acceptodds
Under review as a conference paper at ICLR 2027

Naturalness Is in the Heads: Localizing and Enhancing Translation Naturalness in LLMs

Abstract

Large language models can generally follow translation instructions; the central challenge lies in producing translations that read naturally in the target language. To inform quality improvement, mechanistic analysis should explain how models choose among correct translations, not only how they follow translation instructions. Naturalness unfolds over complete sequences and admits multiple valid wordings, so neither a single-token readout nor a single reference adequately captures the behavior. We find an exceptionally sparse causal circuit: intervening on only the top-ranked 1% of attention heads shifts translation naturalness, while steering the localized circuit enhances it. We localize these heads using the equivalence-class score difference (Equi-DIFF), which contrasts cardinality-normalized scores over classes of valid multi-token translations and reduces exactly to logit difference in the single-token singleton case. Comparisons against word-level translation and sequence-singleton baselines show that complete-sequence scoring materially changes head rankings, while multi-completion classes improve calibration and reproducibility. Held-out knockout reduces natural-style preference before broad quality degradation, whereas activation steering enhances judged naturalness (69–87% model-level decisive win rate). Across three models, 12 translation directions, and two corpora, this effect recurs in most settings, providing consistent causal evidence for naturalness heads despite varying adequacy trade-offs. Code will be made available after publication.

Primary area:
interpretability and explainable AI
Keywords:
Mechanistic Interpretability,Language Models,Interpretability,Machine Translation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Y4siOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

Paper studies study on whether translation naturalness can be attributed to specific attention heads in LLMs. The authors introduced Equi-DIFF, a sequence-level metric which compares sets of natural and literal translations instead of individual tokens or reference sequences. They identify heads associated with naturalness and validate their role through knockout and activation steering using path matching. They have experimented across 3 models, 12 translation directions, and 2 datasets which has shown that ablating these heads reduces natural-style preference, and that steering generally improves judged naturalness.

Critical strengths

The causal validation goes beyond differences in head rankings. I found the cross-ranking knockout experiment in Table 1 pretty useful: word-ranked heads affect the word-translation endpoint more, while naturalness-ranked heads affect the sequence-level naturalness endpoint more. The interaction persists under both zero and mean ablation and is supported by layer/type-matched controls. This is a meaningful evidence for different functional preferences, rather than simply observing that two metrics select different heads. The authors have appropriately distinguished this from claiming completely disjoint circuits.

2. Equi-DIFF addresses a relevant measurement problem. The motivation for scoring multiple complete translations is also convincing: neither a single token nor one particular wording adequately represents the intended behavior. The sequence-level, class-based formulation provides a useful way to study this choice, and the completion-choice analysis demonstrates that the selected wording can materially affect localization. This makes the readout a substantive part of the contribution.

Critical weaknesses

1. The interpretation of the natural/literal completion classes needs stronger validation. Appendix B constructs these classes using the model’s own generations, mechanical output filters, likelihood-based selection, and a prompt-conditioned score-separation criterion. These checks establish that the model distinguishes the two conditions, but do not by themselves establish that both classes preserve the source meaning and differ mainly in naturalness. I could not find an independent adequacy and style check of the discovery classes in the described protocol. Hence, the localized contrast may include semantic differences alongside stylistic ones. The held-out interventions remain informative, but the more specific interpretation of identifying components that choose between equally faithful translation strategies, is less firmly established.

2. The steering evaluation leaves the naturalness–adequacy trade-off difficult to assess. In Table 3, every reported Qwen direction, corpus setting favors the baseline among decisive adequacy judgments, despite favoring steering on naturalness. This does not establish a large overall adequacy loss: ties are excluded, MetricX improves, and COMET changes are small. However, it does require an explanation. Appendix E defers the underlying win/loss/tie counts to an evaluation artifact and states that they aggregate three prompt variants, making the aggregate results difficult to interpret from the manuscript alone. There is also a concrete reason to inspect the adequacy judgments: in Table 6, the steered translation of “Permission to relax” adds “watch some TV,” absent from the source as shown, yet receives a positive adequacy change. This example does not establish a systematic failure, but it makes independent validation of the adequacy results important.

Questions for authors

1.Clarification of the localization protocol: Were the natural and literal discovery completions independently checked for source-meaning preservation and the intended stylistic difference? If such checks were performed, please describe the procedure and summarize the results. This would directly affect my interpretation of what Equi-DIFF localizes.

2.Clarification of the existing evaluation: Could you provide the naturalness and adequacy win/loss/tie counts, together with paired adequacy-score changes, separately for neutral, natural, and literal prompts? Please also summarize the professional-translator assessment mentioned in Section 5 and clarify whether it covered adequacy or only naturalness. These results would help distinguish a limited trade-off from a broader problem with semantic preservation or judge reliability.

5 comments

Part 1/4

Weakness 1 & Question 1: Human Validation of Discovery Completions.

Weakness 1:

  1. The interpretation of the natural/literal completion classes needs stronger validation. Appendix B constructs these classes using the model’s own generations, mechanical output filters, likelihood-based selection, and a prompt-conditioned score-separation criterion.

These checks establish that the model distinguishes the two conditions, but do not by themselves establish that both classes preserve the source meaning and differ mainly in naturalness. I could not find an independent adequacy and style check of the discovery classes in the described protocol. Hence, the localized contrast may include semantic differences alongside stylistic ones. The held-out interventions remain informative, but the more specific interpretation of identifying components that choose between equally faithful translation strategies, is less firmly established.

Question 1:

  1. Clarification of the localization protocol: Were the natural and literal discovery completions independently checked for source-meaning preservation and the intended stylistic difference? If such checks were performed, please describe the procedure and summarize the results. This would directly affect my interpretation of what Equi-DIFF localizes.

Response to Weakness 1 & Question 1:

We thank the reviewer for requesting linguistic validation of the discovery classes. Source-based human review supports both meaning preservation and the intended naturalness contrast in the reviewed text across all 15 discovery units. Reference regeneration provides separate evidence of stable localization of the core heads.

Our discovery pipeline establishes a controlled, model-specific contrast through matched natural/literal prompts and behavioral filtering. To make the linguistic evidence concrete, we report a focused analysis of Qwen3.5-4B English-to-Chinese. The source-based human assessment covers all 15 discovery units and examines the dimensions raised by the reviewer: source-meaning preservation, naturalness, and the types of linguistic differences.

Four Chinese-native annotators proficient in English contributed to the assessment. Each unit contains one English source and two five-member reference classes, generated under the natural (YrY_r) and literal (YcY_c) prompts. Reviewers recorded meaning relation and source faithfulness, checking for omissions, additions, and mistranslations. Naturalness judgments allowed ties, mixed preferences, and uncertainty.

The results are summarized below.

Human-review resultShare of the 15 source units
Both classes judged to preserve source meaning100% (15/15)
Naturalness preference for YrY_r over YcY_c100% (15/15)
Wording differences only26.7% (4/15)
Syntax, word order, or register differences only26.7% (4/15)
Both wording and structural/stylistic differences46.7% (7/15)

The final three rows are mutually exclusive difference types. 73.3% (11/15) of these units involve syntax, word order, or register differences, identifying linguistic choices our multi-token measure is designed to capture. These text-based judgments characterize meaning and expression, complementing the behavioral gate’s check of model responsiveness.

As a separate robustness check, we regenerated the model’s references and repeated localization. Fourteen of the original 15 sources passed the same behavioral criteria; old and new references were compared on this probability-selected subset. Top-15 heads denotes the 15 highest-ranked heads.

Reference-regeneration checkResult
Heads shared by the two Top-15 sets (old vs. regenerated references)12/15
Ranks of the three stable core heads, L18.H11 / L31.H14 / L23.H5, with regenerated references1 / 2 / 6

Ranks are positions in the regenerated head ranking; L and H denote layer and head/group indices.

Human review provides direct linguistic support for meaning-preserving naturalness differences in the reviewed text across all 15 discovery units, while reference regeneration recovers the stable core heads and provides complementary localization evidence.

Part 2/4

Weakness 2(a) & Question 2(a): Naturalness–Adequacy Trade-offs across Prompts.

Weakness 2(a):

  1. The steering evaluation leaves the naturalness–adequacy trade-off difficult to assess.

2 (a). In Table 3, every reported Qwen direction, corpus setting favors the baseline among decisive adequacy judgments, despite favoring steering on naturalness. This does not establish a large overall adequacy loss: ties are excluded, MetricX improves, and COMET changes are small. However, it does require an explanation.

Appendix E defers the underlying win/loss/tie counts to an evaluation artifact and states that they aggregate three prompt variants, making the aggregate results difficult to interpret from the manuscript alone.

Question 2(a):

2.Clarification of the existing evaluation:

2 (a). Could you provide the naturalness and adequacy win/loss/tie counts, together with paired adequacy-score changes, separately for neutral, natural, and literal prompts?

Response to Weakness 2(a) & Question 2(a):

We thank the reviewer for asking us to clarify the results by prompt. Steering improves naturalness across all three models, while its effect on adequacy varies by model and prompt. Below, we report the results by prompt and interpret the adequacy judgments alongside the naturalness gains.

Results by prompt

We use the Qwen family as an example, with Top-15 heads, α=1: the 15 highest-ranked heads, at steering strength α=1. The 144 settings cover 2 models × 2 corpora × 12 directions × 3 prompts. Complete results for Qwen and Gemma are in Appendix E.6, Tables 31–36.

The table shows selected language directions for Qwen3.5-4B on WMT24++, judged by Gemini-3.5-Flash, with 379 pairs per row. Neutral, Natural, and Literal use no style modifier, “naturally,” and “literally.” Nat. and Adeq. denote naturalness and adequacy. W/L/T means steering wins, baseline wins, and ties. Both dimensions use 1–7 scores; Δ is the mean steered-minus-baseline score difference, with a possible range of −6 to +6. Brackets show 95% confidence intervals.

DirectionPromptNat. W/L/TNat. Δ [95% CI]Adeq. W/L/TAdeq. Δ [95% CI]
En→ZhNeutral210/82/87+0.605 [0.468, 0.748]80/129/170−0.134 [−0.214, −0.048]
Natural205/84/90+0.542 [0.417, 0.661]60/160/159−0.315 [−0.419, −0.208]
Literal221/75/83+0.583 [0.454, 0.709]67/135/177−0.212 [−0.300, −0.132]
En→JaNeutral161/127/91+0.164 [−0.001, 0.328]102/140/137−0.148 [−0.268, −0.034]
Natural167/122/90+0.197 [0.040, 0.357]117/132/130−0.082 [−0.196, 0.032]
Literal191/96/92+0.425 [0.291, 0.570]130/109/140+0.079 [−0.029, 0.185]
De→EnNeutral212/56/111+0.665 [0.543, 0.807]43/144/192−0.282 [−0.383, −0.180]
Natural190/60/129+0.504 [0.386, 0.649]40/165/174−0.394 [−0.486, −0.307]
Literal227/44/108+0.786 [0.649, 0.934]55/95/229−0.084 [−0.182, 0.009]

Averaging equally over the 12 directions, Literal gives the largest naturalness gain and smallest adequacy decrease for each Qwen model–corpus combination.

Part 3/4

Understanding the adequacy results

Steering encourages more natural expression. Adequacy evaluation helps identify when these changes also affect a detail of the source meaning, such as a qualifier or the speaker’s certainty.

For example, Qwen3.5-35B-A3B produced these WMT24++ translations under the Neutral prompt:

Source: Floor repair person is doing a vacuum, so hopefully that means he's finishing up.

Baseline: 地板维修人员正在吸尘,希望这意味着他即将完工。

Steered: 地板维修师傅正在吸尘,看来快收尾了。

The steered version uses more conversational phrasing, such as “维修师傅” and “快收尾了.” It also changes “希望” (“hopefully”) to “看来” (“it seems”), shifting a hope to an inference. This illustrates a specific trade-off accompanying the more natural expression: the average steered-minus-baseline adequacy difference was −1.00 under both sets of scoring instructions discussed below.

Human review also supports the naturalness gains. Four annotators independently assessed the same 200 English-to-Chinese pairs (100 sources), preferring steering for naturalness in 72% of comparisons on average. Their average baseline preference for adequacy was 34.88%, compared with 45.00% for the LLM on the same pairs. All percentages include ties in the denominator.

Scoring instructions also matter. The original instructions protected faithful translations from being penalized for awkward wording. The revised version also explicitly accepted meaning-preserving reformulations. On the same 70 pairs, evaluated in both presentation orders, the same Gemini judge’s preference for steering rose from 11.4% to 21.4%. The average steered-minus-baseline score difference increased by 0.117 points [95% CI: −0.001, 0.242]. This suggests that the original instructions were more conservative in accepting some meaning-preserving reformulations. Both versions retained penalties for omissions, mistranslations, and unsupported additions.

Interpreting COMET and MetricX

Both metrics learn from human quality assessments, but differ in their models, training data, and scoring scales. Our COMET model learns from direct quality ratings; MetricX-24 also learns from annotations of translation errors and their severity [1][2].

Across the four Qwen model–corpus combinations, MetricX scores decrease by 0.025–0.124 points, indicating improvement, while COMET changes range from −0.0016 to +0.0005. These overall-quality predictions complement the separate evaluations, which show the naturalness gains and accompanying adequacy trade-offs more directly.

Together, the results show that steering improves translation naturalness and that prompt choice matters for balancing this gain with adequacy.

[1] COMET-22: Unbabel-IST 2022 Submission for the Metrics Shared Task

[2] MetricX-24: The Google Submission to the WMT 2024 Metrics Shared Task


Weakness 2(b): Source Context in the Translation Example.

2(b). There is also a concrete reason to inspect the adequacy judgments: in Table 6, the steered translation of “Permission to relax” adds “watch some TV,” absent from the source as shown, yet receives a positive adequacy change. This example does not establish a systematic failure, but it makes independent validation of the adequacy results important.

Response to Weakness 2(b):

We thank the reviewer for questioning the adequacy judgment in this example. The full source contains a television emoji (📺), which provides the context for “watch some TV.” The original judge received this complete source; the emoji was absent from the PDF example.

The recorded example is Qwen3.5-35B-A3B translating English into Chinese under the Literal (“literally”) prompt. Baseline and Steered denote outputs before and after intervention. The original Gemini-3.5-Flash adequacy scores use a 1–7 scale, with higher scores indicating better preservation of source meaning:

ItemOriginal textOriginal adequacy score
SourcePermission to relax. 📺—
Baseline放松许可。📺3
Steered放松一下,看个电视。📺6

The original judge explained that “看个电视” (“watch some TV”) “naturally incorporates the TV emoji into the text”. It identified “放松许可” as an awkward noun phrase (“relaxation permit”), whereas “放松一下” conveys the intended invitation to relax. The television reference therefore has a basis in the source context. The revised appendix includes the complete source and translations (Table 21).

The separate four-annotator assessment described in our response to Weakness 2(a) & Question 2(a) covers 200 other English-to-Chinese translation pairs (100 source texts). Across its 800 paired judgments, 43.5% favored steering for naturalness while judging adequacy equal or better.

The complete source explains the television reference, and the separate human results support our interpretation: steering can improve natural expression while preserving the source’s intended meaning.

Reviewer_vzAPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether translation naturalness can be localized to specific components within multilingual LLMs. The authors introduce Equi-DIFF, a metric that compares sets of complete natural and literal translations rather than individual tokens or sentences. They use it to identify attention heads associated with naturalness and evaluate their causal role through ablation and activation steering. Experiments across three models, twelve translation directions, and two datasets show that ablating the identified heads reduces preference for natural-style translations, while targeted steering substantially increases judged naturalness.

Critical strengths

The paper introduces Equi-DIFF which extends the standard logit-difference metric to compare sets of complete translations rather than individual tokens or sentences. The authors combine Equi-DIFF with path patching and activation steering to identify heads associated with translation naturalness and test their causal role. Furthr, the experiments show that a small set of heads can influence translation naturalness across different models and language directions.

Critical weaknesses
  1. The authors use path patching to identify individual heads but do not examine possible interactions between them. It is unclear how these interactions might affect the head rankings.
  2. The authors generate completion classes using natural and literal translation prompts and filter examples based on their scores (Appendix B). This could introduce selection bias, as the selected examples already show a clear distinction between the two styles.
  3. The paper focuses on attention heads and head-like groups (including DeltaNet outputs) but does not examine other components such as MLPs. Their contribution to translation naturalness therefore remains unclear.
Questions for authors

Table 3 shows substantial improvements in naturalness, but adequacy win rates are often below 50%. Could the authors analyze where the adequacy losses come from? Are they mainly due to minor changes in wording, or do the steered translations introduce omissions, additions, or mistranslations? How sensitive are the identified heads to the filtering and completion-selection criteria in App? Would the same heads be identified using independently annotated natural and literal translations?

9 comments

Part 1/5

Weakness 1: Head Interactions and Ranking Stability.

  1. The authors use path patching to identify individual heads but do not examine possible interactions between them. It is unclear how these interactions might affect the head rankings.

Response to Weakness 1:

We thank the reviewer for asking how head interactions affect the rankings. Explicit interaction tests support stable prioritization of the leading heads within the tested sets. Conditional rankings and Shapley analyses closely agree with single-head rankings, alongside subadditive joint effects.

Component-wise patching is an established localization approach [1][2], including in translation research [3]. Our knockout and steering jointly intervene on selected head sets during autoregressive generation while the rest of the network recomputes normally, incorporating interactions under those interventions.

Ranking stability

For Qwen3.5-4B English-to-Chinese, we used 15 discovery and 15 held-out source texts. We froze the 15 highest-ranked heads (Top-15 heads), then path-patched each of the original five highest-ranked heads (Top-5 heads) in turn. With that head already patched, we ranked the remaining 14 by their additional path-patching effect and compared this conditional ranking with their single-head ranking.

We also computed exact Shapley values from all 32 subsets of the original Top-5 heads. A head’s Shapley value averages its additional effect across all possible orders of adding these five heads.

Ranking comparisonDiscoveryHeld-out
Conditional vs. single-head ranking: Spearman ρ0.974–1.0000.996–1.000
Leading-five overlap among the remaining 14 heads5/5 in every background5/5 in every background
Path-patching Shapley vs. single-head ranking within the original Top-5: Spearman ρ0.901.00

Here, ρ measures rank agreement (1 = identical ordering), and ranges span the five background-head interventions. The overlap compares the leading five members of the conditional and single-head rankings over the same remaining 14 heads.

Joint effects

On held-out sources, jointly ablating the original Top-5 heads gives a smaller loss than summing their individual losses (subadditivity). LD measures preference for the fixed natural over literal reference class. Loss is the natural-prompt LD decrease from the unablated baseline, divided by each source’s original natural-prompt minus literal-prompt LD gap.

Zero and mean ablation replace head outputs with zeros and mean activations, respectively. Joint − sum is the combined loss minus the sum of individual losses; negative values indicate subadditivity. Brackets show 95% confidence intervals from paired source bootstrap resampling.

Held-out Top-5 ablationJoint lossSum of individual lossesJoint − sum [95% CI]Shapley vs. single-head ρ
Zero0.30070.3804−0.0797 [−0.1424, −0.0282]1.00
Mean0.31440.3923−0.0779 [−0.1419, −0.0264]1.00

Each row’s Shapley ranking uses the corresponding ablation mode.

Our explicit tests therefore support the robustness of the leading-head priorities to the tested interactions, while quantifying how those interactions change their combined effect.

[1] Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small

[2] Towards Best Practices of Activation Patching in Language Models: Metrics and Methods

[3] Exploring the Translation Mechanism of Large Language Models

Part 2/5

Weakness 2 & Question 1(b): Filtering and Reference Selection.

Weakness 2:

  1. The authors generate completion classes using natural and literal translation prompts and filter examples...

Question 1(b):

1 (b). How sensitive are the identified heads to the filtering and completion-selection criteria in App? Would the same heads be identified using independently annotated natural and literal translations?

Response to Weakness 2 & Question 1(b):

We thank the reviewer for asking how filtering and reference selection affect localization. Filtering is a deliberate, task-specific behavioral correctness check: it establishes a measurable, model-specific style contrast before localization. Sensitivity analyses recover a stable core, and independent output evaluations support the resulting heads’ behavioral value.

IOI [1] and Greater-than [2] studies establish the target behavior before tracing circuits. Factual-recall studies analyze correct predictions [3][4], while arithmetic circuit discovery selects correctly completed problems to reduce noise [5]. Here, we check that each instruction favors its corresponding reference class, using the model’s own translations.

Filtering and reference sensitivity (Qwen3.5-4B, English-to-Chinese)

Top-15 heads denotes the 15 highest-ranked heads. Retention counts overlap with the corresponding baseline ranking; split-half overlap compares rankings from two discovery subsets. The separation threshold is the minimum natural-prompt minus literal-prompt Equi-DIFF score gap.

CheckResult
Filtering vs. no behavioral filter, with sample size and domain composition matchedMean split-half Top-15 overlap: 7.97/15 vs. 4.98/15
Remove the behavioral filter11/15 original heads retained
Vary the separation threshold from 0.5 to 2.013–15/15 heads retained; all three stable core heads preserved
Resample five references per class on the same sources10/15 heads retained in each of two draws
Regenerate references and compare old/new localization on the same 14 eligible sources12/15 heads retained; all three stable core heads preserved
External vs. model-generated references, each using one reference per class on the same 15 sourcesOverlap with the original ranking: 8/15 external vs. 11.48/15 on average for model-generated references across reference-member selections, using pooled rankings

Pooled aggregation divides summed raw patching effects by summed natural–literal score gaps. External references are human-written WMT24++ original/revised translations, with style and semantic differences assessed separately by an LLM. They recover all three core heads: L18.H11, L31.H14, and L23.H5. Together, these checks show that filtering improves consistency in selecting the Top-15 heads, while threshold changes, reference regeneration, and external references recover the same three core heads.

Independent generated-output comparison

We compared steering with Top-15 heads localized from Model-generated references (Qwen3.5-4B’s own translations) or External references. Both used the same 15 discovery source texts, one reference per class, and pooled aggregation, with steering directions from the same discovery prompts and matched total injection norms. Head sets were frozen before evaluation.

On 15 held-out FLORES-200 source texts under a neutral prompt, Gemini-3.5-Flash evaluated new translations against the unsteered baseline in both orders, without reference translations:

MeasureModel-generatedExternalPaired difference (Model-generated − External)
Naturalness gain [95% CI]+1.000 [0.100, 1.833]+0.567 [−0.267, 1.333]+0.433 [0.033, 0.933]

These are each head set’s mean steered-minus-unsteered naturalness gains on the 1–7 scale. Paired difference is Model-generated minus External gain on the same sources; positive values favor Model-generated. Brackets show 95% confidence intervals from paired source-level bootstrap resampling.

In our broader evaluation, rankings are frozen before knockout and steering on independent source texts. LLM judges assess new outputs directly, with human checks on a subset.

Filtering improves localization consistency, sensitivity checks recover a stable core, and the matched neutral-prompt comparison supports model-generated references; independent evaluations establish the selected heads’ naturalness contribution beyond the discovery examples.

[1] Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small

[2] How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model

[3] Locating and Editing Factual Associations in GPT

[4] Dissecting Recall of Factual Associations in Auto-Regressive Language Models

[5] Arithmetic Without Algorithms: Language Models Solve Math With a Bag of Heuristics

Part 3/5

Weakness 3: Attention Heads as Intervention Targets.

  1. The paper focuses on attention heads and head-like groups (including DeltaNet outputs) but does not examine other components such as MLPs. Their contribution to translation naturalness therefore remains unclear.

Response to Weakness 3:

We thank the reviewer for asking why we prioritize heads and how MLPs contribute. To our knowledge, this is the first mechanistic study specifically targeting sentence-level translation naturalness in LLMs, connecting component localization with causal interventions that improve naturalness. Our focus on heads is supported by their architectural role and intervention results; the MLP analyses clarify their participation and the need for finer-grained study.

Why prioritize heads?

Attention heads transfer information across token positions [1], providing identifiable sites for studying how style instructions and context influence subsequent expression. MLPs process contextualized representations at each position. Function-vector research supports studying task control through small sets of heads [2], and this granularity enables direct comparison with word-translation heads in the same model [3]. Our analysis also includes DeltaNet’s head-like output groups.

Equi-DIFF, which scores classes of complete natural and literal translations, identifies heads with different functional effects from word-translation heads. Held-out ablation establishes their causal contribution, while steering improves judged naturalness across three models and 12 translation directions. Heads therefore provide a productive starting point for explaining and controlling this sentence-level behavior.

MLP evidence and component cooperation

Our preliminary exploration applied the same reference contrast and normalized preference measure to all 32 Qwen3.5-4B MLP blocks on 15 English-to-Chinese discovery sources. Effects were small in early layers and larger across multiple later layers, including L22–26. The leading layers by mean score were:

MLP layerL31L22L24L25
Normalized patching score−0.0937−0.0886−0.0741−0.0684
95% CI[−0.2089, 0.0318][−0.1229, −0.0574][−0.1061, −0.0400][−0.1003, −0.0354]

Scores measure the change in natural-over-literal reference preference when literal-prompt activations replace natural-prompt activations, divided by the original preference gap between prompts. More negative values indicate larger reductions; brackets give 95% confidence intervals from resampling source documents.

We then froze the rankings and ablated whole MLP blocks on 15 held-out FLORES-200 sources under the natural prompt. Gemini-3.5-Flash assessed naturalness and adequacy in both presentation orders. Zero and mean ablation replace outputs with zeros and discovery mean activations, respectively.

MLP set (layers)Zero: ΔNaturalnessZero: ΔAdequacyMean: ΔNaturalnessMean: ΔAdequacy
Top-1 (31)−1.600−0.650−1.967−1.273
Top-2 (31, 22)−1.767−0.967−3.500−3.233
Top-4 (31, 22, 24, 25)−1.933−1.167−5.733−5.300
Random-1 (26)−0.733−0.100−0.450−0.167
Random-2 (26, 20)−0.333+0.550−0.600+0.467
Random-4 (26, 20, 21, 18)−1.067−0.267−1.733−0.850

Values are mean ablated-minus-unmodified scores on 1–7 scales. Top denotes the highest-ranked blocks; Random denotes randomly selected deep-layer controls.

The MLP contribution is distributed across later layers, and whole-block interventions can affect both expression and meaning preservation. Effects depend on the selected layers and intervention: top-ranked mean ablation produces substantial deterioration, while zero ablation is milder. This supports finer-grained analysis of MLP functions and helps explain our focus on individually addressable heads.

In six selected translation contrasts with identical target prefixes, we also swapped head/group activations between natural and literal prompts in both directions. Fixing same-layer and later MLP outputs to their unperturbed receiving-prompt values reduces L18.H11’s mean effect on next-token preference toward the supplying prompt’s style by 92.1%, while L31.H14 retains 91.6%. These are a DeltaNet output group and an attention head, respectively.

Our study establishes causal and controllable head-level effects within this cooperative system. The distributed MLP effects motivate further study of neurons or feature subspaces together with token positions, to distinguish which representations contribute and where in the sequence they act. Tracing these interactions toward a complete circuit is a natural extension of the present findings.

[1] A Mathematical Framework for Transformer Circuits

[2] Function Vectors in Large Language Models

[3] Exploring the Translation Mechanism of Large Language Models

Reviewer_BeZuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper is a mechanistic study of translation naturalness in LLMs. To do so, the authors introduce a metric, equivalence-class score difference (Equi-DIFF), to differentiate the attention heads the literally translate from the ones that increase naturalness. They study this accross 3 models (Qwen3.5-4B, Qwen3.5-35B-A3B, Gemma-4-26B-A4B-IT) and two translation benchmarks (WMT24++, FLORES-200) across 12 translation pairs (English <=> German, Spanish, French, Japanese, Russian, Chinese). Naturalness is evaluated with pairwise judgment in a LLM-as-a-judge setting. Overall, this paper shows that multiple LLMs have consistent heads that functionally increase the naturalness of the translation, and that ablating them lowers the naturalness of the generations.

Critical strengths
  1. This paper studies an interesting research question not addressed in previous MT interpretability work.
  2. The paper validates its results in multiple settings (7 languages, 3 models).
  3. The metric introduced is well justified and may be used in broader settings for other mechanistic interpretability studies.
Critical weaknesses
  1. The methodology used to build the literal/natural translations via LLM generation may inflate the naturalness effect found.
  2. I would argue that naturalness is also part of translation quality. The steered generations do not show significantly better translation quality scores. This makes it unclear whether the heads increase the naturalness of a translation.
  3. As no LLM-judge/human agreement is reported, the naturalness improvement may reflect only what the LLM judge considers more natural, not what humans perceive as a natural translation.
  4. The heads are identified from very few examples (15), so there may be some noise in the identified heads.
Questions for authors
  1. How do you justify that the naturalness heads are separate from the functional “literal” heads found in previous work, given that both should be linked to translation quality?
  2. Would it be possible to reproduce your experiments with a model used in previous work that identified relevant attention heads for machine translation, in order to compare the heads found?
  3. Could you provide the pairwise agreement scores between the LLM judges and human annotators for naturalness?

Presentation details:

  • The paper would benefit from bigger, more readable plots, especially since the paper has some blank space.
8 comments

Part 1/6

Weakness 1: Model-Generated References for Localization.

  1. The methodology used to build the literal/natural translations via LLM generation may inflate the naturalness effect found.

Response to Weakness 1:

We thank the reviewer for questioning whether reference construction inflates naturalness gains. Our reference construction is a deliberate mechanistic-interpretability design: we establish a controlled, model-specific behavioral contrast before localizing its causal components. The resulting heads’ naturalness benefits are then evaluated independently on new, held-out translations.

Our on-policy references (generated by the model being localized) capture an existing style contrast by changing only naturally to literally for a fixed source and model. Establishing behavior before localization follows IOI and greater-than studies [1][2]; arithmetic circuit discovery uses correctly solved examples to reduce noise [3]. External references may express valid style contrasts while differing from the model’s generation preferences.

Our filter checks the natural/literal preference contrast before localization. For Qwen3.5-4B English-to-Chinese, filtering increases mean split-half overlap from 4.98 to 7.97 heads, with sample size and domain composition matched. This measures how many of the 15 highest-ranked heads (Top-15 heads) are shared between two discovery subsets’ rankings. Natural–literal score-gap thresholds of 0.5–2.0 retain 13–15 original heads, including all three core heads.

We compared steering with Top-15 heads localized using Model-generated references (Qwen3.5-4B’s own translations) or External references (WMT24++ original/revised pairs). Both used the same 15 discovery source texts, one reference per class, and pooled aggregation (summed raw patching effects divided by summed natural–literal score gaps). Steering directions came from the same discovery prompts, with matched total injection norms. Both head sets were frozen before evaluation. On 15 held-out FLORES-200 source texts under a neutral prompt, Gemini-3.5-Flash judged new translations against the unsteered baseline in both orders, without reference translations:

MeasureModel-generatedExternalPaired difference (Model-generated − External)
Naturalness gain [95% CI]+1.000 [0.100, 1.833]+0.567 [−0.267, 1.333]+0.433 [0.033, 0.933]

The Model-generated and External columns report each head set’s mean steered-minus-unsteered naturalness score on the 1–7 scale. Paired difference is their mean within-source difference; positive values favor the Model-generated head set. Brackets give 95% confidence intervals from paired source-level bootstrap resampling.

Separately, all four LLM judges report naturalness gains on Qwen3.5-4B English-to-Chinese WMT24++ outputs. Four independent annotators assessed 200 held-out English-to-Chinese pairs from Qwen3.5-4B and Qwen3.5-35B-A3B. Their 800 naturalness judgments preferred steering in 72.0%, the baseline in 11.0%, with 17.0% ties.

Together, the matched comparison supports our reference design, while independent LLM and human evaluations substantiate the resulting heads’ naturalness benefit on held-out translations.

[1] Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small

[2] How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model

[3] Arithmetic Without Algorithms: Language Models Solve Math With a Bag of Heuristics

Part 2/6

Weakness 2: Naturalness and Translation Quality.

I would argue that naturalness is also part of translation quality. The steered generations do not show significantly better translation quality scores. This makes it unclear whether the heads increase the naturalness...

Response to Weakness 2:

We thank the reviewer for highlighting naturalness as part of translation quality. Steering improves naturalness, with accompanying adequacy trade-offs.

As-Safi and Ash-Sharifi [1] distinguish natural expression (vocabulary, sentence structure, cohesion, idioms) from content preservation. COMET/MetricX predict overall quality, not naturalness separately [2][3]. Liu et al. [4] find that a human-validated measure of translationese (features of translated text) correlates weakly with XCOMET/COMET-Kiwi.

Taking Qwen as an example, steering the Top-15 heads (15 highest-ranked heads) at α=1 yields positive mean naturalness changes on held-out translations in all 144 settings (2 models × 2 corpora × 12 directions × 3 prompts); 134 have 95% confidence intervals above zero.

Gemini-3.5-Flash evaluates these Qwen3.5-4B/WMT24++ directions (379 pairs/row). Neutral/Natural/Literal use no modifier/“naturally”/“literally.” Nat./Adeq. denote naturalness/adequacy; W/L/T denotes steering wins/baseline wins/ties. Scores are 1–7; Δ is mean steered−baseline (range −6 to +6); brackets give 95% CIs.

DirectionPromptNat. W/L/TNat. Δ [95% CI]Adeq. W/L/TAdeq. Δ [95% CI]
En→ZhNeutral210/82/87+0.605 [0.468, 0.748]80/129/170−0.134 [−0.214, −0.048]
Natural205/84/90+0.542 [0.417, 0.661]60/160/159−0.315 [−0.419, −0.208]
Literal221/75/83+0.583 [0.454, 0.709]67/135/177−0.212 [−0.300, −0.132]
En→JaNeutral161/127/91+0.164 [−0.001, 0.328]102/140/137−0.148 [−0.268, −0.034]
Natural167/122/90+0.197 [0.040, 0.357]117/132/130−0.082 [−0.196, 0.032]
Literal191/96/92+0.425 [0.291, 0.570]130/109/140+0.079 [−0.029, 0.185]
De→EnNeutral212/56/111+0.665 [0.543, 0.807]43/144/192−0.282 [−0.383, −0.180]
Natural190/60/129+0.504 [0.386, 0.649]40/165/174−0.394 [−0.486, −0.307]
Literal227/44/108+0.786 [0.649, 0.934]55/95/229−0.084 [−0.182, 0.009]

Across 12 equally weighted directions, Literal gives the largest naturalness gain and smallest adequacy decrease in each Qwen model–corpus combination (Appendix E.6, Tables 31–36).

Separately, four judges assess 1,200 4B En→Zh pairs on WMT24++ at Top-15/α=1:

LLM judgeNat. W/L/TNat. Δ [95% CI]Adeq. W/L/TAdeq. Δ [95% CI]
Qwen3.8-Max609/191/400+0.527 [0.458, 0.595]173/448/579−0.318 [−0.380, −0.258]
GPT-5.5-0424642/250/308+0.451 [0.392, 0.509]152/349/699−0.153 [−0.193, −0.116]
Gemini-3.5-Flash672/236/292+0.618 [0.543, 0.690]173/377/650−0.243 [−0.301, −0.185]
Claude-Opus-4.6510/152/538+0.343 [0.297, 0.389]124/290/786−0.147 [−0.186, −0.108]

All four naturalness CIs lie above zero.

Four annotators independently assessed the same 200 En→Zh pairs (100 sources) from Qwen3.5-4B/35B-A3B at Top-15/α=1. Mean naturalness preferences were 72% steering, 11% baseline, 17% ties. Adequacy favored steering or tied in 65.13%; 34.88% favored baseline. Percentages include ties.

Across four Qwen model–corpus combinations, MetricX decreases by 0.025–0.124 points (improvement); COMET changes by −0.0016 to +0.0005.

Example: Qwen3.5-35B-A3B, WMT24++, Neutral:

Source: Floor repair person is doing a vacuum, so hopefully that means he's finishing up.

Baseline: 地板维修人员正在吸尘,希望这意味着他即将完工。

Steered: 地板维修师傅正在吸尘,看来快收尾了。

The steered version uses more conversational phrasing, such as “维修师傅” and “快收尾了.” It also changes “希望” (“hopefully”) to “看来” (“it seems”), shifting a hope to an inference. This illustrates a specific trade-off: more natural wording accompanies a meaning change (mean adequacy change: −1.00), helping explain why overall quality scores may change little.

For 4B, Gemini-3.5-Flash evaluates weaker settings across 12 WMT24++ directions and three prompts (3,600 pairs/row). W+T means steering preferred or tied.

InterventionNat. W/L/TNat. Δ [95% CI]Nat. W+TAdeq. W/L/TAdeq. Δ [95% CI]Adeq. W+T
Top-5, α=0.25532/356/2712+0.057 [0.036, 0.078]90.1%252/341/3007−0.026 [−0.043, −0.010]90.5%
Top-5, α=0.5852/505/2243+0.125 [0.097, 0.153]86.0%352/538/2710−0.053 [−0.080, −0.031]85.1%

Intervals use document-cluster bootstrap with equal direction weights.

Direct assessments reveal naturalness gains and adequacy trade-offs alongside small overall quality changes.

[1] Naturalness in Literary Translation

[2] COMET-22: Unbabel-IST 2022 Submission for the Metrics Shared Task

[3] MetricX-24: The Google Submission to the WMT 2024 Metrics Shared Task

[4] Translationese-index: Using Likelihood Ratios for Graded and Generalizable Measurement of Translationese

Part 3/6

Weakness 3 & Question 3: Human–LLM Agreement on Naturalness.

Weakness 3:

As no LLM-judge/human agreement is reported, the naturalness improvement may reflect only what the LLM judge considers more natural, not what humans perceive as a natural translation.

Question 3:

Could you provide the pairwise agreement scores between the LLM judges and human annotators for naturalness?

Response to Weakness 3 & Question 3:

We thank the reviewer for requesting human validation and pairwise agreement scores, which help assess whether the observed naturalness improvement aligns with human perceptions. Professional translators corroborate the naturalness improvement. Taking English-to-Chinese translation with the Qwen family as an example, we observe 90.63% exact agreement (725/800) with Gemini-3.5-Flash across the two evaluated models.

We conducted human evaluations across all 12 translation directions. Taking English-to-Chinese as an example, four professional translators, all native Chinese speakers proficient in English, independently assessed the same 200 translation pairs for naturalness and adequacy (preservation of the source meaning). These cover 100 source IDs from WMT24++ and FLORES-200, with 100 pairs each from Qwen3.5-4B and Qwen3.5-35B-A3B. Each pair compares the baseline with steering using the Top-15 highest-ranked heads and a steering-strength multiplier of α = 1.

Table 1 reports the preference distributions. Human annotators preferred the steered translation for naturalness in 72.0% of judgments, compared with 65.5% for Gemini-3.5-Flash, a difference of 6.5 percentage points. For adequacy, both humans and the LLM preferred the baseline more often than the steered translation, indicating a trade-off at this steering strength.

DimensionAssessmentSteered preferredTieBaseline preferred
NaturalnessGemini-3.5-Flash65.50%14.50%20.00%
NaturalnessFour human annotators72.00%17.00%11.00%
AdequacyGemini-3.5-Flash14.00%41.00%45.00%
AdequacyFour human annotators16.25%48.88%34.88%

Table 1. English-to-Chinese preference distributions across both models. Each human annotator and the LLM evaluated the same 200 pairs. Human percentages are averaged across the four annotators. Percentages include ties and may not sum to 100% due to rounding.

To address the agreement question directly, we compared each annotator’s naturalness judgment with the LLM’s judgment on the same translation pair. Table 2 reports the individual agreement scores and their averages, both by model and across both models.

Rater pairQwen3.5-4B (%)Qwen3.5-35B-A3B (%)Both models (%)
Gemini-3.5-Flash vs. Human 191.0094.0092.50
Gemini-3.5-Flash vs. Human 289.0095.0092.00
Gemini-3.5-Flash vs. Human 392.0093.0092.50
Gemini-3.5-Flash vs. Human 482.0089.0085.50
Mean88.5092.7590.63

Table 2. Pairwise human–LLM exact agreement on English-to-Chinese naturalness. Each annotator evaluated 100 pairs per model and 200 pairs in total. The mean is taken across the four annotators.

Across the 200 English-to-Chinese pairs from both models, mean pairwise agreement among human annotators was 94.42%, with Fleiss’ κ = 0.873. Among the 40 English-to-Chinese pairs where the LLM preferred the baseline for naturalness, the 160 human judgments comprise 54 steering preferences, 19 ties, and 87 baseline preferences.

For Qwen3.5-4B, human–LLM naturalness agreement across all 12 translation directions ranged from 87.68% to 94.13%, as reported in Table 3.

DirectionAgreement (%)DirectionAgreement (%)DirectionAgreement (%)
En→De89.21En→Fr90.25En→Ru88.13
De→En92.20Fr→En94.13Ru→En90.78
En→Es93.13En→Ja87.68En→Zh88.50
Es→En91.77Ja→En88.30Zh→En92.22

Table 3. Human–LLM exact agreement on naturalness across 12 translation directions for Qwen3.5-4B, using Gemini-3.5-Flash as the LLM judge. En→Zh matches the Qwen3.5-4B mean in Table 2. En: English; De: German; Es: Spanish; Fr: French; Ja: Japanese; Ru: Russian; Zh: Chinese.

These results show that the naturalness improvement observed with steering is also perceived by professional translators.

Comments on the paper

1 comment

Dear Reviewers, Area Chairs, and Program Chairs,

We thank all three reviewers for helping us strengthen the paper and clarify its contributions. Our paper combines Equi-DIFF, a measure over classes of model-generated outputs, with sparse causal interventions to uncover the mechanisms underlying translation naturalness. Experiments across three models, 12 translation directions, and two corpora demonstrate the effectiveness of our approach.

Reviewers have acknowledged the motivation and contributions of our method and the breadth of our analysis.

[Scientific motivation and contribution]

  • Reviewer BeZu: “This paper studies an interesting research question not addressed in previous MT interpretability work.”
  • Reviewer BeZu described the metric as “well justified” and potentially useful in broader mechanistic interpretability studies.
  • Reviewer Y4si: “Equi-DIFF addresses a relevant measurement problem.”

[Causal evidence and breadth]

  • Reviewer Y4si: “The causal validation goes beyond differences in head rankings.”
  • Reviewer vzAP noted that “a small set of heads can influence translation naturalness across different models and language directions.”

During the response period, we carefully addressed all reviewer comments and conducted supplementary experiments. We summarize our responses, additional evidence, and manuscript revisions below.

[Method clarification and stability]

  • We clarified the rationale for using filtered model-generated references for localization and demonstrated that filtering substantially improves localization stability, with the same core heads recovered across filtering thresholds and reference sources.
  • We conducted human review for all 15 discovery sentences, validating the references and identifying syntax, word order, and register as key natural–literal differences.
  • We matched interventions on 15 held-out sources to demonstrate a +0.433-point naturalness advantage for heads localized with model-generated rather than external references on a 1–7 scale.
  • We verified sampling stability, recovering 14/15 heads on average under leave-one-out analysis and 11/15 with fresh discovery sources.

[Component roles]

  • We conducted crossed ablations in Qwen3.5-4B to distinguish our sentence-level naturalness heads from lexical-translation heads localized using the method of prior work, highlighting the novelty of our findings.
  • We evaluated all 32 Qwen3.5-4B MLP blocks and demonstrated that whole-MLP ablation affects both naturalness and meaning preservation, supporting our focus on attention heads as primary research targets.

[Human validation]

  • We clarified that four LLM judges consistently reported positive naturalness gains when evaluating translations across language pairs and different generation models.
  • We clarified that four professional translators independently assessed naturalness and adequacy for translations in each language pair, validating our LLM-based evaluation.
  • We clarified that human judgments further strengthen our results, with translators preferring steered translations more often than the LLM judge.

[Naturalness–Adequacy Trade-offs]

  • We clarified the naturalness gains and adequacy trade-offs accompanying small changes in COMET and MetricX by reporting win/loss/tie counts and paired score changes under neutral, natural, and literal prompts.
  • We clarified that an LLM-assisted audit of error severity after steering showed that minor semantic errors were more frequent than major ones among flagged adequacy losses, supporting the effectiveness of steering the localized naturalness heads.
  • We clarified that results from milder steering on 3,600 Qwen3.5-4B pairs across 12 directions and three prompts demonstrate an adjustable balance between naturalness gains and adequacy costs.

[Presentation and discussion]

  • We enlarged the figures to improve readability.
  • We restored the source's television emoji to explain why “watch some TV” was supported by the input (Appendix E, Table 21).
  • We thank Y4si for reassessing the paper and raising the rating to Weak Acceptance.
  • We appreciate vzAP's acknowledgment that “The new analyses address many of my concerns” and that the additional evidence was helpful.
  • We remain available to address further questions from BeZu, from whom we have not yet received a follow-up response.

All noted changes have been incorporated into the revised manuscript. We thank the Area Chair for overseeing the discussion and considering the revised manuscript and additional evidence.

With thanks and best regards,

Authors of #58583

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.