward layer's native residual write is a signed, task-conditioned steering signal: the same layer can push retrieval margins up or down depending on retrieval mode, and the write-gradient sign predicts the direction.","keywords":["feed-forward networks","residual stream","mechanistic interpretability","long-context retrieval","directional derivative","signed roles","retrieval margin","attenuation"],"falsifier":"A central finite-difference check of the derivative at α=1 (for example, comparing margins at α=0.9 and 1.1) on every suppressor and amplifier layer, computed without invoking the first-order formula; if the sign of the measured derivative agrees with the endpoint-defined role at near-chance rates across a full layer grid, the claim that s_l anchors the monotone roles fails. Alternatively, one suppressor-classified layer with a U-shaped response (both endpoints hurt but a local maximum at α=1) would show the endpoint taxonomy can mislabel non-monotone layers.","tokens_in":17528,"feed_emoji":"🧠","tokens_out":6493,"duration_ms":53983,"temperature":0.7,"pith_summary":"The paper tries to establish that the feed-forward networks (FFNs) of large language models are not only parametric memories but also directional actuators in the residual stream: in long-context retrieval, a layer's native write can push the model toward or away from the correct answer. By scaling each layer's own FFN write up and down without editing weights or injecting external vectors, the author finds signed response surfaces: some layers are suppressors (removing them helps, amplifying them hurts), some are amplifiers, and most layers switch role when retrieval changes from literal to semantic. The local directional derivative of retrieval margin along the native write, s_l = ΔF_l^T ∇M, separates the two monotone roles almost perfectly (34/35 suppressors negative, 18/18 amplifiers positive) and predicts attenuation damage on a natural-language QA probe with R²≈0.79. A sympathetic reader would care because it gives a compact diagnostic for which FFN layers to attenuate or replace in retrieval settings, and because it reframes FFN interpretability from storage to direction.","feed_headline":"FFN write direction—not size—predicts retrieval damage","feed_subtitle":"Scaling a model's own feed-forward write shows some layers suppress retrieval; a simple gradient-sign test finds them.","key_machinery":"The central object is the native FFN residual write ΔF_l at each layer, probed by the gain-scaled forward pass h_{l+1} = u_l + αΔF_l and summarized by D_l(α) = M_base − M_l(α). The explanatory quantity is s_l = ΔF_l^T ∇M, the inner product of the native write with the retrieval-margin gradient—the local directional derivative of margin along the write. It anchors the taxonomy: negative s_l means the write points against retrieval (suppressor), positive means with retrieval (amplifier), and the zero-parameter predictor D_l(α)≈(1−α)s_l quantifies attenuation damage.","core_discovery":"On the paper's own terms, the discovery is that the vectors a feed-forward sublayer adds to the residual stream—its native writes—have causal, signed, layer-specific effects on retrieval. Scaling one layer's own write by a gain, without weight edits or external steering, produces response surfaces that are not uniform degradation: some layers are suppressors (removing the write helps, amplifying it hurts), some are amplifiers, and many are calibrated or weak. In controlled literal and semantic retrieval, the final layer is a suppressor in 7 of 8 model–suite pairs and 60% of layers switch role between retrieval modes. The local directional derivative s_l = ΔF_l^T ∇M separates the two monotone","pith_inferences":["If the directional-derivative diagnostic holds across more architectures, a cheap post-hoc pass could flag individual FFN layers whose writes oppose the current task objective, without needing a full gain-scaling grid.","The same sign logic could extend beyond retrieval: whenever a readout margin or loss gradient is available, s_l may characterize whether a layer amplifies or suppresses that objective, making the method a general per-layer objective-alignment audit.","Because the paper leaves training dynamics open, one testable extension is to track s_l across checkpoints to see when suppressor and amplifier roles form and whether they are stable or reorganized during training.","The role-switch result suggests attention-focused explanations of literal-versus-semantic retrieval gaps are incomplete: some of the divergence is localized in FFN write direction, which could be validated by ablating the opposite-sign layers reported in the paper."],"forward_implications":["Final-layer FFNs in controlled long-context retrieval are usually readout-adjacent suppressors: removing their write improves retrieval margins, so the output-adjacent FFN can inject information that competes with the retrieved answer rather than only storing facts.","An FFN layer's role is not a fixed property of what it stores: about 60% of layers flip between suppressor and amplifier roles when the retrieval mode shifts from literal to semantic.","The sign of s_l = ΔF_l^T ∇M gives a compact, parameter-free diagnostic for the two monotone roles and for attenuation damage (R²≈0.79 on natural-language QA), so write direction—not just write size—should be measured when evaluating FFN importance.","A held-out policy that attenuates the strongest suppressor layers improves retrieval margins over random and norm-matched layer sets, indicating suppressors are causally useful control points.","For architectures that decouple, sparsify, or replace FFN computation, the paper implies lookup accuracy alone is insufficient; the replacement write's direction relative to the retrieval gradient should also be checked."],"fun_headline_variants":["Signed FFN writes predict retrieval suppression","Gradient sign flags harmful FFN layers","FFN write sign, not magnitude, shapes retrieval","Layer-by-layer FFN scaling exposes suppression","Diagnostic: negative write gradient means suppressor"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the retrieval margin depends approximately linearly on the scaling gain between α=0 and α=2 for monotone layers, so endpoint signs and the local derivative s_l are measuring the same directional fact; if many layers are strongly nonlinear across that interval, the near-perfect sign separation would be an artifact of role definitions rather than an independent confirmation.","fun_headline_variants_meta":{"raw":{"variants":["Signed FFN writes predict retrieval suppression","Gradient sign flags harmful FFN layers","FFN write sign, not magnitude, shapes retrieval","Layer-by-layer FFN scaling exposes suppression","Diagnostic: negative write gradient means suppressor"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000171,"raw_usage":{"total_tokens":1139,"prompt_tokens":808,"completion_tokens":331,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":278}},"tokens_in":552,"tokens_out":331,"duration_ms":3356,"temperature":1.0,"reasoning_tokens":278,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T09:43:51.421964+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A central finite-difference check of the derivative at α=1 (for example, comparing margins at α=0.9 and 1.1) on every suppressor and amplifier layer, computed without invoking the first-order formula; if the sign of the measured derivative agrees with the endpoint-defined role at near-chance rates across a full layer grid, the claim that s_l anchors the monotone roles fails. Alternatively, one suppressor-classified layer with a U-shaped response (both endpoints hurt but a local maximum at α=1) would show the endpoint taxonomy can mislabel non-monotone layers.","supporting_citations":[],"review_version":1}