Pith. sign in

REVIEW 5 major objections 4 minor 12 references

Multilingual LLM evaluators score semantically identical content differently across languages, favoring lower-resource languages, and this bias is invisible to pairwise accuracy yet changes threshold-based decisions.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Multilingual LLM evaluators systematically inflate scores for lower-resource languages, and the standard pairwise-accuracy metric cannot detect the resulting safety-threshold disparities.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Worth reading and refereeing: the pairwise-blindness finding is real, but the strongest version of the claim rests on rejected responses alone, so the paper needs a chosen-response control before the abstract's generalization can stand. the 5 major comments →

arxiv 2607.14480 v2 pith:KERZ3COJ submitted 2026-07-16 cs.CL

Lower-Resource, Higher Scores: Language Bias in LLM Evaluators

classification cs.CL
keywords LLM evaluatorsreward modelsLLM-as-a-judgelanguage biasmultilingual evaluationpairwise accuracyuncertainty calibrationsafety thresholds
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

LLM evaluators—reward models and prompted judges—are usually trusted because they rank preferred over rejected responses with high pairwise accuracy. This paper argues that pairwise accuracy is the wrong lens for multilingual use: evaluators assign systematically different pointwise scores to semantically identical content depending on the evaluation language, and the direction is counterintuitive—lower-resource languages are scored more generously. Across eight open-weight evaluators and two frontier judges, the bias is consistent, statistically significant, and large enough that under a single global threshold acceptance rates differ by up to 43 percentage points across languages, meaning harmful content in lower-resource languages is more likely to pass safety filters. The paper also shows the bias correlates with model uncertainty but is not explained by it: language identity remains predictive after controlling for uncertainty, and instance-level difficulty does not account for the effect. A sympathetic reader would take away that multilingual evaluator validation must include pointwise, per-language consistency checks, not just ranking accuracy.

Core claim

The paper establishes that multilingual LLM evaluators—both prompted judges and trained reward models—judge semantically identical content on language-dependent scales: lower-resource languages receive consistently higher pointwise scores. The bias is large (roughly 0.5 points on a 1–5 scale), consistent across evaluator families and training paradigms, and persists in frontier models, and it correlates with web-resource availability (Spearman rho about −0.81 for reward models). Crucially, pairwise accuracy remains above 90% and stable, so standard validation cannot see the bias; under a fixed global threshold, acceptance rates differ by up to 43 percentage points across languages. The paper

What carries the argument

Semantically parallel, professionally translated instruction–response pairs across 23 languages serve as the controlled stimulus: the same content in each language, so any score difference is attributable to language. The central decomposition is score = content difficulty + β·NLL (negative log-likelihood uncertainty) + language-specific baseline + noise, used to separate uncertainty from language identity. A global-threshold acceptance-rate analysis exposes the practical blind spot that pairwise accuracy misses, and a code-switching demonstration shows how per-language thresholds can be attacked.

Load-bearing premise

All pointwise analyses use only the rejected (unchosen) responses; the paper generalizes from those to safety filtering and RLHF without showing that preferred responses show the same language-dependent score shifts.

What would settle it

Compute per-language mean scores on the preferred responses of the same parallel benchmark. If the language-dependent shift disappears or reverses there, the paper's central generalization—that threshold-based safety filtering is skewed across languages—would not hold for the chosen-response regime.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Pairwise accuracy above 90% is insufficient evidence that a multilingual evaluator is language-neutral; pointwise score distributions must be compared across languages.
  • Under a fixed global threshold, lower-resource languages can see acceptance rates tens of percentage points higher, so safety filters calibrated on high-resource languages can let harmful content through in other languages.
  • Per-language thresholds are only as safe as the language-identification step: code-switched prompts can be mislabelled and judged against a lenient threshold, raising acceptance from 50% to 75% in the paper's demonstration.
  • A per-language additive offset correction removes all cross-language variance in mean scores and cuts the acceptance-rate gap by 60.9% on average, but residual dispersion and language-by-item interactions remain.
  • The bias is driven partly by uncertainty but has a structural language-level component, so training-time calibration with multilingual data is needed, not just post-hoc score shifts.
  • Multilingual RLHF should condition reward computation on language or explicitly penalize cross-lingual score gaps; otherwise policies may learn to shift languages to harvest higher rewards.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the language-level prior is a stable property of an evaluator, multilingual RLHF should condition reward computation on language or explicitly penalize cross-lingual score gaps; otherwise policies may learn to shift languages to harvest higher rewards. This is my inference, not a claim the paper tests.
  • The code-switch attack likely has a larger surface than the paper's single-wrapper example: nested or multi-language code-switching could defeat stronger language-identification systems, so the 75% acceptance rate is probably a lower bound. My inference.
  • The paper uses web-resource availability as its resource proxy; a sharper test would correlate the bias with the language distribution in each model's own pretraining corpus, which may explain why some models show weaker resource correlations. My inference.
  • The paper's evidence is built entirely on rejected responses; I would not assume the safety-filtering implications transfer to chosen responses until per-language scores on preferred responses are measured. My inference.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper studies multilingual LLM evaluators (four prompted judges and four reward models) on semantically parallel instruction-response pairs from RewardBench/M-RewardBench across 23 languages. It reports that pointwise scores assigned to the same content shift with the evaluation language: lower-resource languages receive systematically higher scores, with a strong monotonic resource correlation for reward models (Spearman rho = -0.81). These shifts are invisible to pairwise accuracy, which stays above 90%, yet produce acceptance-rate differences up to 43 percentage points under a global threshold, including for safety-relevant content; code-switched prompts defeat per-language thresholds. The paper further investigates the mechanism, finding that summed NLL and other uncertainty measures correlate with the language-level score shift, but language identity remains a significant predictor after controlling for uncertainty, and within-language NLL-score slopes are inconsistent across domains.

Significance. The finding, if robust, challenges standard practice of validating multilingual evaluators by pairwise accuracy alone and has direct safety and RLHF implications. The paper uses public benchmarks, multiple open-weight models with different architectures and training paradigms, and offers a concrete code-switching attack demonstration. The resource-level correlation is a falsifiable, cross-model pattern, and the uncertainty analyses use multiple operationalizations with explicit appendices. These are substantial strengths that make the paper potentially important for evaluation methodology and multilingual safety.

major comments (5)
  1. [§3.3.1 / Appendix A] All pointwise score analyses—mean shifts, resource correlations, acceptance-rate gaps, and the NLL/uncertainty analyses—are computed exclusively on unchosen (rejected) responses, as stated in Appendix A: "we take prompt and unchosen as instruction-response pairs for our pointwise score analyses." The abstract and §1 state an unqualified claim that "multilingual evaluators assign significantly different scores to different evaluation languages." The chosen-response regime is central to standard evaluation and RLHF reward modeling; if the language-dependent shift is concentrated in rejected responses, the broad claim and the proposed per-language offset correction are overgeneralized. A pointwise analysis of chosen responses (or a careful restriction of the claim and downstream implications to rejected-content filtering) is needed as a load-bearing control.
  2. [§3.4 / Figure 4] The headline acceptance-rate disparity (43.0 pp max-min, and 34.0 pp in the >95% pairwise-accuracy regime) is reported without confidence intervals or per-language sample sizes. Thresholds are calibrated on a held-out set and language sample sizes are finite, so the disparity could partially reflect sampling noise. Report bootstrap CIs for the acceptance-rate gap and per-language acceptance rates, and state per-language n. This is necessary to support the claim that pairwise-validated evaluators are 'structurally blind' to decision bias.
  3. [Appendix D / Table 13] The claim that the additive correction eliminates 100% of cross-language variance is tautological: Eq. (7) defines b-hat(lang) = mu_lang - mu_global from the same data it then corrects, so subtracting per-language means by construction zeroes between-language variance. The substantive result is the 60.9% reduction in the acceptance-rate gap, which is meaningful, but the '100% elimination' and the wording 'confirming that differences ... are perfectly captured by a single scalar offset' are not empirical. Reframe as a decomposition identity or validate the correction on held-out languages/items.
  4. [§4 / Eq. (1)] The summed-NLL uncertainty measure is proportional to response length, and semantically parallel translations can differ in length across languages. The claim that higher summed NLL means "the model finds the language harder, not the material" (Section 4) requires demonstrating that cross-language length differences do not drive the NLL-score association. Report per-language mean token/character lengths, or include length as a covariate in the item-level regressions (Eq. 3-5), before accepting the uncertainty mechanism.
  5. [§3.3.1 / Table 6] Because M-RewardBench provides semantically identical items in every language, language is a within-item factor. The reported one-way ANOVAs treat observations as independent across languages, ignoring item-level pairing. This can inflate significance and does not isolate a language effect from content-by-language interaction. Use repeated-measures ANOVA or a mixed-effects model with item as a random effect, or at least paired tests, to support the 'statistically significant and consistent' claim.
minor comments (4)
  1. [Table 7] The caption says pairwise accuracy has 'minimal variation (≤5%)' within each evaluator, but the table reports a standard deviation of 8.0 for one evaluator column. Reconcile the caption with the reported values.
  2. [Table 4] The table caption/header uses 'Unbabel' for a model that is elsewhere called M-Prometheus; standardize model names across all tables and figures.
  3. [§3.4] The text reports a 44 percentage point gap between English and Ukrainian in the Skywork-LLaMA example, while the aggregate max-min gap is reported as 43.0% earlier in the same section. Please check consistency.
  4. [Appendix D / Eq. (7)] The notation b-hat(lang) is introduced in Appendix D but the main-text decomposition in Eq. (2) uses b(lang). Define the relationship explicitly to avoid confusion in the mechanism discussion.

Circularity Check

1 steps flagged

Central claim is empirical and not circular; only the Appendix D additive-bias '100% elimination' reduces to a definitional identity.

specific steps
  1. self definitional [Appendix D, 'Experiment: additive-bias decomposition' (Eq. 7 and Table 13)]
    "For each reward model–benchmark combination, we estimate b̂(ℓ) as the deviation of the per-language mean from the global mean: b̂(ℓ) = μℓ − μ_global ... We then subtract b̂(ℓ) from every score ... First, the additive correction eliminates 100% of the cross-language variance in every run, confirming that differences in per-language mean scores are perfectly captured by a single scalar offset b̂(ℓ) per language."

    The estimator b̂(ℓ) is defined as exactly the per-language mean deviation from the global mean. Subtracting a group's own mean from every observation mechanically sets all group means to zero, so the reported '100% elimination of cross-language variance' is an algebraic identity rather than an empirical confirmation of the additive-bias model. The accompanying 60.9% acceptance-rate-gap reduction is a substantive empirical result, which limits the circularity, but the 'perfectly captured by a single scalar offset' claim is a restatement of how the offset was constructed.

full rationale

The paper's central derivation is not circular. It evaluates public reward models and prompted judges on externally constructed, human-validated parallel benchmarks (RewardBench/M-RewardBench), comparing semantically identical content across 23 languages. The language-dependent pointwise shifts, resource-level correlations, pairwise-accuracy blindness, and uncertainty analyses are all measured from model outputs rather than derived from the paper's own assumptions, so they constitute independent empirical evidence. The appendix's additive-bias correction is the one step that reduces by construction: b̂(ℓ) is fit as the per-language mean shift and then subtracted, making '100% variance reduction' a tautology. However, this appears only in a supporting appendix and does not drive the main claims; the acceptance-rate gap reduction is a real empirical consequence, and the central findings would stand without the '100%' framing. The paper's reliance on unchosen/rejected responses for all pointwise analyses is a genuine scope limitation—generalization to chosen responses is not demonstrated—but it is an external-validity concern, not a circularity. Self-citations to the authors' prior work (e.g., Zhou et al. 2025; Resck et al. 2025, 2026; Hui et al. 2024a,b, 2025, 2026; Occhini et al. 2026) are contextual and not load-bearing; no uniqueness theorem or ansatz is imported from them to force the conclusions. Overall, the core contribution is self-contained against external benchmarks, with one localized definitional step in Appendix D.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The paper introduces no new entities. Its empirical claims depend on standard benchmark assumptions, the resource-level proxy, and the uncertainty operationalization. The only fitted quantity with a tautological flavor is the per-language offset in Appendix D; the global thresholds are standard but data-derived.

free parameters (2)
  • Per-language additive offset b-hat(lang) = mu_lang - mu_global for each language in each model-benchmark split
    Estimated in Appendix D Eq. 7 from the same data it is then removed from; makes the '100% variance reduction' an identity rather than a discovery.
  • Global decision threshold (50th percentile) = per model-category split, calibrated on a held-out set
    Used in Section 3.4 acceptance-rate analysis; threshold is data-derived and could affect the magnitude of the reported acceptance-rate gaps.
axioms (4)
  • domain assumption M-RewardBench is professionally human-translated and human-validated, so content is semantically identical across 23 languages.
    Section 3.1. If translations diverge semantically, score differences could reflect content rather than language.
  • domain assumption CC-MAIN page distribution percentage is a valid proxy for language resource level.
    Section 3.3.3, Figure 3. The 'lower-resource languages are scored more generously' conclusion depends on this proxy; a different resource ranking could change the correlation.
  • ad hoc to paper Summed NLL is a valid measure of model uncertainty; the token-length confound is handled by arguing numerator and denominator increase together.
    Section 4, Equation 1. If summed NLL mostly reflects tokenizer segmentation rather than true uncertainty, the uncertainty mechanism is not identified as cleanly as claimed.
  • domain assumption Pointwise scores on unchosen responses are the appropriate population for studying evaluator bias and threshold decisions.
    Appendix A, Figure 8. The selection of unchosen responses to avoid ceiling effects is not justified for all deployment settings, and no chosen-response analysis is reported.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Lower-Resource, Higher Scores: Language Bias in LLM Evaluators." pith.science (2026). https://pith.science/paper/KERZ3COJ

@misc{pith2026260714480,
  author       = {Pith},
  title        = {Pith review of: Lower-Resource, Higher Scores: Language Bias in LLM Evaluators},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KERZ3COJ}},
  note         = {Machine review of arXiv:2607.14480}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

LLM evaluators (trained reward models and prompted LLM-as-a-Judge) are routinely validated via pairwise accuracy. In a multilingual setting, this operates under the premise that high pairwise accuracy implies reliable, language-neutral scoring. We show that this assumption does not hold. We conduct experiments with semantically identical instruction-response pairs across 23 languages, and find that multilingual evaluators assign significantly different scores to different evaluation languages. The bias is statistically significant and consistent across eight open-weight evaluators of different architectures and training paradigms, persists in frontier judges, and is strongly correlated with language resource level: lower-resource languages are scored more generously. Meanwhile, these biases are invisible to pairwise accuracy: evaluators achieve above 90% pairwise accuracy, yet have up to 43% difference in acceptance rate across languages under a global decision threshold, meaning, for instance, that harmful content in lower-resource languages is more likely to pass safety filters. Per-language thresholds would require language identification, which can be defeated by code-switched prompts. We then investigate why lower-resource languages receive higher rather than lower scores, and we find that model uncertainty is linked with the effect: models tend to give higher scores when less confident, both under negative log-likelihood and under token-free uncertainty measures; however, language identity remains a significant predictor after controlling for uncertainty, and the bias cannot be explained away by content difficulty alone, but is a structural, language-level misalignment.

Figures

Figures reproduced from arXiv: 2607.14480 by Anna Korhonen, Ej Zhou, Lucas Resck, Zheng Hui.

Figure 1
Figure 1. Figure 1: Scoring distribution across languages. Left: mean pointwise scores assigned by prompted LLM-AS-A-JUDGE models (1–5 scale), averaged across evaluators; right: mean z-normalized scores from trained multilingual reward models. Error bars indicate inter-model variability [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Cross-model correlations of language-dependent scoring patterns. Each cell shows the Pearson correlation between per-language mean scores assigned by two evaluators, high correlation cells highlighted. Left: correlations aggregated across all benchmarks. Right: correlations computed on Chat-Hard, the subset with the highest correlation. 10 0 10 1 10 2 10 3 10 4 CC-MAIN-2025-51 pages (millions, log scale) 3… view at source ↗
Figure 3
Figure 3. Figure 3: Relationship between language evaluation scores and training data availability. Each [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-language pairwise accuracy (x-axis) vs. acceptance rate (y-axis) under a fixed global [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Relationship between model uncertainty (total NLL) and reward score. Left: language￾level, where each point is one of 23 languages with NLL and score averaged across all reward models and benchmark categories. Right: item-level, where every individual instance is plotted per model × task, pooled across all four reward models; color indicates evaluation language. For visualization, the right panel clips the… view at source ↗
Figure 6
Figure 6. Figure 6: Stacked variance decomposition of reward model scores into three components: variance [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Distribution of within-language NLL→Score slopes (βℓ ) across 4 models × 4 domains. Each dot is one model–domain pair; the x-axis shows the mean standardized slope. To test this, we fit a separate regression within each of the 23 languages: Scoreiℓ = αℓ + βℓ NLLiℓ + εiℓ , (6) and examine the distribution of per-language slopes βℓ across models and domains. If uncertainty were the driver, slopes should be c… view at source ↗
Figure 8
Figure 8. Figure 8: Distribution of reward scores for unchosen responses across benchmark domains. Stacked bars show the count of instances assigned to each discrete score level by different reward models. Unchosen responses exhibit wider score dispersion than chosen responses, reducing ceiling effects and revealing language-dependent variation. Model Scores Task Averages Lang Chat Chat-Hard Reasoning Safety C H R S O Aya Q M… view at source ↗
Figure 9
Figure 9. Figure 9: Correlation of language-dependent scoring patterns across judge models, aggregated over [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Between-language (orange) vs. within-language (blue) NLL [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

12 extracted references · 7 linked inside Pith

  1. [1]

    Output a score that is an integer between 1 and 5

  2. [2]

    SCORE: (an integer number between 1 and 5)

    The output format should look as: "SCORE: (an integer number between 1 and 5)"

  3. [3]

    role": "user

    Please do not generate any other opening, closing, or explanations. ###The instruction to evaluate: {prompt} ###Response to evaluate: {answer} For each of the remaining 22 languages (ar, cs, de, el, es, fa, fr, he, hi, id, it, ja, ko, nl, pl, pt, ro, ru, tr, uk, vi, zh), both the system prompt and user prompt are translated into the target language while ...

  4. [6]

    Zheng Hui, Zhaoxiao Guo, Hang Zhao, Juanyong Duan, Lin Ai, Yinheng Li, Julia Hirschberg, and Congrui Huang

    URL https://proceedings.neurips.cc/paper_files/paper/2023/ file/e425b75bac5742a008d643826428787c-Paper-Conference.pdf. Zheng Hui, Zhaoxiao Guo, Hang Zhao, Juanyong Duan, Lin Ai, Yinheng Li, Julia Hirschberg, and Congrui Huang. Toxilab: How well do open-source llms generate synthetic toxicity data?arXiv preprint arXiv:2411.15175, 2024a. Zheng Hui, Zhaoxiao...

  5. [7]

    José Pombal, Dongkeun Yoon, Patrick Fernandes, Ian Wu, Seungone Kim, Ricardo Rei, Graham Neubig, and André F

    URLhttps://arxiv.org/abs/2305.15425. José Pombal, Dongkeun Yoon, Patrick Fernandes, Ian Wu, Seungone Kim, Ricardo Rei, Graham Neubig, and André F. T. Martins. M-prometheus: A suite of open multilingual llm judges, 2025. URLhttps://arxiv.org/abs/2504.04953. Lucas Resck, Isabelle Augenstein, and Anna Korhonen. Explainability and Interpretability of Multi- l...

  6. [8]

    doi: 10.18653/v1/2024.naacl-long.301

    Association for Computational Linguistics. doi: 10.18653/v1/2024.naacl-long.301. URL https://aclanthology.org/2024.naacl-long.301/. Phillip Rust, Jonas Pfeiffer, Ivan Vuli´c, Sebastian Ruder, and Iryna Gurevych. How good is your tokenizer? on the monolingual performance of multilingual language models. In Chengqing Zong, Fei Xia, Wenjie Li, and Roberto Na...

  7. [9]

    doi: 10.18653/v1/2021.acl-long.243

    Association for Computational Linguistics. doi: 10.18653/v1/2021.acl-long.243. URL https://aclanthology.org/2021.acl-long.243/. Rickard Stureborg, Dimitris Alikaniotis, and Yoshi Suhara. Large language models are inconsistent and biased evaluators, 2024. URLhttps://arxiv.org/abs/2405.01724. Peiyi Wang, Lei Li, Liang Chen, Zefan Cai, Dawei Zhu, Binghuai Li...

  8. [587]

    Yarin Gal and Zoubin Ghahramani

    URLhttps://aclanthology.org/2025.findings-emnlp.587/. Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Maria Florina Balcan and Kilian Q. Weinberger (eds.),Proceedings of The 33rd International Conference on Machine Learning, volume 48 ofProceedings of Machine 11 Learning Research, p...

  9. [2021]

    doi: 10.18653/v1/2021.acl-long.131

    Association for Computational Linguistics. doi: 10.18653/v1/2021.acl-long.131. URL https://aclanthology.org/2021.acl-long.131/. Xiyan Fu and Wei Liu. How reliable is multilingual LLM-as-a-judge? In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng (eds.),Findings of the Association for Compu- tational Linguistics: EMNLP 2025, ...

  10. [2022]

    doi: 10.18653/v1/2022.acl-long.376

    Association for Computational Linguistics. doi: 10.18653/v1/2022.acl-long.376. URL https://aclanthology.org/2022.acl-long.376/. Paul Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences, 2023. URL https://arxiv.org/abs/ 1706.03741. Juntao Dai, Xuehai Pan, Ruiyang Sun, Jiaming ...

  11. [2023]

    doi: 10.18653/v1/2023.emnlp-main.614

    Association for Computational Linguistics. doi: 10.18653/v1/2023.emnlp-main.614. URL https://aclanthology.org/2023.emnlp-main.614/. Kabir Ahuja, Harshita Diddee, Rishav Hada, Millicent Ochieng, Krithika Ramesh, Prachi Jain, Akshay Nambi, Tanuja Ganu, Sameer Segal, Mohamed Ahmed, Kalika Bali, and Sunayana Sitaram. MEGA: Multilingual evaluation of generativ...

  12. [2024]

    URLhttps://arxiv.org/abs/2412.04261. A. Seza Do˘gruöz, Sunayana Sitaram, Barbara E. Bullock, and Almeida Jacqueline Toribio. A survey of code-switching: Linguistic and social perspectives for language technologies. In Chengqing Zong, Fei Xia, Wenjie Li, and Roberto Navigli (eds.),Proceedings of the 59th Annual Meeting of the Association for Computational ...

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.