{"id":"bec0ab19-50d7-420d-8612-2e4b30caa7a4","arxiv_id":"2509.02075","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Instruction-tuned Llama 3.1 controls word count far better than the base model, and attribution scores point to later layers, but the scoring rule mishandles outputs that are too short.","lead":"This paper compares a base Llama 3.1 model with its instruction-tuned version on writing sentences with an exact number of words, in English and Italian, and inspects which internal parts push the model toward correct length. It finds the tuned model is much better and that later layers matter most, but the main scoring rule mistakes too-short answers for correct ones, so the mechanistic conclusion is not yet supported.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CWA's sign rule (Eq. 5) scores premature EOS as +1, contradicting the paper's own too-short failure definition; since ENG-IT errors are mostly -1, positive deep-layer CWA cannot support the claimed length-control specialization.","rationale":"The reader's weakest assumption is exactly the load-bearing flaw. Eq. (5) is internally inconsistent with the paper's own failure definition, and the error statistics make the consequence concrete: ENG-IT fails mostly by being one word short, so those failures are scored positive. Because CWA is the sole basis for the component-level attribution, the central mechanistic claim cannot be distinguished from a trivial artifact of generation length. The performance result (IT improves length control) is credible, but the mechanistic narrative — late-layer attention for English, final-layer MLP compensation for Italian — is not supported. A corrected-sign recomputation or a truncation control would settle the matter; until then the reader's REJECT is appropriate. I therefore keep the verdict unchanged.","tokens_in":11631,"tokens_out":6468,"duration_ms":68635,"concrete_test":"Control: truncate ENG-BASE generations at N words (with EOS) and recompute attention/MLP CWA using Eq. 5; if the deep-layer BASE profile becomes positive under mere truncation, the IT-vs-BASE contrast is an artifact of length distribution. Direct check: recompute CWA with a sign function matching Section 4's taxonomy — +1 for word steps while m<N, +1 for EOS only when m=N, -1 for premature EOS (m<N) and for any token after m>N — and regenerate Figure 4. If the ENG-IT later-layer positive pattern disappears or reverses, Eq. 5 is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 defines too-short failure and states the premature EOS is a failure step, but Eq. (5) assigns sign=+1 whenever the running word count m ≤ N. At a premature EOS, m<N, so the step is scored +1. In the paper's N=3 example 'The dog . <eos>', all four steps receive +1, making a too-short sequence indistinguishable from a successful one under CWA. Table 1 shows ENG-IT errors are mostly non-positive (mean -0.07, min -1, 75th pct 0); the IT model's dominant failure mode is too-short. CWA therefore has almost no negative signal for the errors the IT model actually makes, and its positive signal simply reflects that IT outputs do not exceed N. The BASE model overgenerates (mean +34.8), accumulating many -1 steps once m>N. The IT-vs-BASE deep-layer contrast in Figure 4 is thus confounded with output length distribution: components are scored positively whenever the model stays within N, regardless of whether they implement length control. The paper's central mechanistic claim (later-layer attention/MLP specialization) rests entirely on CWA and is unsupported; the limitations section does not acknowledge this sign inconsistency.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies how instruction tuning changes the internal mechanisms that let Llama 3.1 8B follow explicit word-count constraints. It compares the base and instruction-tuned (IT) models on English and Italian prompts with target lengths N = 3–9, scoring outputs as success, too short, or too long. The authors introduce Cumulative Weighted Attribution (CWA), derived from Direct Logit Attribution (DLA), which aggregates per-step DLA scores with signs determined by whether the running word count is within the target. The main claims are that instruction tuning substantially improves length control, that this improvement is implemented mainly in deeper model layers, that later-layer attention heads are especially important in English, and that final-layer MLPs play a compensatory role in Italian. Performance results show the IT model is far more accurate than the base model, which overgenerates. The mechanistic claims rest on CWA patterns in Figures 4 and 5.","tokens_in":11907,"tokens_out":6789,"duration_ms":83596,"significance":"If valid, the paper would be a useful step toward understanding where length-control behavior lives in an open-weight model and how it differs across languages. The task is important, the comparison between base and instruct versions is natural, and the cross-lingual design is a genuine addition over prior work. The performance findings are plausible and supported by Table 1. However, the central mechanistic contribution is undermined by a fundamental problem in the CWA sign rule, which contradicts the paper's own definition of too-short failures and makes the main figures uninterpretable as evidence about length-control specialization. The paper also leaves a key methodological detail unspecified: how a vector-valued DLA is reduced to the scalar used in CWA. For these reasons, the mechanistic claims are not currently supported.","major_comments":[{"comment":"The sign function in Eq. (5) is inconsistent with the failure definitions given earlier in §4. The text defines a premature EOS as a failure step when fewer than N words have been generated, but Eq. (5) assigns sign = +1 whenever the running word count m ≤ N. In the paper's own too-short example, 'The dog . <eos>' for N=3, all four steps satisfy m ≤ 3, so every step is scored +1 and the sequence is indistinguishable from a successful one. Since Table 1 shows that ENG-IT and ITA-IT errors are mostly non-positive (mean −0.07 and −0.08, with ITA-IT 25th percentile −1), the CWA contains almost no negative signal for the dominant IT error mode. The positive deep-layer CWA in Figures 4–5 therefore reflects only that IT outputs stay within the requested bound, and the IT-versus-BASE contrast is confounded with output length distribution. This undermines the central mechanistic claim that later-","section":"§4, Eq. (5)"},{"comment":"CWA is not fully defined. In Eqs. (2) and (3), DLA_Cl is a vector over the vocabulary: it is the component output projected by the unembedding matrix W_U. Eq. (4), however, appears to treat DLA_Cl as a scalar when computing CWA_Cl. The paper does not state whether the scalar is the logit of the sampled token, the logit of the gold token, an average over tokens, or some other reduction. Without this specification, the numerical CWA values reported in Figures 4 and 5 are not reproducible. In addition, the sign function uses 'the sentence generated from tokens (t1,...,ti) collectively constituting m words,' but the paper does not define how word counts are computed from token sequences—e.g., how punctuation, subword tokens, or the EOS token are handled. These details are load-bearing for the proposed metric and must be clarified before the results can be accepted.","section":"§4, Eqs. (2)–(4)"}],"minor_comments":[{"comment":"The figures show layer-level CWA values without error bars or confidence intervals, even though §5 states each configuration was repeated 20 times. It is also unclear whether each curve is an average over attention heads within a layer, over prompt templates, over repetitions, or over all of these. Adding this information would make the visual language differences (e.g., ITA-IT final-layer MLP) more interpretable.","section":"§6, Figures 4–5"},{"comment":"The text says prompts were built with integers from 0 to 9, but the reported experiments cover N ∈ [3, 9]. Please clarify what was done with N = 0, 1, 2, or state that they were excluded from analysis.","section":"§5, prompt construction"},{"comment":"The limitations paragraph mentions single-digit N, the Llama 3.1 architecture, and the two languages, but it does not mention the metric's sign inconsistency or the confound between CWA and output length. Adding this would improve the paper's transparency.","section":"§7, Limitations"}],"recommendation":"reject","confidential_remarks":"The stress-test concern lands directly: Eq. (5) misclassifies premature EOS as a positive step, and because the IT model's errors are predominantly too-short, the CWA evidence for deep-layer specialization is invalid. This is not a presentation issue; the main mechanistic claim is unsupported. A future revision that corrects the sign rule, redefines the scalarization, and reruns the analysis could be considered anew, but the current manuscript does not meet the bar for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper asks a question worth asking: where does instruction tuning put word-count control in Llama 3.1 8B, and does the answer differ for English and Italian? The performance half is decent and the cross-lingual angle is a genuine novelty. The mechanistic half, however, rests on Cumulative Weighted Attribution, and CWA has a sign inconsistency that guts the main claim.\n\nWhat is new and good: the first DLA-style attribution comparison for word-count control across two languages, as far as I can tell from the cited literature. CWA is a simple aggregation of direct logit attribution, but that doesn't make it useless. The performance results are credible: 20 repeats per configuration, and Table 1 gives means and standard deviations. The paper also has an honest limitations section on scope (one model, one task, two related languages), though it misses the sign problem.\n\nNow the soft spot, and it is load-bearing. Eq. (5) defines sign(t_i) = +1 whenever the running word count m is <= N, and -1 otherwise. The paper's own failure definition says a premature EOS (m < N) is a failure step. Put the two together: a too-short sequence like \"The dog . <eos>\" gets +1 on every step, including the EOS. So CWA cannot distinguish a too-short failure from a success. And Table 1 shows that the IT model's errors are mostly -1 word (ENG-IT mean -0.07, 75th percentile 0). That means CWA has essentially no negative signal for the dominant error mode of the IT model. The positive deep-layer CWA for IT models is simply reflecting that IT outputs stay within N, while the BASE model overgenerates and accumulates -1 steps. The IT-vs-BASE contrast in Figure 4 is therefore confounded with output length distribution. The central claim—that later-layer attention heads (English) and final-layer MLPs (Italian) specialize for length control—is not supported.\n\nAlso worth noting: no code or data released, no causal validation despite the limitations section mentioning that as future work, and the CWA figures have no error bars. The cross-lingual differences could be real, but this analysis doesn't show it.\n\nWho gets value from this paper? People working on length-control benchmarks might cite the performance numbers. The interpretability contribution should not be relied on until the sign rule is fixed and the analysis redone. That said, this is not a desk-reject-quality paper. The question is important and the flaw is technical and fixable. I would send it to a serious referee, with the expectation of major revision—a corrected sign function, a re-analysis, and ideally code release.\n\nRecommendation: send to peer review, but reject in current form.","headline":"The sign function in Eq. 5 is a load-bearing flaw: it scores premature EOS as +1, so CWA cannot see the IT model's dominant failure mode; the performance results are credible, but the mechanistic claims rest on that broken metric.","tokens_in":12393,"tokens_out":1844,"would_cite":false,"duration_ms":22078,"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":"Instruction tuning gives Llama 3.1 8B near-zero word-count error by rewiring deep-layer attention in English, with final-layer MLPs compensating in Italian.","keywords":["word-count constraint","instruction tuning","mechanistic interpretability","direct logit attribution","cumulative weighted attribution","cross-lingual","Llama 3.1 8B","attention heads"],"falsifier":"Re-run the CWA computation with the too-short stop (premature end-of-sentence) scored as a failure step rather than a success; if the late-layer attention attribution for instruction-tuned models no longer exceeds the base model's, the claimed deep-layer specialization is an artifact of the sign rule.","tokens_in":11514,"feed_emoji":"🎯","tokens_out":7779,"duration_ms":74110,"temperature":0.7,"pith_summary":"This paper asks where a language model's ability to hit an exact word count lives inside the network, and how instruction tuning creates that ability. Using Llama 3.1 8B in English and Italian, it shows that the instruction-tuned version nearly nails requested lengths (average error near zero) while the foundation model over-generates by tens of words. To see why, the authors introduce Cumulative Weighted Attribution (CWA), which scores each internal component by whether its logit push keeps the running word count within target. The main result: instruction tuning reconfigures the deeper layers (roughly layer 24 onward) for the task—later attention heads become strongly positive in English, while Italian relies on the final-layer MLP, a pattern the authors read as a language-dependent compensatory division of labor. If correct, this locates length control in specific, modifiable components rather than diffuse model-wide behavior, and it cautions that the mechanism is not language-neutral.","feed_headline":"Instruction tuning rewires deep layers to enforce word counts","feed_subtitle":"Cross-lingual trace: late attention heads carry English, final MLPs compensate in Italian.","key_machinery":"Cumulative Weighted Attribution (CWA), a signed aggregate of Direct Logit Attribution (DLA) over all generation steps of an output. At each step, DLA projects a component's residual-stream output onto the unembedding to measure its push on the winning token; CWA averages these pushes after signing each step +1 if the running word count is at most the target N and −1 otherwise. The sign rule is what turns raw logit attribution into a 'helps or hinders length control' score, and the per-layer aggregation is what allows the paper to localize the effect in deeper layers.","core_discovery":"Comparing the base and instruction-tuned versions of Llama 3.1 8B on 'write a sentence with exactly N words' prompts, the paper finds that instruction tuning reduces mean absolute error from tens of words to near zero. Using CWA, the paper traces this to the deep layers: components at layers 24–31 that show negative (constraint-hindering) attribution in the base model flip to positive attribution after tuning, especially attention heads in English. In Italian, attention contributions stay weaker and the final-layer MLP instead shows a strong positive CWA, which the authors propose is a compensatory mechanism for less effective attention-based length tracking. The paper also reports that the","pith_inferences":["The sign rule in Eq. 5 labels a premature end-of-sentence as a success step (running count still below N), yet the paper's own failure definition counts too-short outputs as failures; with English tuned errors mostly −1, the positive deep-layer attention CWA may substantially reflect the model's ordinary tendency to generate short texts, not length adherence.","A direct test of the mechanism: score the same outputs with a sign rule that marks premature EOS as failure, and with a rule that marks all non-EOS word tokens after the count equals N as failure; the claimed layer-24+ specialization should survive only if it tracks true constraint-following.","The compensatory-MLP reading for Italian could be checked by putting Italian lexical content into English prompts (or vice versa): if the final-layer MLP remains the strong component, the effect is tied to the language of content, not the language of instruction.","The paper's attribution is correlational; a causal probe that zeroes or amplifies the late-layer heads (already flagged as future work) would convert the localization claim into a verified circuit claim."],"forward_implications":["The base model's failure is mostly a stopping problem (it runs ~35 words long in English, ~27 in Italian), not a counting problem, since the tuned model with the same architecture is near-zero-error.","If CWA is correct, length control in Llama 3.1 8B sits in late layers (roughly 24–31), so targeted interventions such as activation steering, pruning, or fine-tuning of those heads should have outsized effects on the capability.","The English/Italian split—late attention heads vs final MLP—implies instruction-following strategies are language-conditional, so cross-lingual transfer of control improvements is not automatic.","The deep-layer specializations appear only after instruction tuning, suggesting tuning adds a task-adherence circuit rather than strengthening a pre-existing one.","Because the base model's deep MLPs are strongly negative for the constraint, instruction tuning must override rather than reuse those computations."],"supporting_citations":[{"why":"Supplies the Llama 3.1 8B base and instruction-tuned checkpoints whose behavior and internal components are analyzed.","marker":"(Llama Team, AI Meta, 2024)"},{"why":"Establishes the residual-stream additivity that makes Direct Logit Attribution (and hence CWA) a valid decomposition.","marker":"(Elhage et al., 2021)"},{"why":"Supplies the cross-lingual DLA protocol this paper adapts to length control, including the English/Spanish comparison approach.","marker":"(Ferrando and Costa-Jussà, 2024)"},{"why":"Prior benchmark documenting that LLMs perform poorly on controlled length tasks and tend to generate short outputs; the paper's performance results are checked against it.","marker":"(Sun et al., 2023)"},{"why":"Shows instruction tuning strengthens attention connections to instruction verbs, the prior result the authors invoke to explain the English attention-head pattern.","marker":"(Wu et al., 2024)"},{"why":"Supports the claim that instruction tuning teaches superficial instruction-following patterns, which the paper links to its observed deep-layer specialization.","marker":"(Kung and Peng, 2023)"}],"fun_headline_variants":["Deep-layer rewiring explains length control after tuning","Italian length control leans on MLPs, not attention","Instruction tuning flips deep-layer sign for word counts","Late layers specialize for exact word counts after tuning","Cross-lingual: English uses attention, Italian uses MLPs"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The central measurement assumption is that a generation step counts as correct whenever the running word count is at or below the target, so a too-short finish is scored as a success and the positive deep-layer signal may just be the model's habit of generating short text.","fun_headline_variants_meta":{"raw":{"variants":["Deep-layer rewiring explains length control after tuning","Italian length control leans on MLPs, not attention","Instruction tuning flips deep-layer sign for word counts","Late layers specialize for exact word counts after tuning","Cross-lingual: English uses attention, Italian uses MLPs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000518,"raw_usage":{"total_tokens":2307,"prompt_tokens":664,"completion_tokens":1643,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":408,"completion_tokens_details":{"reasoning_tokens":1566}},"tokens_in":408,"tokens_out":1643,"duration_ms":10396,"temperature":1.0,"reasoning_tokens":1566,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:53:34.848607+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the CWA computation with the too-short stop (premature end-of-sentence) scored as a failure step rather than a success; if the late-layer attention attribution for instruction-tuned models no longer exceeds the base model's, the claimed deep-layer specialization is an artifact of the sign rule.","supporting_citations":[],"review_version":1}