{"id":"8f03186c-1ef7-4237-9e9c-00193e87bbd7","arxiv_id":"2505.21224","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"NMT encoders detect grammatical errors in early layers and move the error's representation toward the clean form in later layers; fine-tuning on noisy text increases reliance on the attention heads that do this work.","lead":"What happens inside a translation model when it meets a sentence with a grammatical mistake? This paper peers into neural machine translation encoders and reports that they first detect the wrong word and then nudge its internal representation toward the correct version. It also names a set of attention heads, called Robustness Heads, that it claims become more important after fine-tuning on noisy data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Correction evidence in §5.3 lacks a shared-context control; Figure 4's Robustness/Influential overlap inherits the same convergence confound.","rationale":"The reader's verdict is CONDITIONAL, and my analysis supports that. The load-bearing point is the Section 5.3 interpretation: the CKA distance curves are compatible with a generic context-pooling null hypothesis, and the paper explicitly offers that explanation for Base and Clean-Finetuned models but does not run the control that would distinguish it for Noise-Finetuned models. Without such a control, the phrase 'moving its representation toward the correct form' (RQ1) is an interpretation rather than a demonstrated mechanism. I also verified the Reader's additional concern about Figure 4: under the definitions in Section 3.3.4, when the noisy and clean representations converge (the Section 5.3 result), the ranking of heads by 1-CKA(masked_noisy, clean) and by 1-CKA(masked_noisy, noisy) coincide, so the reported high accuracy for Noise-Finetuned models is at least partly definitional. This does not make the paper's other contributions invalid: the fine-tuning results, GED probing curves, and POS attention analyses are independently useful, and the code is released. The concern is addressable with a control experiment; hence the verdict stays CONDITIONAL rather than REJECT or UNVERDICTED. I propose one concrete check that would settle it.","tokens_in":23238,"tokens_out":7983,"duration_ms":82883,"concrete_test":"Run a control experiment on En-Es with, say, NLLB and OPUS-MT: for each noisy sentence, compute the Section 5.3 CKA distance not only between the ungrammatical word and its clean form, but also between the ungrammatical word and a matched control word in the same sentence (same POS, different surface form), layer by layer, for Base and Noise-Finetuned models. If the Noise-Finetuned advantage is specific to the clean form (distance decreases substantially faster than to the control word), targeted correction is supported; if both decrease at similar rates, the effect is generic context integration. In the same pass, recompute the Robustness-vs-Influential head overlap using the control word as the reference; if the overlap is similarly high for control pairs, the RQ3 claim should be downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central evidence for RQ1 and RQ3 rests on the Section 5.3 CKA distance between the set of ungrammatical word representations and the set of their clean forms. Because each noisy sentence differs from its clean counterpart only in the target word, the two representation sets share nearly all context; any encoder that pools context will drive the sets together with depth, as the paper itself notes for Base and Clean-Finetuned models. The near-zero distances for Noise-Finetuned models are claimed to show targeted correction, but no control—such as distance to another word in the same sentence, or a paired per-token cosine distance—rules out the null hypothesis of generic context integration. This is load-bearing because it is the only direct evidence for 'moving its representation toward the grammatical form.' It also contaminates Section 5.4.2/Figure 4: when the noisy and clean representations converge, ranking heads by distance to the clean form (Robustness Heads) and by distance to the original noisy representation (Influential Heads) become the same ranking, so the high overlap for Noise-Finetuned models is largely a mathematical consequence of the Section 5.3 convergence, not independent evidence of a dedicated mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes how NMT encoder representations respond to synthetic grammatical errors. It introduces three error types (article, preposition, noun-number) plus MORPHEUS perturbations, and compares Base, Clean-Finetuned, and Noise-Finetuned variants of four translation models across five language directions. The authors use GED probing to track error detection through layers, 1-CKA distance between noisy and clean word representations to track correction, and a head-masking method to define Robustness Heads and Influential Heads. The central claims are that NMT encoders implement a detect-then-correct GEC process and that robustness fine-tuning increases reliance on Robustness Heads in deeper layers.","tokens_in":23320,"tokens_out":5165,"duration_ms":58015,"significance":"If the central interpretation holds, the paper provides a model-agnostic, layer-resolved account of how NMT encoders handle ungrammatical input, and it identifies an attention-level mechanism that fine-tuning amplifies. The study is broad in coverage: four model families, five language directions, three targeted error types plus one black-box attack, and it ships code. The descriptive findings—GED probing curves, CKA distance trends, and attention-to-POS patterns—are internally consistent and potentially useful for interpreting robustness in NMT. The main risk is that the evidence for the 'correction' component and for the Robustness-Head overlap is confounded by shared context between the noisy and clean sentence pairs, as detailed below.","major_comments":[{"comment":"The CKA distance between the noisy word and its clean form decreases with depth even for Base and Clean-Finetuned models, and the paper itself explains this by noting that the two words share the same context and that representations move closer as the encoder integrates context. Therefore the near-zero distances observed for Noise-Finetuned models do not by themselves establish targeted correction; a control such as the CKA distance between the noisy word and a different word in the same sentence, or paired per-token distances for unchanged words, is needed to rule out generic context integration. This is load-bearing for RQ1 and for the interpretation of RQ3, and the paper should add such a control or substantially weaken the correction claim.","section":"§5.3, Figure 2"},{"comment":"The Robustness Head distance 1-CKA(gw_hi, w) and the Influential Head distance 1-CKA(w_hi, w) coincide whenever the un-masked noisy representation is close to the clean representation, which is exactly the regime observed for Noise-Finetuned models in Figure 2. Thus the high Robustness/Influential overlap reported in Figure 4 may be a mathematical consequence of the convergence measured in §5.3 rather than independent evidence that fine-tuning recruits special heads. The authors should compare the observed overlap against a null model that preserves the marginal head distances, for example random head subsets matched on the distribution of 1-CKA values.","section":"§3.3.4, §5.4.2, Figure 4"},{"comment":"The procedure for selecting the sets of Robustness Heads and Influential Heads is not specified. Figure 4 reports an 'accuracy' between the two sets, but the manuscript does not state whether heads are selected by a threshold, a top-k rule, a statistical test, or some other criterion, nor does it report sensitivity to that choice. Without this, the result is not reproducible and the magnitude of the overlap cannot be interpreted. Please specify the selection rule and report how Figure 4 changes under reasonable alternatives.","section":"§3.3.4, §5.4.2"},{"comment":"The claim that NMT encoders 'inherently implement' a GEC setup is stronger than the evidence supports. In §5.2, GED probing accuracy decreases in deeper layers only for Noise-Finetuned models, while Base and Clean-Finetuned models plateau; in §5.3, the distance decrease for Base and Clean-Finetuned models is attributed to shared context rather than correction. The Discussion already hedges by saying the behavior is 'more distinguishable in Noise-Finetuned models.' I recommend moving the RQ1 claim to that more modest level, or providing direct evidence for inherent detection-then-correction in Base models.","section":"Abstract, §1 RQ1, §6 Discussion"}],"minor_comments":[{"comment":"The heading 'Respresentation Similarity' contains a typo and should read 'Representation Similarity'.","section":"Appendix C.3"},{"comment":"The y-axis label and caption say 'Accuracy' but the figure reports agreement between two sets of heads; please define this quantity in the text and state whether it is head-level accuracy, overlap proportion, or another measure.","section":"Figure 4, §5.4.2"},{"comment":"The notation fW for the ungrammatical word representations is confusing because f usually denotes a function; consider using X_noisy or W_noisy for consistency with W.","section":"§3.3.2"},{"comment":"Durrani et al. 2023a and 2023b appear to be the same paper with identical titles and bibliographic details; please verify and correct the duplicate reference.","section":"References"},{"comment":"The sentence beginning 'Figure shows the accuracy...' appears to be a leftover placeholder; it should refer to a specific figure number.","section":"Appendix C.4"}],"recommendation":"major_revision","confidential_remarks":"I agree with the conditional assessment: the descriptive analysis is valuable and the paper is within scope, but the central interpretive claims require the proposed control experiments and a more cautious formulation of what is 'inherent' versus what emerges after robustness fine-tuning."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a serious empirical study: four models, five language directions, three hand-crafted error types plus Morpheus, with Base/Clean-Finetuned/Noise-Finetuned comparisons and an encoder-only fine-tuning ablation. That breadth is real and the code is available. Second, the headline claim—that NMT encoders 'inherently' detect then correct grammatical errors—rests on weaker evidence than the paper suggests. The main correction evidence, Section 5.3, is a decreasing CKA distance between noisy and clean word representations. Since the noisy and clean sentences differ only in one word, that distance can fall simply because the encoder pools nearly identical context. The authors acknowledge this explanation for Base and Clean-Finetuned models, then attribute the near-zero distances for Noise-Finetuned models to robustness without a control against other words in the same context or a paired per-token distance. So the detect-then-correct narrative is plausible but not nailed down.\n\nWhat the paper does well: the GED probing results are internally consistent and interesting—Noise-Finetuned models keep high detection accuracy in early layers, then probing accuracy drops in deep layers, which is what you'd expect if correction happens there. The POS attention analysis is a nice descriptive contribution, and the cross-linguistic differences (English vs French adjective ordering) are well reasoned. The encoder-only fine-tuning finding is practically useful, and the Morpheus generalization is a good sanity check.\n\nThe soft spots are real but fixable. Figure 4's Robustness vs Influential Heads overlap is confounded: by the paper's own equations, when the noisy representation converges to the clean one, the Robustness Head distance 1-CKA(gw_hi, w) and the Influential Head distance 1-CKA(gw_hi, gw) become the same ranking. So the high overlap for Noise-Finetuned models is largely a mathematical consequence of the Section 5.3 convergence, not independent evidence of a dedicated mechanism. The head selection rule (threshold or top-k) is never specified, and there are no error bars or significance tests anywhere. The abstract's 'inherently' overclaims.\n\nWho it's for: people working on MT robustness and interpretability. It's a good starting point for thinking about how to audit robustness, but the central causal claims need tightening. I would send it to peer review—the empirical base is valuable and the fixes are tractable—but I'd ask for a shared-context control in Section 5.3, a de-confounded head-overlap analysis, and softer wording.","headline":"Serious empirical breadth, but the detect-then-correct narrative and Robustness Heads claim rest on a confound that needs controls before the headline strength is justified.","tokens_in":24010,"tokens_out":3238,"would_cite":true,"duration_ms":33367,"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 NMT encoders first detect a grammatical error, then move that word's representation toward its correct form, and that fine-tuning for robustness mainly amplifies this existing behavior.","keywords":["neural machine translation robustness","grammatical error correction","GED probing","CKA representation similarity","attention heads","synthetic noise fine-tuning","multilingual NMT"],"falsifier":"A control experiment that would refute the targeted-correction claim: compute $1-\\text{CKA}$ between the noisy word's representation and (a) its clean form and (b) a third word of the same POS category placed in the identical sentence context, across encoder layers. If both distances shrink at the same rate, the representation change is generic context pooling rather than error correction; the paper's claim predicts a significantly larger drop toward the clean form, especially in deeper layers of noise-finetuned models.","tokens_in":22891,"feed_emoji":"🧠","tokens_out":4562,"duration_ms":44673,"temperature":0.7,"pith_summary":"This paper proposes that neural machine translation (NMT) encoders already implement an implicit grammatical-error-correction loop: as layers deepen, the encoder first marks an ungrammatical word as detectable by probes, then adjusts that word's representation so that it moves toward the representation of the correct word. The authors test this on four multilingual NMT models across five language directions using synthetic article, preposition, and noun-number errors plus a black-box inflectional attack. They find that GED probing accuracy rises in the first half of the encoder and then plateaus or drops, while a CKA-based distance between the noisy word and its clean counterpart shrinks through the layers, especially after fine-tuning on noisy data. They also isolate attention heads, called Robustness Heads, whose masking moves the noisy representation toward its clean form, and show these heads attend to linguistically relevant part-of-speech categories. A sympathetic reader would care because it suggests that robustness to grammatical noise is not an extra skill but a re-weighting of a structure that translation models already possess.","feed_headline":"Translation models correct grammar inside the encoder, layer by layer","feed_subtitle":"Probing shows encoders spot a grammar slip, then pull its representation toward the correct form","key_machinery":"The load-bearing objects are: (i) GED probes, single-layer linear classifiers trained on each encoder layer's word representations to detect the ungrammatical word; (ii) CKA (Centered Kernel Alignment) representational similarity, used as $1-\\text{CKA}(fW, W)$ between noisy-word representations and their clean-form counterparts to measure correction; and (iii) Influential Heads and Robustness Heads, constructed by masking one attention head at a time and measuring how much the word representation moves, where Robustness Heads are the heads whose masking pulls the noisy representation toward its grammatical form. The framework's work is to convert the unobservable 'correction' into two measurable quantities, detection accuracy per layer and representation displacement per layer, and to tie the displacement to specific attention heads.","core_discovery":"The central claim is that an NMT encoder operates as a detect-then-correct pipeline for grammatical errors. For each ungrammatical input, the encoder learns a representation of the erroneous word that is progressively steered toward the representation the same word would have in a clean sentence; the paper supports this with two measurements. Grammatical error detection probing shows that the erroneous word's identity is readily linear-decoded in lower-to-middle layers, with F1 rising early and then falling in deeper layers, which the authors read as the detection stage being superseded by correction. The representational distance $1-\\text{CKA}$ between the noisy word representation and the clean-form representation decreases layer by layer for base models and collapses toward zero for noise-finetuned models, which they interpret as active correction. Fine-tuning for robustness on noisy data does not invent new behavior; it amplifies existing behavior, since noise-finetuned models rely more heavily on Robustness Heads in deeper layers, where the correction is hypothesized to occur.","pith_inferences":["If the detect-then-correct account is right, the same mechanism may appear in decoder-only language models, where early layers might localize the error and later layers blend it away; the paper itself flags this as future work.","The Robustness Head identification method could be turned into an audit tool: given any new model, mask heads to locate which linguistic regularities it silently repairs, without needing adversarial fine-tuning.","Because shared context would also shrink CKA distance, a stronger test would compare the noisy representation's distance to its clean form against its distance to a randomly chosen word in the same sentence; the correction claim predicts a specific advantage, not just general convergence.","The language asymmetry suggests that multi-language robustness fine-tuning may suffer interference, and language-specific adapters could isolate the correction heads per language rather than letting languages compete."],"forward_implications":["Fine-tuning only the encoder on synthetic grammatical errors is nearly as effective as full-model fine-tuning for closing the clean-versus-noisy COMET gap, while preserving performance on clean data.","Base, clean-finetuned, and noise-finetuned models share the same Robustness Heads and similar attention patterns over POS tags; robustness fine-tuning mainly increases reliance on these heads in deeper layers.","The detect-then-correct pattern holds across four model families and across article, preposition, noun-number, and MORPHEUS inflectional errors, so it is not an artifact of a single noise type.","Error handling differs across source languages; for instance, noun-number detection peaks at a higher F1 for French, where articles and adjectives carry number, than for English, suggesting robustness depends on the source language's linguistic signaling.","Since fine-tuning on grammatical errors can even slightly improve clean-data translation, the authors conclude that such fine-tuning acts as a regularizer and can be a practical, interpretable robustness strategy."],"supporting_citations":[{"why":"Supplies the Transformer architecture whose attention heads and layers are the objects of analysis.","marker":"(Vaswani et al., 2017)"},{"why":"Provides the masking-based influence method that is redefined to identify Robustness Heads.","marker":"(Voita et al., 2019)"},{"why":"Supplies CKA, the representational similarity measure used to quantify distance between noisy and clean word representations.","marker":"(Kornblith et al., 2019)"},{"why":"Provides the GED probing methodology for detecting ungrammatical words from encoder representations.","marker":"(Davis et al., 2022)"},{"why":"Provides the implementation for introducing article, preposition, and noun-number replacement errors.","marker":"(Anastasopoulos et al., 2019)"},{"why":"Supplies MORPHEUS, the black-box inflectional attack used to test generalizability to other error types and multiple errors per sentence.","marker":"(Tan et al., 2020)"},{"why":"Provides the attention aggregation scheme used to analyze Robustness Heads' attention to POS tags.","marker":"(Clark et al., 2019)"}],"fun_headline_variants":["NMT encoders detect then correct grammar errors internally","Encoder layers first spot grammar slips, then fix them","Robustness heads drive grammar correction inside NMT encoders","Probing shows encoders move bad grammar toward correct form","Grammar errors are corrected inside the encoder, not just output"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The detect-then-correct reading assumes that the shrinking CKA distance between a noisy word and its clean-form twin measures a targeted repair, rather than just both representations converging because the surrounding sentence is nearly identical; the paper itself flags this shared-context alternative and runs no control against another word in the same context.","fun_headline_variants_meta":{"raw":{"variants":["NMT encoders detect then correct grammar errors internally","Encoder layers first spot grammar slips, then fix them","Robustness heads drive grammar correction inside NMT encoders","Probing shows encoders move bad grammar toward correct form","Grammar errors are corrected inside the encoder, not just output"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000531,"raw_usage":{"total_tokens":2529,"prompt_tokens":889,"completion_tokens":1640,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":1560}},"tokens_in":505,"tokens_out":1640,"duration_ms":12673,"temperature":1.0,"reasoning_tokens":1560,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:33:16.681422+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A control experiment that would refute the targeted-correction claim: compute $1-\\text{CKA}$ between the noisy word's representation and (a) its clean form and (b) a third word of the same POS category placed in the identical sentence context, across encoder layers. If both distances shrink at the same rate, the representation change is generic context pooling rather than error correction; the paper's claim predicts a significantly larger drop toward the clean form, especially in deeper layers of noise-finetuned models.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the GED probing methodology for detecting ungrammatical words from encoder representations."}],"review_version":1}