{"id":"54a222bb-f7ef-4d44-a654-7ad1b55a2d93","arxiv_id":"2411.16252","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"NormXLogit multiplies a token's input-embedding norm by the logit its final-layer representation receives from the head, yielding faithful, per-token attributions in one forward pass.","lead":"NormXLogit is a new method for identifying which words most influenced a language model's prediction, using each word's embedding size and the model's own output head. It runs in a single forward pass, making it much cheaper than gradient-based alternatives.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The core assumption that a [CLS]-trained head yields meaningful logits for all tokens is unvalidated; a control-head AOPC test would determine whether NormXLogit's faithfulness comes from the head or from norm/position artifacts.","rationale":"The reader's weakest assumption correctly identifies the pivotal step: LogAt applies a head trained exclusively on [CLS] to all token positions, and NormXLogit multiplies that score by an input-norm term. If the head outputs on non-[CLS] tokens are meaningless, then neither LogAt nor NormXLogit is a task-specific attribution; the method would reduce to a norm-weighted heuristic. The existing Table 1 provides partial, indirect evidence that the head contributes: LogAt often outperforms the norm baseline, but in some configurations (e.g., DeBERTa on STS-B) the norm alone is better than NormXLogit, so the contribution is inconsistent. The BLiMP result also relies on LogAt applied per layer, but its headline comparison uses the token 'plural' selected for semantic fit to number agreement, and the random-word controls are only mentioned qualitatively. I do not treat these as evidence of bad faith; they are addressable with controls. The control-head experiment directly isolates whether the trained head carries the task signal, which is the condition on which the central claim rests. Since the paper currently does not provide that control, the conditional verdict is appropriate; no change to the reader's verdict is needed.","tokens_in":16493,"tokens_out":4802,"duration_ms":49925,"concrete_test":"Use the same fine-tuned SST-2 BERT/DeBERTa model. Shuffle the weights of the classification head (or retrain it on randomly permuted labels on the same [CLS] representations), recompute LogAt and NormXLogit attributions on the test set, and measure AOPC over K = 10-90%. Compare with the original head and the random baseline. If control-head AOPC stays close to the original (within about 0.02), the head is not the source of faithfulness and the central claim weakens; if it drops to the random baseline, the head-on-top assumption is validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Eq. (2) and Section 3.3: LogAt applies the task head, trained only on the [CLS] (or final) token, to every token's final-layer representation and reads off the predicted-class logit. The paper's only justification is the residual-stream 'preserving effect' from Kobayashi et al. (2021), which concerns token identity in attention outputs, not whether a head trained on [CLS] produces meaningful task logits on non-[CLS] representations. If those logits are arbitrary or dominated by position and frequency artifacts, NormXLogit (Eq. 4) is a norm-weighted heuristic rather than a task-specific attribution. The regression variant (Eq. 3) is even more fragile: absolute distance from the [CLS] output can be high for any out-of-distribution token representation without indicating task contribution. A control experiment is needed to establish that the head, not the norm, is what makes NormXLogit faithful.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NormXLogit, a token-attribution method for transformer models. The attribution of token i is the product of the ℓ2 norm of its input embedding and a logit-attribution score (LogAt), obtained by applying the model's classification or regression head to the token's final-layer representation and reading off the predicted-class logit (or, in regression, the absolute distance from the head's output for the [CLS] token). The method requires only a single forward pass. The authors evaluate faithfulness via AOPC and accuracy-drop metrics on SST-2, MNLI, QNLI, and STS-B with BERT, DeBERTa, and LLAMA 2, comparing against gradient-based baselines, DecompX, a random baseline, and an ℓ2-norm-only baseline. They also evaluate plausibility by measuring alignment between attributions and known evidence tokens in a BLiMP subset with RoBERTa. The central claims are that NormXLogit is more faithful than gradient-based methods and competitive with architecture-specific methods, while being computationally much cheaper.","tokens_in":95,"tokens_out":5009,"duration_ms":114881,"significance":"If the head-on-top assumption is valid, NormXLogit is an appealing method: it is model-agnostic, task-specific, per-label, and extremely efficient (one forward pass, no backpropagation, with reported speedups and large memory savings). The paper provides a broad empirical comparison across three model families, four tasks, and several baselines, which is a useful contribution. The method's simplicity and the clear presentation of the AOPC and accuracy-drop protocols are strengths. However, the core assumption that a head trained only on [CLS] produces meaningful logits for all token representations is not validated, and several empirical claims are stronger than the reported results warrant. The BLiMP plausibility experiment is weakened by a post hoc probe-word selection. These issues are addressable but require additional experiments and careful rephrasing.","major_comments":[{"comment":"The central assumption that applying the [CLS]-trained classification head to every token's final-layer representation yields meaningful task-specific logits is not validated. The cited 'preserving effect' from Kobayashi et al. (2021) concerns the dominance of the residual connection in the decomposition of attention outputs; it does not establish that a head trained only on the [CLS] token produces valid out-of-distribution logits for non-[CLS] token representations. If those logits are arbitrary or dominated by position/frequency artifacts, NormXLogit in Eq. (4) reduces to a norm-weighted heuristic. I request a control-head AOPC experiment: compare NormXLogit with the actual task head against NormXLogit with (a) a randomly initialized head and (b) a head trained on a different task. If the random-head variant achieves comparable AOPC, the task-specificity claim is falsified. This is the load-bearing step for the paper's central claim.","section":"Section 3.3, Eq. (2)"},{"comment":"The claim that NormXLogit 'surpasses' gradient-based methods is not supported in all reported configurations. In DeBERTa/MNLI, Integrated Gradients achieves AOPC 0.571 versus NormXLogit's 0.566; in DeBERTa/STS-B, IG has Accuracy 0.227 versus NormXLogit's 0.320 (lower is better). Several other cells differ by less than 0.01. The paper should either qualify the claim (e.g., 'in most settings') or provide statistical significance testing and error bars. Without variance estimates, differences of 0.005-0.01 in AOPC cannot be distinguished from noise, and the current presentation overstates the consistency of the advantage.","section":"Table 1 and Section 4.2"},{"comment":"The BLiMP evidence-alignment experiment is weakened by post hoc probe selection. The paper reports that LogAt('plural') achieves notably high alignment and attributes this to the number-agreement phenomena in the data, but 'plural' is semantically related to the phenomena under study and is not a random probe. The comparison to two other randomly selected words does not constitute a systematic evaluation. The plausibility claim should be supported by a pre-specified set of probe words or by aggregating over a larger vocabulary subset, rather than by selecting a single word after observing the results.","section":"Section 5.2 and Appendix A.3"}],"minor_comments":[{"comment":"The title and abstract contain a spacing typo: 'Head-on-T opNever Lies' should read 'Head-on-Top Never Lies'.","section":"Abstract and title"},{"comment":"The regression attribution, defined as the absolute distance of the head's output for token i from its output for [CLS], needs a more explicit justification; as written, any out-of-distribution token representation will tend to have a large distance, and the paper does not explain why this measures task contribution rather than representation shift.","section":"Section 3.3, Eq. (3)"},{"comment":"No code or detailed implementation settings are provided for the baselines (e.g., the exact parameters for Integrated Gradients beyond the number of steps, or the handling of subword tokens in aggregation); releasing code would substantially improve reproducibility.","section":"Section 4.1"},{"comment":"The table reports single values without error bars or information about multiple runs; adding variance estimates (or at least a statement that results are deterministic given a fixed seed) would help readers interpret small differences.","section":"Table 1"},{"comment":"The explanation for the strength of the ℓ2 norm on STS-B ('similar openings... less frequent words toward the end') is speculative; either provide supporting analysis (e.g., token-level norm statistics) or soften the claim.","section":"Section 4.2"},{"comment":"The captions report the layer-0 alignment for the ℓ2 norm but not for LogAt or NormXLogit; including those values would make the layer-wise comparison more complete.","section":"Figures 4 and 5"}],"recommendation":"major_revision","confidential_remarks":"The comparison set includes several baselines (GlobEnc, ALTI, DecompX) from the same research group as the authors. This is not a circularity issue, but the editor may want to verify that the implementations of these baselines are standard and that no unintended advantage or disadvantage arises from implementation choices. Additionally, the paper does not mention code availability; for an empirically driven interpretability paper, releasing code would strengthen the submission. The central conceptual concern (head-on-top validity) is testable with a control experiment, so I do not see the issues as fatal, but they are substantial enough to require a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: NormXLogit is a single-forward-pass attribution that combines input embedding norm with the logit produced by applying the task head to each token's final representation. The combination is new, and the experiments mostly support the claim that it beats gradient-based baselines on faithfulness for the tested configs. The gain is not huge everywhere—DeBERTa on STS-B is a clear counterexample—but the consistency across models and tasks is enough to take seriously. Per-label attributions and the runtime/memory table are genuinely useful; 750 samples per batch vs 1–2 for DecompX/IG is a real practical win.\n\nThe soft spots are the usual ones plus one specific one. No code, no error bars on the central tables, so we don't know whether the reported margins survive noise. The BLiMP 'plural' result is presented as a striking finding but it's one word chosen after the fact, and the random-word controls are only mentioned in prose in the appendix. The regression formula (absolute distance from CLS output) is ad hoc and the mixed results show it: on DeBERTa STS-B, NormXLogit is at 0.320 accuracy vs IG's 0.227. The bigger conceptual issue is the load-bearing assumption in Eq. (2): that a head trained on [CLS] produces meaningful logits on every token's final representation. The 'preserving effect' they cite is about attention outputs, not about the task head. The stress-test note's control experiment—randomizing the head or comparing against a norm-only baseline with a scrambled head—is exactly what would settle whether the head is doing work or whether NormXLogit is mostly a norm-weighted heuristic. The ℓ2 norm baselines in Table 1 are already suspiciously strong, which makes that control necessary rather than nice-to-have.\n\nStill, the central claim holds up in the limited sense tested: their attributions remove more prediction probability than gradients on most settings, and the method costs a single pass. The limitations section is honest about the lack of a gold standard. This deserves a serious referee. I'd ask for code, repeated-run error bars, the head-scrambling control, and random-word controls for BLiMP before accepting. As is, it's a useful method to cite and compare against.","headline":"A cheap single-pass token attribution that mostly works, but the headline numbers need error bars and the head-on-top assumption needs a control before I'd sign off.","tokens_in":17158,"tokens_out":2193,"would_cite":true,"duration_ms":21536,"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":"Applying the model's own classification head to every token yields faithful attributions in a single forward pass.","keywords":["NormXLogit","token attribution","explainability","interpretability","head-on-top","transformer interpretability","faithfulness","single forward pass"],"falsifier":"Replace the fine-tuned head with a randomly initialized head of the same shape and rerun the AOPC experiment: if the rankings still match the model's sensitivity to token deletion as well as the trained-head version, then the logit term is not carrying task information; a second check is to measure, across many examples, whether the per-token logit magnitude correlates with the actual drop in the predicted-class probability when that token is deleted.","tokens_in":16212,"feed_emoji":"🔍","tokens_out":6433,"duration_ms":106815,"temperature":0.7,"pith_summary":"NormXLogit claims that token importance in a transformer can be read directly from two cheap quantities: the $\\ell^2$ norm of each token's input embedding and the logit produced when the model's task head is applied to that token's final representation. The paper argues that because the residual connection preserves each token's identity through the layers, the head trained on the [CLS] token also acts as a meaningful per-token interpreter, and that multiplying the resulting logits by the embedding norm sharpens them into per-label attributions with signs. On SST-2, MultiNLI, QNLI and STS-B across BERT, DeBERTa and LLAMA 2, the paper reports that NormXLogit beats Gradient Norm, Gradient×Input and Integrated Gradients on faithfulness, and is competitive with the architecture-specific DecompX; on the BLiMP language-modeling tasks its target-token attributions align better with known linguistic evidence than the comparison methods. A sympathetic reader would care because this would make faithful token attribution a single forward pass, with no backpropagation and up to hundreds of times lower memory use.","feed_headline":"One forward pass beats gradient token attributions in transformers","feed_subtitle":"NormXLogit scores each token from its embedding norm and the model's own task head, no backprop needed.","key_machinery":"The load-bearing object is the product of two per-token scalars: the $\\ell^2$ norm of the input word embedding, $\\|x_i^0\\|_2$, and the logit assigned to token $i$ by applying the head-on-top (the classification, regression, or language-modeling head trained on the [CLS] or final token) to the token's last-layer representation $x_i^L$. The norm term imports the information-gain property of embeddings, while the logit term imports the task-specific decision boundary; the preserving effect of the residual connection is what justifies treating per-token head outputs as reflections of each token's causal contribution rather than noise. The same machinery yields per-label attributions by indexing logits for classes other than the predicted one, and yields per-layer explanations by applying it to intermediate layers.","core_discovery":"The central claim is that the head-on-top never lies: for a model with $L$ layers, computing $\\mathrm{Att}_{\\mathrm{LogAt}}(x_i) = \\mathrm{HoT}_{\\mathrm{clas}}(x_i^L)[\\hat{p}]$ for classification (or the absolute distance from the [CLS] prediction for regression) turns the task head into an attribution function over every input token, and the full method $\\mathrm{Att}_{\\mathrm{NormXLogit}}(x_i) = \\|x_i^0\\|_2 \\cdot \\mathrm{Att}_{\\mathrm{LogAt}}(x_i)$ adds the information content carried by the word embedding norm. The paper argues this works because attention's preserving effect keeps token identity in the final-layer representations, so the logits reflect each token's contribution to the decision, and because higher-norm embeddings carry more information and therefore should weight attributions more heavily. The paper then claims, with the perturbation-based AOPC/Accuracy experiments and the BLiMP evidence-alignment experiments, that these attributions are more faithful than gradient-based methods and competitive with vector-based, architecture-specific methods.","pith_inferences":["A natural test the paper does not run: randomize or swap the task head; if NormXLogit still tracks the model's behavior after perturbation, the logit term is not doing task-specific work, so this experiment would isolate how much of the attribution comes from the head versus the norm.","Because the method reads hidden states rather than gradients, it is immune to the gradient-manipulation attacks the paper cites as a motivation; however, adversarial perturbations of the input still alter the embeddings and therefore the attributions, so 'manipulable' explanations may simply move to input space.","The same construction should transfer to non-text transformers with a pooling token, e.g., vision transformers, by substituting the [CLS] convention with the equivalent pooled representation; this is a testable extension in the spirit of the paper's stated future work.","The BLiMP layer curves suggest a diagnostic use: the layer at which target-vs-foil logit alignment diverges may localize where a grammatical feature becomes task-decisive, which could be compared across models or fine-tuning checkpoints."],"forward_implications":["NormXLogit turns any transformer with a task head into its own explainer: attributions for a prediction cost one forward pass and no gradients, so they scale to models where Integrated Gradients or DecompX would be impractical.","Per-label scoring means one can ask which tokens support a specific class, not just the predicted one; in language modeling this extends to arbitrary vocabulary items, as the paper's LogAt(\"plural\") result demonstrates.","Applying the same formula at each layer produces layer-wise explanations without rollout-style aggregation, and on BLiMP these layer-wise alignments increase with depth for the target token.","Since only the head and the last-layer representations are used, the method transfers across BERT-like, DeBERTa-like, and decoder-only architectures with no architecture-specific code."],"supporting_citations":[{"why":"Supplies the finding that word-embedding norms encode information gain, the foundation of the norm factor.","marker":"Oyama et al. (2023)"},{"why":"Establishes the preserving effect of the residual connection, which justifies interpreting per-token logits as contributions.","marker":"Kobayashi et al. (2021)"},{"why":"Provides Integrated Gradients, a principal gradient-based baseline that NormXLogit aims to beat.","marker":"Sundararajan et al. (2017)"},{"why":"Supplies BLiMP, the linguistic minimal-pair benchmark used to measure plausibility and evidence alignment.","marker":"Warstadt et al. (2020)"},{"why":"Supplies the BLiMP paradigm subset and evaluation setup for evidence alignment.","marker":"Mohebbi et al. (2023)"},{"why":"Defines the AOPC faithfulness metric used across the perturbation experiments.","marker":"Samek et al. (2015)"},{"why":"Provides BERT and the [CLS] token convention on which LogAt is built.","marker":"Devlin et al. (2019)"},{"why":"Provides LLAMA 2, the decoder-only model used to demonstrate generality beyond encoder models.","marker":"Touvron et al. (2023)"}],"fun_headline_variants":["The head-on-top never lies: token attribution with one pass","NormXLogit: token weight from embedding norm and head logits","One forward pass, zero backprop: better token attributions","LLM token importance via norm and final logits, no gradients","Why the task head reveals which tokens matter in LLMs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that a classification or regression head trained only on the [CLS] token still outputs meaningful task-specific scores when applied to the final representation of any other token; if those outputs are arbitrary, NormXLogit shrinks to a norm-weighted noise signal.","fun_headline_variants_meta":{"raw":{"variants":["The head-on-top never lies: token attribution with one pass","NormXLogit: token weight from embedding norm and head logits","One forward pass, zero backprop: better token attributions","LLM token importance via norm and final logits, no gradients","Why the task head reveals which tokens matter in LLMs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000807,"raw_usage":{"total_tokens":3535,"prompt_tokens":928,"completion_tokens":2607,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":2519}},"tokens_in":544,"tokens_out":2607,"duration_ms":17577,"temperature":1.0,"reasoning_tokens":2519,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:19:24.225228+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the fine-tuned head with a randomly initialized head of the same shape and rerun the AOPC experiment: if the rankings still match the model's sensitivity to token deletion as well as the trained-head version, then the logit term is not carrying task information; a second check is to measure, across many examples, whether the per-token logit magnitude correlates with the actual drop in the predicted-class probability when that token is deleted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies BLiMP, the linguistic minimal-pair benchmark used to measure plausibility and evidence alignment."}],"review_version":1}