acceptodds
Under review as a conference paper at ICLR 2027

Shared Computation Is Not Shared Access: Diagnosing and Improving Multilingual Reasoning

Abstract

Multilingual models can solve a problem in one language yet fail on the same problem in another. Final-answer accuracy leaves unclear whether the difficulty lies in executing calculations or in identifying the required quantities and operations and using intermediate results. To distinguish these difficulties, we develop a controlled framework that varies cross-language correspondence and the information supplied to the model. We measure autonomous solving from the question, conditional computation of a supplied operation, and continued solving from a correct partial solution. Experiments with synthetic languages, English–Chinese tasks, and two open-source model families show that accurate local computation can coexist with unreliable autonomous solving. Further diagnostics reveal difficulties in retrieving numerical inputs and completing solutions beyond a correct start. Guided by these diagnostics, we train Qwen3.5-4B first on ordered arithmetic steps extracted from natural-language solutions and then on full target-language chain-of-thought. Compared with chain-of-thought training in both stages, this route achieves higher mean autonomous accuracy under matched source exposure and training updates. These findings motivate evaluating and training multilingual models for both accurate computation and the ability to build complete solutions from a question.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multilingual reasoning, Cross-lingual transfer, Large language model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_FV82OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper explains why LLMs failed to solve the math problems in multilingual settings. It found that LLMs can execute supplied calculations correctly but have difficulty retrieving numerical inputs and subsequent solutions. Inspired by these findings, this paper used a two-stage pipeline, AccessPath-then-TargetLanguage SFT, to improve multilingual math ability.

Critical strengths
  1. Beyond Acc, this paper analyses the reason why LLMs fail in multilingual reasoning. The distinction among autonomous solving, conditional computation, and continued solving provides a more informative diagnosis than final-answer accuracy alone.
  2. Compared with Full-CoT training in both stages, this approach improves mean MGSM accuracy by approximately 3 pp in Swahili and Bengali under matched data order, source exposure, and training updates.
Critical weaknesses
  1. The controlled models are trained from scratch with source-language solutions but only target-language questions. Why only target-language questions? These designs are useful for diagnosis but differ substantially from realistic multilingual training.
  2. The three synthetic configurations also show highly variable and non-monotonic autonomous accuracy, from 10.55% to 89.55% under full lexical sharing (§3–4; Figure 3).
  3. The mean gains are only 2.93 and 3.07 points, and the 95% seed-level intervals in Figure 7 cross zero. The post-training diagnostic intervals also include zero. The improvement is not so solid.
Questions for authors

See above.

3 comments
AuthorsOpenReview ↗

Response to Reviewer FV82 [Part.1]

Thank you for recognizing the value of separating arithmetic execution from autonomous solving. We clarify the purpose of the controlled setting and the configuration variability, and report the expanded training comparison below.

W1: Why train on target-language questions without paired solutions when this setting differs substantially from realistic multilingual training?

The controlled setting tests whether source-language solution supervision transfers without paired target-language solutions. Providing such pairs would introduce direct task supervision in the target language, making successful solving harder to attribute to transfer. The English–Chinese models receive Chinese questions and independent solution-format text, providing exposure to Chinese and its solution format while withholding question–solution pairing.

The Chinese-supervised calibration confirms that the tasks and formats are learnable with paired supervision: all five calibration models achieve at least 98.24% autonomous accuracy on both Chinese question forms (Table 5). Section 5 then examines the same diagnostic distinction in pretrained Qwen and Phi models, extending the analysis beyond the controlled supervision setting.

W2: Autonomous accuracy varies substantially across the three synthetic configurations and changes non-monotonically with lexical sharing, ranging from 10.55% to 89.55% at full sharing.

Full lexical sharing does not guarantee successful autonomous transfer in these experiments. The three configurations are independently constructed settings that differ in language forms, sharing rankings, task instances, and training randomness. Across these settings, autonomous accuracy varies substantially, while conditional computation and continued solving remain near ceiling.

Table F1. Target-language accuracy at full lexical sharing (%), reproduced from Table 3.

MeasurementConfiguration 1Configuration 2Configuration 3
Autonomous solving78.4289.5510.55
Conditional computation100.00100.0099.83
Continued solving99.8099.7998.95
Value retrieval82.6293.5517.19

Value retrieval identifies a specific difficulty in Configuration 3: accuracy is 99.61% in the source language but only 17.19% in the target language. Thus, poor recovery of stated inputs coexists with accurate computation from supplied states and successful continuation from a correct prefix.

Within each sharing sweep, the source corpus, sample order, tokenizer, initialization, and optimizer settings are fixed (Appendix B.1). Full lexical sharing still leaves source–target word-order differences. Configuration 3 also exhibits a non-monotonic response: autonomous accuracy falls from 48.83% at 50% sharing to 10.55% at full sharing (Table 2). These runs do not isolate the cause of this reversal. We will clarify that the supported finding is the coexistence of accurate supplied computation and continuation with poor input retrieval and autonomous solving, even under full lexical sharing.

AuthorsOpenReview ↗

Response to Reviewer FV82 [Part.2]

W3: The mean training gains are only 2.93 and 3.07 percentage points, and the confidence intervals for both these gains and the post-training diagnostic differences include zero.

We extended both the main evaluation and the post-training diagnostics from three to eight paired training seeds, keeping the training recipe and fixed endpoint unchanged. The combined mean gains are 1.60 percentage points in Swahili and 2.70 in Bengali.

Table F2. MGSM results across eight paired seeds, with 250 questions per language and a 512-token limit. Differences and intervals are in percentage points.

LanguageFull-CoT→Full-CoTAccessPath→Full-CoTDifference95% paired-seed CI95% question-bootstrap CI
Swahili60.85%62.45%+1.60[−0.46, 3.66][−1.45, 4.55]
Bengali63.15%65.85%+2.70[0.63, 4.77][−0.15, 5.55]

Paired-seed CIs assess training variability on fixed questions. Question-bootstrap CIs use 10,000 paired resamples with the trained models fixed.

The expanded comparison retains positive mean gains in both languages. Bengali’s paired-seed interval now excludes zero, providing stronger evidence for its gain across training repetitions on the fixed test set. Swahili’s paired-seed interval and both question-bootstrap intervals include zero.

Table F3. Post-training differences on the same 165 Swahili and 164 Bengali diagnostic instances, across eight paired seeds. Values are AccessPath-route minus Full-CoT-route accuracy in percentage points, with 95% paired-seed intervals.

LanguageAutonomous solvingContinued solving
Swahili+3.18 [0.63, 5.73]+0.30 [−2.52, 3.13]
Bengali+1.37 [−1.96, 4.71]+0.84 [−1.57, 3.24]

The diagnostics also show higher mean autonomous accuracy: +3.18 points in Swahili and +1.37 in Bengali, with the Bengali effect still uncertain. Continued-solving differences are small and show no clear route advantage. The expanded results retain modest positive mean differences in autonomous solving, with the strength of the evidence varying by language and evaluation cohort.

AuthorsOpenReview ↗

Looking Forward to Your Feedback

Dear Reviewer FV82,

Thank you again for recognizing the value of our diagnostic approach and for raising questions about the controlled setting, configuration variability, and training gains. Our responses in Parts 1–2 address these points:

  • We explained why the controlled experiments withhold paired target-language solutions and discussed the Chinese-supervised calibration that confirms the tasks are learnable with paired supervision.
  • We clarified how the three synthetic configurations differ and examined value retrieval alongside autonomous solving to interpret their contrasting results (Table F1).
  • We expanded the training comparison and post-training diagnostics from three to eight paired seeds, reporting both the mean improvements and their uncertainty (Tables F2–F3).

As the discussion period draws to a close, we would appreciate your feedback on whether these responses address your concerns. We would be happy to provide further clarification on any remaining point.

Thank you again for your time and thoughtful review. Have a nice day!:)

Best regards,

Authors

Reviewer_CEjtOpenReview ↗
Rating
2
Confidence
3(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks why multilingual models solve a math problem in one language but fail on the same problem in another. They have three measurements: autonomous solving, conditional computation, and continued solving. They use synthetic language pairs with different lexical sharing, English-Chinese models trained from scratch, and Qwen3.5-4B / Phi-3.5-mini on GSM8K, and cases in Swahili, Bengali and Thai. The main finding of the paper is that local computation is accurate while autonomous solving is not. Then they propose two-stage training on Qwen3.5-4B: first on ordered arithmetic states from English solutions, then on full target-language CoT.

Critical strengths
  • Splitting final-answer accuracy into three measurements is a clean idea. The controlled settings are good.
  • The training comparison is well matched (same source problems, order, updates, replay, and the same Stage B data).
Critical weaknesses
  • The improvement of training is small, in my point of view. On Swahili and Bengali, the mean gain over Full-CoT is about 3 points (+2.93 and +3.07), with only 3 seeds and 95% CIs that include zero (Table 13). There is also a cost: English drops from 91.6% to about 83% for both routes, even with English replay. And the paper says something strange in the appendix: "This evaluation plan was specified after training and before the reported generation round; the test and development sets had already been used earlier in the project." So the MGSM test set was not really held out, and I am not sure the scores are reliable.
  • The diagnostics and the training don't really match. After training both routes are the same on conditional ( about 100%) and continued solving (73%), and in Bengali the AccessPath stage even reduces continued solving. In English, the correct prefix reduces accuracy and Phi hits the token cap in more than 50% of continued generations in Swahili/Bengali, so the prefix interface may be measuring format mismatch rather than "access". Also, simple alternatives are missing, like translated English CoT in Stage A, or just a longer Full-CoT stage, to show the gain is from arithmetic states and not from shorter targets.
Questions for authors
  • Can you explain or report more about the improvement and the evaluation, to show the results are reliable and useful for others?
  • Since conditional computation doesn't depend on the question (B.3 controls), what does it tell us beyond arithmetic ability?
10 comments
AuthorsOpenReview ↗

Response to Reviewer CEjt [Part.1]

Thank you for recognizing the diagnostic framework and matched training design. We have added training repetitions, prefix-interface checks, and the two requested training controls to address the reliability and interpretation of the results.

W1 & Q1: The training gains are small and based on only three seeds, with 95% confidence intervals that include zero (Table 13).

We extended the comparison from three to eight paired training seeds, keeping the training recipe, fixed Stage-B endpoint, and evaluation protocol unchanged. Across all eight paired seeds, AccessPath→Full-CoT exceeds Full-CoT→Full-CoT by 1.60 percentage points in Swahili and 2.70 in Bengali.

Table C1. MGSM accuracy across eight paired seeds, with 250 test questions per language and a 512-token generation limit. Differences and intervals are in percentage points.

LanguageFull-CoT→Full-CoTAccessPath→Full-CoTDifference95% paired-seed CI95% question-bootstrap CI
Swahili60.85%62.45%+1.60[−0.46, 3.66][−1.45, 4.55]
Bengali63.15%65.85%+2.70[0.63, 4.77][−0.15, 5.55]

Paired-seed CIs assess training variability on fixed questions. Question-bootstrap CIs use 10,000 paired resamples with the trained models fixed.

The additional runs retain positive mean gains in both languages. Bengali's paired-seed interval now excludes zero, while Swahili's and both question-bootstrap intervals still include zero. The evidence supports a modest, language-dependent average benefit; we will update Table 13 and the corresponding claims accordingly.

W1 & Q1: Finalizing the evaluation plan after training and previously using the test and development sets raise concerns about whether the MGSM evaluation was held out and whether its scores are reliable.

Thank you for carefully reading the appendix and identifying this ambiguity. MGSM test examples were not used in our fine-tuning, including English replay. Both target-language training examples and English replay came from the GSM8K training split.

The appendix states:

“This evaluation plan was specified after training and before the reported generation round; the test and development sets had already been used earlier in the project.”

“Specified after training” refers to finalizing the common evaluation protocol after Stage-B training, but before generating the reported outputs. We used the same questions, prompts, generation limits, and scoring rules across routes and stages because the two-stage experiment asks whether the AccessPath benefit persists after both routes receive the same Full-CoT training. The common protocol allows paired comparisons of first-stage and final checkpoints under consistent evaluation conditions.

“Used earlier” refers to previous evaluations for the training study in Section 6, including evaluations of first-stage checkpoints. MGSM was therefore a reused evaluation benchmark, rather than a previously untouched test set. The 320-update Stage-B endpoint was fixed before Stage-B training; it was not selected as the best checkpoint from the reported MGSM results. We will revise the appendix sentence to make this purpose and timeline explicit.

W1: English accuracy drops from 91.6% to about 83% for both training routes despite English replay.

Our training study compares arithmetic-state and Full-CoT supervision for improving target-language reasoning under matched adaptation conditions. We evaluate English retention to quantify the accompanying trade-off.

Table C2. English accuracy after adaptation, extending Table 15 to eight paired seeds. The base model scores 91.60% on the same evaluation.

Adaptation languageFull-CoT→Full-CoTAccessPath→Full-CoTDifference (pp; 95% paired-seed CI)
Swahili83.25%82.40%−0.85 [−2.88, 1.18]
Bengali83.50%82.90%−0.60 [−3.15, 1.95]

Both routes show English-accuracy losses of similar magnitude despite replay, which accounts for one in ten training examples (Appendix E.1). AccessPath is 0.60–0.85 percentage points lower on average, while its additional retention cost remains uncertain. The substantial decline therefore occurs with Full-CoT adaptation as well, rather than being unique to AccessPath.

More effective replay strategies may help mitigate these losses, although this requires further evaluation. Optimizing English retention is outside the scope of the present supervision comparison; we report this cost alongside the target-language gains so that the benefit and trade-off are both explicit.

AuthorsOpenReview ↗

Response to Reviewer CEjt [Part.2]

W2.1: The link between diagnosis and training is unclear because the two routes perform similarly on conditional computation and continued solving, while AccessPath performs worse than Full-CoT on Bengali continued solving after Stage A.

Our diagnosis distinguishes executing a supplied calculation from constructing a solution from the question. AccessPath supervises the selected quantities, operations, and their order through arithmetic trajectories. We test whether this Stage-A supervision improves autonomous solving after a common Full-CoT Stage B, compared with Full-CoT training in both stages.

Table C3. AccessPath-route minus Full-CoT-route accuracy after Stage B, across eight paired seeds on the same 165 Swahili and 164 Bengali diagnostic instances. Values are percentage-point differences with 95% paired-seed intervals.

LanguageAutonomous solvingContinued solving
Swahili+3.18 [0.63, 5.73]+0.30 [−2.52, 3.13]
Bengali+1.37 [−1.96, 4.71]+0.84 [−1.57, 3.24]

On this diagnostic cohort, the evidence for an autonomous-solving gain is clearer in Swahili, while the Bengali estimate remains uncertain. Conditional computation is already at 99.92–100% for both routes. Because it supplies the operation and inputs, it bypasses their selection from the question; similar conditional scores are therefore compatible with differences in autonomous solving.

At Stage A, the natural-language continuation prefix more closely matches Full-CoT supervision than AccessPath’s arithmetic-state chains. This alignment may contribute to the lower Bengali continuation accuracy for AccessPath in the original three-seed comparison: 69.11% versus 73.58% (Table 16). After the common Full-CoT stage, continued-solving differences are small and uncertain, while mean autonomous accuracy favors AccessPath. This pattern is consistent with the motivation for supervising question-to-trajectory construction. We will clarify this diagnosis–training link in Section 6.

W2.2: Correct prefixes reduce English accuracy, and Phi reaches the token cap in over 50% of Swahili and Bengali continuations, raising concerns that the interface measures format mismatch rather than access to computation.

We increased the original-prefix generation limit from 1,024 to 2,048 tokens and compared three interfaces at the 2,048-token limit: the original assistant prefix, the same prefix with the autonomous-solving user instruction, and a partial solution placed in the user message. The user-context variant explicitly requests continuation without restarting, changing both placement and wording.

Table C4a. English final-answer accuracy on the same 221 diagnostic cases, with a 2,048-token limit.

Input conditionQwen3.5-4BPhi-3.5-mini-instruct
Autonomous solving93.21%90.50%
Original assistant prefix85.97%69.68%
Assistant prefix with matched user instruction87.78%85.97%
Partial solution as user context91.86%88.24%

The English results show substantial sensitivity to the prefix interface, especially for Phi. Matching the user instruction raises Phi’s accuracy from 69.68% to 85.97%, and the user-context variant brings both models closer to their autonomous accuracy. Doubling the original-prefix limit provides no improvement: Qwen remains at 85.97%, while Phi changes from 70.14% to 69.68%.

We also examined how the generation limit and interface affect Phi’s target-language continuations.

Table C4b. Phi-3.5-mini-instruct continued-solving accuracy and cap-hit rates on 183 Swahili and 178 Bengali diagnostic instances. Cap-hit denotes reaching the generation-token limit.

LanguagePrefix interfaceToken limitAccuracyCap-hit rate
SwahiliOriginal1,02414.75%50.82%
SwahiliOriginal2,04814.75%50.27%
SwahiliMatched instruction2,04814.75%9.84%
BengaliOriginal1,02417.98%51.12%
BengaliOriginal2,04817.98%50.00%
BengaliMatched instruction2,04816.29%10.67%

The lower cap-hit rates do not translate into higher target-language accuracy under the tested conditions. Doubling the budget leaves accuracy unchanged, while matching the instruction reduces cap-hit rates from approximately 50% to 10% without improving accuracy in either language. We will include these controls and interpret continued-solving accuracy together with the prefix interface and cap-hit rate.

AuthorsOpenReview ↗

Response to Reviewer CEjt [Part.3]

W2.3: Controls using translated English CoT in Stage A and longer Full-CoT training are needed to distinguish the effect of arithmetic-state supervision from that of shorter targets.

We added both requested controls. The longer-training control increases Full-CoT Stage A from 320 to 640 updates. The translation control uses machine translations of complete English reference solutions for 320 updates, with the same training questions and order as the original Full-CoT control. All routes then receive the same target-language Full-CoT examples for 320 Stage-B updates.

Table C5. Final MGSM accuracy on 250 questions per language. All rows use the same three paired seeds; the longer Stage A deliberately adds source exposure and updates.

Stage-A targetStage-A updatesSwahiliBengali
Original Full-CoT32060.80%64.67%
AccessPath32063.73%67.73%
Original Full-CoT64056.93%60.80%
Translated complete English CoT32061.60%65.87%

AccessPath achieves the highest observed mean accuracy among the tested routes. Doubling Full-CoT Stage A lowers final accuracy under the tested schedule, while translated CoT improves the original baseline and narrows the gap. These results support ordered arithmetic trajectories as a useful Stage-A supervision target under the tested conditions.

Q2: Since conditional computation remains accurate when the question content is replaced (Appendix B.3), what does it reveal beyond arithmetic ability?

Conditional computation measures whether a model can execute a calculation when the operation and operands are supplied. Its diagnostic value comes from comparison with value retrieval and autonomous solving. In Configuration 3 at full lexical sharing, conditional accuracy reaches 99.83%, while value retrieval and autonomous accuracy are only 17.19% and 10.55%, respectively (Table 3).

The question-replacement result follows from this design: the supplied operation and operands already specify the calculation, so its execution does not require information from the question. Conditional computation therefore serves as an execution control. Its contrast with low retrieval and autonomous accuracy shows that the model can perform supplied calculations while struggling to recover the inputs and construct a solution from the question. This distinction is not visible from final-answer accuracy alone.

Reviewer_3kMoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Multilingual language models often solve a math problem correctly in one language but fail the same problem in another, and final-answer accuracy alone cannot reveal whether the breakdown is in doing the arithmetic or in identifying which calculations to perform and chaining them together. To separate these, the authors measure three things on identical problems (1) solving from the question alone, (2) computing a result when the operation and inputs are supplied, and (3) finishing a solution from a correct partial one. In controlled experiments with synthetic language pairs and English–Chinese renderings, models trained from scratch could compute supplied operations at 90–99% accuracy in the target language while autonomous solving ranged from near-zero to highly inconsistent, with further probes showing they struggled to retrieve numbers stated in the question and to carry a solution through to the end. The same pattern appears in Qwen3.5-4B and Phi-3.5-mini on parallel GSM8K problems.

Critical strengths

The problem setup is very smart and clever for studying this type of research question. The authors break the question down into three experiments: solving the question alone, computing a supplied operation, and continuing from a correct partial solution. This makes it possible to tell apart failures in arithmetic from failures in finding and connecting the required steps. It was interesting to read.

The paper contains thorough and well-controlled experiments, including training language models from scratch on synthetic language pairs, and evaluating Qwen3.5-4B and Phi-3.5-mini on parallel GSM8K questions. The experiments are well designed and easy to understand.

Critical weaknesses

The work could have benefitted from studies on bigger models. Probably scaling change how models behave?

Questions for authors

see weakness.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 3kMo

We thank the reviewer for the positive assessment of our diagnostic design and controlled experiments, and for suggesting an evaluation of larger models.

W1: The study would benefit from evaluating larger models to determine whether the diagnostic pattern changes with model size.

We extended the Section 5 diagnostics to Qwen3.5-9B and Qwen3.5-27B. All three checkpoints are evaluated on the same 716 instances using identical rendered prompts and greedy non-thinking decoding, with generation limits of 24 tokens for conditional computation and 1,024 tokens for autonomous and continued solving.

Table M1. Accuracy (%). Each cell reports Qwen3.5-4B → Qwen3.5-9B → Qwen3.5-27B.

LanguageCasesAutonomous solvingConditional computationContinued solving
English22192.76 → 93.67 → 95.48100.00 → 99.55 → 100.0085.97 → 89.14 → 95.48
Swahili18355.74 → 70.49 → 89.0799.45 → 99.45 → 100.0056.28 → 67.76 → 89.62
Bengali17851.12 → 60.11 → 79.2199.44 → 100.00 → 100.0056.18 → 65.17 → 83.71
Thai13444.78 → 71.64 → 82.0997.76 → 97.76 → 98.5153.73 → 63.43 → 83.58

Larger checkpoints substantially narrow the gap between conditional computation and autonomous solving, while the gap remains evident at 27B. From 4B to 27B, autonomous accuracy increases by 28.09–37.31 percentage points across the target languages, while conditional computation remains near ceiling. At 27B, the remaining gaps are 10.93 points in Swahili, 20.79 in Bengali, and 16.42 in Thai, with all three 95% paired-source bootstrap intervals above zero. Thus, near-perfect execution of supplied calculations still does not ensure equally accurate solving from the question. The diagnostic distinction remains informative at larger checkpoints, even as the magnitude of the gap decreases.

Continued-solving accuracy also improves across checkpoints. At 27B, however, the evaluated prefix interface provides no clear additional accuracy benefit. Its mean differences from autonomous solving are +0.55, +4.49, and +1.49 points in Swahili, Bengali, and Thai, respectively, with all three 95% paired intervals including zero. These results show how the diagnostic pattern changes across checkpoint sizes: the conditional–autonomous gap persists, while an additional benefit from a correct prefix is not established at 27B. We will add this comparison and its interpretation to Section 5.

AuthorsOpenReview ↗

Looking Forward to Your Feedback

Dear Reviewer 3kMo,

Thank you again for your positive assessment of our diagnostic framework and for suggesting experiments with larger models.

Following your suggestion, our response reports evaluations of Qwen3.5-9B and Qwen3.5-27B alongside the 4B model on the same 716 diagnostic instances under matched evaluation settings (Table M1). The larger checkpoints substantially improve autonomous solving and narrow the gap with conditional computation, while a gap of 10.93–20.79 percentage points remains across the three target languages at 27B.

As the discussion period draws to a close, we would appreciate your feedback on whether this comparison addresses your concern about how the diagnostic pattern changes with model size. We would be happy to clarify any remaining questions.

Thank you again for your time and consideration. Have a wonderful day! :)

Best regards, Authors

Thank you for your response. Since I already gave a positive assessment. I keep my score.

AuthorsOpenReview ↗

Thank You for Your Positive Assessment!

Dear Reviewer 3kMo,

Thank you for your response and positive assessment. We appreciate your thoughtful feedback, particularly your suggestion to examine larger models, which helped strengthen our study. We will incorporate the additional results and clarifications into the revised manuscript.

Thank you again for your time and effort in reviewing our work.

Have a nice day! :)

Best regards, Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

Dear AC and reviewers,

Thank you for your constructive comments and suggestions. We have incorporated the rebuttal experiments and clarifications into the revised manuscript. New and revised text is marked in blue. The main updates are:

  • Added 9B/27B evaluations and prefix-interface and generation-limit controls (Section 5, Appendix D).
  • Expanded the primary training comparisons and post-training diagnostics to eight paired seeds, with updated figures and confidence intervals (Section 6, Figures 7–8, Appendix E).
  • Added state-order, concise-language, translated-CoT, and longer-training controls, alongside English-retention results (Appendix E).
  • Corrected decimal-normalization scoring, with prompts, generated outputs, and evaluation cases unchanged (Appendix D).
  • Clarified benchmark reuse, evaluation protocols, and the scope and limitations of our conclusions.

We appreciate your time and suggestions. Have a nice day! :)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.