acceptodds
Under review as a conference paper at ICLR 2027

ANCHOR: Cross-Lingual Reasoning with Anchored Reinforcement Learning Under Accuracy-Language Trade-off

Abstract

Chain-of-Thought LLMs support multilingual reasoning, yet on low-resource language questions, they often rely on high-resource languages such as English for accuracy. Forcing reasoning in the question language improves readability for question-language speakers but may compromise accuracy. To address this trade-off, we propose Anchor, a representation anchoring and behavioral framework for cross-lingual reasoning. We structure the response into three segments: a multilingual reasoning segment, a question-language public explanation, and a verifiable answer. At the representational level, layer-wise probing reveals a recovery pattern in question language signals, based on which we introduce a one-sided anchoring loss to preserve question language routing; at the behavioral level, we design a multi-dimensional reward with two language-gated rewards to shape language choices under answer correctness and format validity. Both levels are jointly optimized in a DAPO-based policy optimization process. Experiments show Anchor enables flexible cross-lingual reasoning on multiple benchmarks. On MGSM, Anchor reaches 86.07% average accuracy across 11 languages, delivering 1.90% and 1.94% gains over the baseline for 4 low-resource and 7 high-resource languages respectively, alongside 74.2% public-segment question-language alignment. On MMLU-ProX-Lite and Global-MMLU-Lite, Anchor boosts performance on 41 of 52 languages and modest alignment improvements.

Primary area:
interpretability and explainable AI
Keywords:
Cross-Lingual Reasoning, Complex Reasoning, Reinforcement Learning, Low-Resource Languages

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HVxqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

ANCHOR combines DAPO-based reinforcement learning, language rewards, and a vocabulary-margin anchoring loss applied to the final three transformer layers. The goal is to support mixed-language reasoning while providing an explanation in the question language. Experiments with Qwen3-8B-Instruct show improved accuracy and language alignment on MGSM, accuracy gains on MMLU-ProX-Lite, and little accuracy change on Global-MMLU-Lite.

Critical strengths

The comparison with R4, which retains the behavioral rewards but removes anchoring, provides useful evidence that the auxiliary loss contributes beyond reward design. On MGSM, adding anchoring raises accuracy from 84.76% to 86.07% and public-language alignment from 69.1 to 74.2. Improving both measures is a meaningful result for the problem the paper addresses.

Critical weaknesses
  1. The public-explanation requirement largely fails on the knowledge benchmarks. Tables 17–18 report a 97.5% empty-public-segment rate on both Lite benchmarks under the reported sampling setting, with alignment scores of only 2.2% and 1.9%. Providing an explanation that question-language speakers can read is a central motivation of the paper. These results substantially weaken the claim that ANCHOR addresses the accuracy–language trade-off across tasks.
  2. The evidence does not establish why the proposed recovery band is important. Figures 3–4 show improvements in the vocabulary margin that the loss directly optimizes. The R4 comparison establishes the benefit of adding the loss, but does not distinguish recovery-band anchoring from a more general language regularizer. There is no final-layer-only control or comparison with another layer band. The transfer of thresholds from question tokens to generated reasoning is also insufficiently validated. I find the empirical gain more convincing than the proposed explanation for it.
Questions for authors
  1. Evaluation clarification: What causes the 97.5% empty-public rate: the evaluation prompt, response parsing, or the learned behavior? Does explicitly requesting a public explanation preserve the reported accuracy while substantially improving alignment?
  2. Evidence for the central mechanism: How does the proposed recovery band compare with final-layer-only anchoring or a matched alternative layer band under the same training setup? This comparison would materially affect my assessment of the technical contribution.
5 comments

We sincerely thank you for your time and constructive feedback on our work. We provide point-by-point responses to all your questions and concerns below.

Response to Q1: Evaluation Clarification

We appreciate your attention to this interesting phenomenon. We first clarify that the Lite benchmarks were introduced to evaluate cross-task generalization, rather than to serve as the primary setting for method development. To maintain consistency with the MGSM experiments, we initially used the same response structure and language requirements across these benchmarks. However, unlike MGSM, both Lite benchmarks consist of multiple-choice questions with option letters as final answers. Under this prompt, the model often outputs the answer directly and omits the public explanation, as we also observed when inspecting the raw responses. This phenomenon reflects a mismatch between the prompt structure and the multiple-choice task format, rather than necessarily indicating a limitation of ANCHOR.

To directly address your concern, we conducted an additional comparison. Specifically, we revised the prompt to explicitly require the model to provide a complete explanation in the question language before giving its final answer, while keeping the model and all other evaluation conditions unchanged. The results are presented below.

Results on Global-MMLU-Lite:

ModelAAMMpgp_gpenp_{en}pop_oEmpty-public rateAccuracy
ANCHOR: Original prompt2.469.534.3988.347.2797.2969.93
ANCHOR: Revised prompt88.028.213.0889.667.262.5166.98

Results on MMLU-ProX-Lite:

ModelAAMMpgp_gpenp_{en}pop_oEmpty-public rateAccuracy
ANCHOR: Original prompt2.077.285.1885.978.8597.3254.73
ANCHOR: Revised prompt75.966.362.0689.028.9114.7554.47

The revised prompt substantially improves language behavior on both benchmarks. On Global-MMLU-Lite, the language alignment score (AA) increases from 2.46 to 88.02, while the empty-public rate decreases from 97.29% to 2.51%. On MMLU-ProX-Lite, AA increases from 2.07 to 75.96, and the empty-public rate decreases from 97.32% to 14.75%. Therefore, the original results should not be interpreted as evidence that ANCHOR fails to achieve its intended language behavior. Rather, they highlight the importance of adapting the output requirements to the task format when evaluating cross-task generalization. We promise to add this important experiment to the main text in the revised version.

Response to Q2: Evidence for the Central Mechanism

We sincerely thank you for pointing out the need for a more comprehensive ablation of the Recovery Band. To address this concern, we have conducted three groups of additional experiments: shifted bands outside the identified recovery region, a reduced band, and an expanded band. The only variable in all experiments is the selection of the anchor layer, and the rest of the training and inference settings are exactly the same as the Anchor.The first two groups have been completed, while the expanded band experiment is still running and is expected to finish within approximately three days. We will provide its results in the updated response.

ConfigenesfrdezhjaswbnruthteAVGΔAVG\Delta_{\mathrm{AVG}}A‾\overline{A}ΔA‾\Delta_{\overline{A}}M‾\overline{M}ΔM‾\Delta_{\overline{M}}bo
ANCHOR (L34–L36)96.0093.2086.4088.8088.4086.0062.0085.2090.8089.2080.8086.07—74.18—10.04—8.40
L36 only93.6090.0085.6085.2088.0083.6055.6084.0092.0086.8077.2083.78−2.2974.45+0.279.72−0.339.60
L29–L31 (shifted 5 layers earlier)94.4091.2087.2088.0088.0086.8056.8085.2093.2088.8080.4085.45−0.6269.61−4.577.84−2.206.40
L24–L26 (shifted 10 layers earlier)94.8091.2086.0088.0084.8084.4056.0084.4091.6089.2079.6084.55−1.5369.13−5.058.12−1.928.00

For the reduced-band setting, we retain only L36. For the shifted-band controls, we move the three-layer band five and ten layers earlier, resulting in L29–L31 and L24–L26, respectively. Compared with the original L34–L36 configuration, all three controls yield lower average accuracy, with decreases of 2.29, 0.62, and 1.53 points, respectively. The two shifted bands also exhibit notably lower A‾\overline{A} and M‾\overline{M}, while the L36-only setting achieves a slightly higher A‾\overline{A} with only a marginal difference, but shows a decrease in M‾\overline{M}. Overall, these results suggest that both the location and coverage of the constrained layers matter, and that the recovery band identified by our initial layer-wise analysis provides the best overall performance among the tested configurations. We promise to add this important experiment to the main text in the revised version.

We sincerely hope you can reconsider our work. We are readily available for further discussion until all your concerns are fully resolved. We commit that all key supplementary materials from this discussion period will be fully incorporated into the final version of the paper.

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

This paper proposes ANCHOR, a reinforcement learning framework designed to address the trade-off between reasoning accuracy and question-language fidelity in multilingual Chain-of-Thought (CoT) reasoning.

The authors structure model responses into three components: a multilingual reasoning segment, a public explanation written in the question language, and a canonical answer for automatic verification. To optimize these components, the proposed framework combines two mechanisms. First, the authors identify a recovery pattern in language-related representations within the final Transformer layers and introduce a Representation Anchoring Loss to preserve these signals during reinforcement learning. Second, they design a multi-dimensional reward function that considers answer correctness, output format, question-language alignment in public explanations, and mixed-language reasoning involving the question language and English. Both mechanisms are jointly optimized using DAPO.

Experimental results show that ANCHOR, based on Qwen3-8B-Instruct, achieves an average accuracy of 86.07% on MGSM, improving upon the initial model by approximately 1.92 percentage points while substantially improving public-segment language alignment. The authors also report a statistically significant improvement of 1.31 percentage points over the variant without Representation Anchoring Loss (R4). Additional improvements are observed on MMLU-ProX-Lite, although no statistically significant accuracy improvement is found on Global-MMLU-Lite.

The combination of output-level behavioral rewards and representation-level constraints is an interesting direction for multilingual reasoning. However, the validity of the language representation signals used for anchoring and the necessity of explicitly rewarding mixed-language reasoning are not sufficiently established. These concerns are directly related to the fundamental design choices and claimed contributions of the proposed method.

Critical strengths

S1. Meaningful problem formulation addressing accuracy and language fidelity

The paper addresses an important practical challenge in multilingual reasoning. While reasoning in high-resource languages such as English may improve accuracy, generating explanations in the user's question language is essential for accessibility and usability. The proposed framework explicitly distinguishes these objectives and attempts to optimize them jointly.

S2. Integration of behavioral rewards and representation-level constraints

The introduction of an auxiliary loss that directly constrains language-related representations in selected Transformer layers is an interesting methodological contribution. Combining this constraint with DAPO-based reinforcement learning offers a potentially useful direction for preserving specific model behaviors or representations during post-training.

S3. Evaluation across multiple languages and benchmarks

The authors construct a training corpus covering 50 languages and evaluate the proposed method on MGSM, Global-MMLU-Lite, and MMLU-ProX-Lite. The paper also includes ablation studies of the major components. In particular, the use of paired statistical tests and multiple-comparison corrections to evaluate the contribution of Representation Anchoring Loss on MGSM is a positive aspect of the experimental methodology.

Critical weaknesses

W1. Insufficient justification for using Latin-script vocabulary as the reference in the Vocabulary Margin

A central component of the proposed Representation Anchoring mechanism is the Vocabulary Margin defined in Equation (1), which measures the difference between the log-sum-exp scores of question-language tokens and Latin-script tokens.

However, the paper does not sufficiently justify why Latin-script vocabulary constitutes an appropriate reference distribution for measuring question-language representations.

First, Latin script is shared by numerous languages, including English, German, French, and Spanish. Consequently, an increase in the probability mass assigned to Latin-script tokens cannot be directly interpreted as a shift toward English reasoning.

Second, the proposed margin measures only the relative probability mass between two predefined token sets. An increase in this margin does not necessarily imply an increase in the absolute probability mass assigned to the question language or greater use of that language during reasoning. Moreover, differences in vocabulary size and token composition may influence the measured values.

Although the authors acknowledge that the Vocabulary Margin is a proxy rather than a complete language classifier, this limitation is particularly important because the proposed Recovery Band and Anchoring Loss are both constructed based on this measure.

The paper would benefit from additional experiments examining whether the same recovery pattern emerges when alternative reference distributions are used, such as language-specific English token sets or normalized measures relative to the full vocabulary.

W2. Limited validation of the representation-level mechanism for languages sharing the same script

Although the proposed framework is trained on 50 languages, Representation Anchoring Loss is applied only to 21 non-Latin-script languages. The remaining 29 Latin-script languages are directly optimized only through behavioral rewards.

Figure 1(b) includes German, French, Spanish, and other Latin-script languages. However, this analysis relies on the same script-based Vocabulary Margin and does not directly distinguish the internal representations of different languages sharing the Latin script.

For example, when a model receives a question in German but performs its reasoning in English, it remains unclear whether the proposed representation-level analysis can identify the distinction between these two languages or determine whether the original question-language representation is preserved.

The paper provides language-specific accuracy and output-language behavior measurements, including results for German. However, these behavioral evaluations do not directly validate the proposed representation-level hypothesis.

A more convincing evaluation would involve language-specific hidden-state probes capable of distinguishing languages that share the same script. Such experiments could determine whether the proposed Recovery Band exists independently of script-level differences.

Furthermore, Chinese and Tibetan do not exhibit the same concentrated recovery pattern as most of the other non-Latin-script languages. This observation raises additional questions about whether the Recovery Band represents a general phenomenon in multilingual language representations or is specific to certain languages and measurement conditions.

W3. Insufficient justification for explicitly rewarding mixed-language reasoning

The primary objective of the proposed framework is to maintain reasoning accuracy while ensuring that public explanations are generated in the question language.

From this perspective, the language used during the reasoning process should not necessarily matter, provided that the model achieves high accuracy and produces a correct, high-quality explanation in the target language.

However, the Language Reward in Equation (3) explicitly encourages the joint use of the question language and English during reasoning. Therefore, the proposed method does not merely allow flexible language selection; it introduces an additional preference for mixed-language reasoning.

The paper does not adequately establish why mixed-language reasoning should be preferred over monolingual reasoning or whether explicitly encouraging language mixing is necessary for improving reasoning accuracy or final explanation quality.

Table 1 includes comparisons with English-CoT and Native-CoT baselines. However, these experiments involve different models and training configurations rather than a controlled comparison using the same ANCHOR model. Consequently, the results do not isolate the effect of reasoning-language selection.

Furthermore, the ablation study in Table 2 does not independently evaluate the contribution of the reasoning-mixing reward. In particular, the language reward used in R3 combines public-segment question-language alignment and reasoning-segment language mixing. Therefore, the individual contribution of the mixed-language reasoning reward cannot be determined from the reported experiments.

An additional ablation that removes the reasoning-mixing reward while preserving answer correctness and public-segment language alignment rewards would be valuable. Since the group-filtering mechanism also incorporates language-mixing criteria, its contribution should be controlled accordingly.

Moreover, a direct comparison of Native-only, English-only, and Flexible Reasoning using the same trained model would help establish whether mixed-language reasoning provides any measurable advantage in accuracy or explanation quality.

Without such experiments, it is difficult to determine whether the proposed mixed-language reward contributes to the intended objectives or merely imposes an additional behavioral preference.

W4. Limited achievement of the public-explanation objective on knowledge-oriented benchmarks

According to Table 18 in Appendix D, ANCHOR produces empty public explanations for approximately 97.5% of responses on both Global-MMLU-Lite and MMLU-ProX-Lite.

These results are reported under a sampling configuration with a repetition penalty and should therefore be distinguished from the greedy-decoding setting used for the main accuracy comparisons.

Nevertheless, this observation is important because generating a public explanation in the question language is one of the central objectives of the proposed framework.

If most responses contain only a canonical answer without a meaningful public explanation, the method cannot be considered to have fully achieved its intended language-fidelity objective on these tasks.

Additionally, the improvement in accuracy on Global-MMLU-Lite is not statistically significant. Therefore, the reported results do not consistently demonstrate simultaneous improvements in reasoning accuracy and public-explanation language fidelity across different task types.

The authors should investigate the causes of empty public explanations and report the proportion of responses that contain valid, meaningful explanations in the question language. It would also be useful to evaluate whether the performance improvements persist when non-empty public explanations are explicitly required.

Questions for authors

Q1. Justification of the reference vocabulary in the Vocabulary Margin

What is the specific motivation for using Latin-script vocabulary as the reference distribution in Equation (1)? Would the same Recovery Band emerge if the question-language representations were compared against language-specific English vocabulary or alternative normalized vocabulary measures? How sensitive are the observed margins to differences in token-set size and composition?

Q2. Representation-level validation for languages sharing the same script

Can the proposed representation-level mechanism distinguish languages that share the Latin script, such as German and English? Would language-specific hidden-state probes reveal a similar Recovery Band for these languages? Additionally, how do the authors explain the performance improvements observed in Latin-script languages that do not directly receive Representation Anchoring Loss?

Q3. Independent contribution of the mixed-language reasoning reward

What happens when the reasoning-mixing reward, R_reas, is removed while retaining the correctness and public-segment language alignment rewards? Does explicitly encouraging mixed-language reasoning improve accuracy or public-explanation quality after controlling for the language-mixing conditions in group filtering?

Q4. Direct comparison of monolingual and mixed-language reasoning

Can the authors compare Native-only, English-only, and Flexible Reasoning using the same ANCHOR checkpoint? Reporting accuracy, actual reasoning-language usage, and public-segment language alignment under these controlled conditions would provide stronger evidence for the necessity of mixed-language reasoning.

Q5. Empty public explanations on knowledge-oriented benchmarks

Why does ANCHOR generate empty public explanations for most questions in Global-MMLU-Lite and MMLU-ProX-Lite? Would the reported accuracy improvements persist if the model were required to generate a valid, non-empty explanation in the question language?

7 comments

We sincerely thank you for your time and constructive feedback on our work. We provide point-by-point responses to all your questions and concerns below.

Response to Q1: Justification of the Reference Vocabulary in the Vocabulary Margin

We thank you for pointing out this lack of clarity, and apologize for the insufficient explanation in the original manuscript. We clarify this distinction below.

We use the Latin-script vocabulary as a common reference when measuring the relative generation preference of non-Latin-script languages. Specifically, the Vocabulary Margin is defined as

μl(t)=log⁡P(Vg)−log⁡P(Vlat),\mu_l(t)=\log P(V_g)-\log P(V_{\mathrm{lat}}),

where VgV_g denotes the token set associated with the question language and VlatV_{\mathrm{lat}} denotes the Latin-script token set. This metric compares the probability mass assigned to these two token sets; it neither identifies the Latin-script vocabulary exclusively with English nor measures the absolute probability of generating the question language.

During training, the lower bound αμˉl,g\alpha\bar{\mu}_{l,g} is computed separately for each language and layer based on the initial model's margin, with the tokenizer and token sets kept fixed. Therefore, the constraint regulates changes in the relative generation preference with respect to the initial model, rather than imposing a universal threshold across languages or layers.

We will clarify the motivation, definition, and scope of the Vocabulary Margin in the revised manuscript.

Respons to Q2. Representation-level validation for languages sharing the same script

We thank the reviewer for raising this concern. We apologize that the original manuscript did not sufficiently distinguish the applicability of representation-level anchoring from that of behavior-level reinforcement learning. We clarify this distinction below.

The Vocabulary Margin is constructed using script-based token sets and cannot reliably distinguish languages sharing the Latin script, such as German and English. Representation Anchoring is therefore applied directly only to non-Latin-script languages with distinguishable token sets. Latin-script languages still participate in behavior-level reinforcement learning: fastText predicts specific language labels and can distinguish German, English, French, and other languages, so their samples receive feedback for correctness, formatting, public-segment language alignment, and reasoning-language behavior. Improvements in these languages can thus arise from language-specific behavioral rewards and cross-lingual transfer through shared parameter updates. We will clearly distinguish these two training mechanisms in the revision, rather than attribute improvements in Latin-script outputs to a representation-level constraint applied directly to those languages.

Respons to Q3. Independent contribution of the mixed-language reasoning reward

Our motivation for the reasoning-language reward is to preserve the model's problem-solving ability while encouraging the question language to participate in reasoning. Prior work and our early experiments suggest that requiring the model to reason entirely in the question language can noticeably reduce accuracy. [Cross-lingual Collapse](https://arxiv.org/html/2506.05850v3#S3.SS4) observes that strong constraints on target-language reasoning reduce some of the accuracy gains, while [AdaMCoT](https://ojs.aaai.org/index.php/AAAI/article/download/40678/44639) shows benefits from selecting reasoning languages adaptively for each problem rather than using a single fixed language. Our experiments build on this idea: we allow the reasoning segment to draw on high-resource languages while encouraging participation of the question language, and require the public segment to provide a complete explanation in the question language. Specifically, RreasR_{\mathrm{reas}} encourages joint participation of the question language and English, preserving the ability to reason with English while reducing cases in which the question language is excluded from reasoning altogether.

Respons to Q4. Direct comparison of monolingual and mixed-language reasoning

We added MGSM reasoning-language comparisons using the same best ANCHOR checkpoint, prompting it to reason only in the question language or only in English, and compared them with the existing free-reasoning main result. The two new conditions change only the reasoning-language instruction; the questions, output budget, answer grading, and requirement for the public segment to use the question language remain unchanged. The tables below report per-language accuracy, public-segment alignment, and reasoning-language distributions using the main paper’s evaluation definitions.

ModelenesfrdezhjaswbnruthteAVGA‾\overline{A}M‾\overline{M}bo
ANCHOR: Native-only prompt95.2093.6086.4087.2088.8084.4055.6082.8091.2088.8079.6084.8776.286.039.60
ANCHOR: English-only prompt95.2092.4085.6086.4089.2086.0060.0084.0092.0090.4080.0085.5662.148.379.20
ANCHOR: Free reasoning96.0093.2086.4088.8088.4086.0062.0085.2090.8089.2080.8086.0774.1810.048.40
ModelAAMMpgp_gpenp_{en}pop_oEmpty-public rate
ANCHOR: Native-only prompt76.286.0349.1446.174.6913.13
ANCHOR: English-only prompt62.148.373.2790.146.5931.16
ANCHOR: Free reasoning74.1810.047.4985.916.6015.24

Free reasoning achieves the highest mean accuracy and higher public-segment alignment than the English-only prompt, while omitting public explanations less often. Requiring reasoning in the question language reduces mean accuracy, with the decrease concentrated mainly in low-resource languages. The language distributions show that the question-language prompt increases its participation, although substantial English usage remains. The monolingual labels in the tables therefore describe the prompting instructions, while the distributions reflect actual language use. Here, pg,pen,pop_g,p_{en},p_o are mean fastText probabilities over valid reasoning segments. These results support preserving flexibility in reasoning-language choice while setting a separate question-language requirement for public explanations.

Respons to Q5. Empty public explanations on knowledge-oriented benchmarks

We appreciate your attention to this interesting phenomenon. We first clarify that the Lite benchmarks were introduced to evaluate cross-task generalization, rather than to serve as the primary setting for method development. To maintain consistency with the MGSM experiments, we initially used the same response structure and language requirements across these benchmarks. However, unlike MGSM, both Lite benchmarks consist of multiple-choice questions with option letters as final answers. Under this prompt, the model often outputs the answer directly and omits the public explanation, as we also observed when inspecting the raw responses. This phenomenon reflects a mismatch between the prompt structure and the multiple-choice task format, rather than necessarily indicating a limitation of ANCHOR To directly address your concern, we conducted an additional comparison. Specifically, we revised the prompt to explicitly require the model to provide a complete explanation in the question language before giving its final answer, while keeping the model and all other evaluation conditions unchanged. The results are presented below.

Results on Global-MMLU-Lite:

ModelAAMMpgp_gpenp_{en}pop_oEmpty-public rateAccuracy
ANCHOR: Original prompt2.469.534.3988.347.2797.2969.93
ANCHOR: Revised prompt88.028.213.0889.667.262.5166.98

Results on MMLU-ProX-Lite:

ModelAAMMpgp_gpenp_{en}pop_oEmpty-public rateAccuracy
ANCHOR: Original prompt2.077.285.1885.978.8597.3254.73
ANCHOR: Revised prompt75.966.362.0689.028.9114.7554.47

The revised prompt substantially increases the non-empty public-segment rate and improves A, showing that explicit instructions can elicit question-language explanations. Accuracy decreases on Global and changes little on ProX. We will report both the improvement in explanation coverage and the accuracy changes, and revise the cross-task conclusion accordingly: explicit explanation instructions improve public-segment output, but the accuracy–explanation trade-off varies across tasks.

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

Multilingual reasoning models often answer correctly while reasoning in English, and optimizing accuracy makes this worse. The authors build a probe that measures, at specific layers and token positions, whether the target-language signal outweighs the English one, and define the losing positions as a "recovery band." A light one-sided penalty (weight 0.01) is applied only inside that band, alongside RL. MGSM goes 84.15 => 86.07, target-language alignment rises sharply, and English/knowledge ability does not regress.

Critical strengths

Acc +1.92 accuracy gain alone could easily be a surface-format artifact. But the same objective applied as pure supervised learning loses 42.83 points, and adding RL recovers most of it — which rules out the format explanation and shows the penalty must be applied on-policy inside the RL loop. The authors also honestly report null generalization costs (all p > 0.2) and mark clearly which contrast straddles zero.

The probe is a real conceptual step. Prior work adds a language-consistency reward, this paper instead localizes where the language decision is made and intervenes only there, with intervention strength decaying as expected (active ratio 1.00 → 0.81). The probe is reusable independently of whether ANCHOR's numbers hold.

Critical weaknesses

Critically, there is no layer-band ablation. The paper never says which layers form the recovery band or whether a different band reproduces the effect. The entire mechanistic story rests on the band being the right place to intervene. So I cannot tell whether ANCHOR works because the probe found a real decision locus, or because a small penalty on any band happens to regularize language choice.

Questions for authors

Which layers form the recovery band, how were they chosen, and what happens at ±5 and ±10 layers? This is the load-bearing ablation and it is missing.

Is the 0.7 threshold coefficient stable across a sweep?

Do high-scoring traces code-switch in a way that satisfies fastText while staying conceptually English?

6 comments

We sincerely thank you for your time and constructive feedback on our work. We provide point-by-point responses to all your questions and concerns below.

Response to Q1: Which layers form the recovery band, how were they chosen, and what happens at ±5 and ±10 layers? This is the load-bearing ablation and it is missing.

The Recovery Band in Qwen3-8B-Instruct consists of L34–L36, where L0 denotes the embedding output and L36 is the final Transformer layer. As mentioned in the Introduction Section, we identified this range through a layer-wise probe of the initial model. For most eligible non-Latin-script languages, the Vocabulary Margin decreases in the middle and later layers and then recovers in the final three layers, which motivated our choice of intervention layers. We apologize for our insufficient description and promise to explicitly specify layer indices, indexing convention and selection procedure in the revision for better readability.

We sincerely thank you for pointing out the need for a more comprehensive ablation of the Recovery Band. To address this concern, we have conducted three groups of additional experiments: shifted bands outside the identified recovery region, a reduced band, and an expanded band. The only variable in all experiments is the selection of the anchor layer, and the rest of the training and inference settings are exactly the same as the Anchor.The first two groups have been completed, while the expanded band experiment is still running and is expected to finish within approximately three days. We will provide its results in the updated response.

ConfigurationenesfrdezhjaswbnruthteAVGΔAVG\Delta_{\mathrm{AVG}}A‾\overline{A}ΔA‾\Delta_{\overline{A}}M‾\overline{M}ΔM‾\Delta_{\overline{M}}bo
ANCHOR (L34–L36)96.0093.2086.4088.8088.4086.0062.0085.2090.8089.2080.8086.07—74.18—10.04—8.40
L36 only93.6090.0085.6085.2088.0083.6055.6084.0092.0086.8077.2083.78−2.2974.45+0.279.72−0.339.60
L29–L31 (shifted 5 layers earlier)94.4091.2087.2088.0088.0086.8056.8085.2093.2088.8080.4085.45−0.6269.61−4.577.84−2.206.40
L24–L26 (shifted 10 layers earlier)94.8091.2086.0088.0084.8084.4056.0084.4091.6089.2079.6084.55−1.5369.13−5.058.12−1.928.00

For the reduced-band setting, we retain only L36. For the shifted-band controls, we move the three-layer band five and ten layers earlier, resulting in L29–L31 and L24–L26, respectively. Compared with the original L34–L36 configuration, all three controls yield lower average accuracy, with decreases of 2.29, 0.62, and 1.53 points, respectively. The two shifted bands also exhibit notably lower A‾\overline{A} and M‾\overline{M}, while the L36-only setting achieves a slightly higher A‾\overline{A} with only a marginal difference, but shows a decrease in M‾\overline{M}. Overall, these results suggest that both the location and coverage of the constrained layers matter, and that the recovery band identified by our initial layer-wise analysis provides the best overall performance among the tested configurations. We promise to add this important experiment to the main text in the revised version.

Response to Q2: Is the 0.7 Threshold Coefficient Stable Across a Sweep?

ANCHOR uses α=0.7\alpha=0.7 as the default threshold coefficient, and our original ablation study already included α=0.5\alpha=0.5. To further examine its sensitivity, we conducted an additional experiment with α=0.9\alpha=0.9, keeping all other training settings unchanged.

ConfigenesfrdezhjaswbnruthteAVGΔAVG\Delta_{\mathrm{AVG}}A‾\overline{A}ΔA‾\Delta_{\overline{A}}M‾\overline{M}ΔM‾\Delta_{\overline{M}}bo
α=0.5\alpha=0.595.2092.0085.6086.0087.6084.4061.6084.4091.6090.0076.8085.02−1.0573.54−0.6410.11+0.077.20
α=0.7\alpha=0.7 (ANCHOR)96.0093.2086.4088.8088.4086.0062.0085.2090.8089.2080.8086.07—74.18—10.04—8.40
α=0.9\alpha=0.996.0092.0086.0087.6088.8085.6058.0085.2090.8089.2078.8085.27−0.8075.53+1.3610.04+0.005.20

The results show that performance is relatively stable across the three settings, with α=0.7\alpha=0.7 achieving the highest mean accuracy. While α=0.9\alpha=0.9 yields a higher A‾\overline{A}, this improvement does not translate into higher accuracy. We will include the new α=0.9\alpha=0.9 results in the ablation study of the revised manuscript.

Respons to Q3: Do high-scoring traces code-switch in a way that satisfies fastText while staying conceptually English?

We inspected ANCHOR’s original outputs and found examples in which the question language carries calculations and verification, rather than merely restating the problem. Both examples below are taken from inside <think>: the German example calculates individual clothing costs, sums them, and checks the result; the Japanese example calculates the cost of crayons and subtracts it from the budget to find the change. Both final answers are correct. We will include complete traces of this kind in the revision so that the reasoning steps performed in the question language can be checked directly.

ExampleExcerpt from the reasoning segmentReasoning step
German, MGSM sample 26“Zuerst die Shorts: 3 * 16,50. 16,50 mal 3. 16 * 3 ist 48, und 0,50 * 3 ist 1,50. Also 48 + 1,50 = 49,50 Dollar für die Shorts.” “Zuerst 49,50 + 67,50. 49 + 67 ist 116, und 0,50 + 0,50 ist 1,00. Also 116 + 1 = 117 Dollar. Dann 117 + 126 = 243 Dollar insgesamt.”The first excerpt breaks down the cost of the shorts; the second adds the costs of all three types of clothing. German carries the arithmetic decomposition and total-cost calculation, leading to the correct answer.
Japanese, MGSM sample 194“なので、5本買うと、合計の金額は5かける2で$10になります。それに対して、彼女は$20用意しているので、おつりは20から10を引けばいいはず。20-10=10。だからおつりは$10です。”The excerpt derives the cost of the crayons, relates it to the budget, and explains the subtraction used to calculate the change. Japanese carries the inference from the given information to the result.

The excerpts retain the original wording and arithmetic. We will add this case study to the appendix.

Comments on the paper

1 comment
Area_Chair_bdiLOpenReview ↗

Please engage in the discussion

Reviewers: if you have not responded to the authors' reply, please read it and engage in the discussion. Thanks!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.