Cross-language transfer of token-reference tables

Determine whether ReWEIGH’s token-reference table can transfer across languages or whether separate calibration is required for each language.

Background

All calibration, development, and evaluation in the paper are conducted in English, and token-specific references are constructed over each model’s native vocabulary. Differences in morphology, tokenization, candidate frequencies, and visual-evidence baselines may therefore affect multilingual behavior. The paper leaves unresolved whether an English-derived reference table can be transferred across languages or whether language-specific calibration is necessary.

References

Third, some Science and Geography reference answers use English transliterations, and we do not assess whether judges score these consistently against native Telugu equivalents.

— VākQA: A Benchmark and Evaluation Study for Telugu Spoken Factoid Question Answering  (2609.19879 - Akkiraju et al., 17 Sep 2026) in Section Conclusions

We could repeat calibration in another language, but we do not know whether a reference table can transfer across languages or requires separate calibration.

— ReWEIGH the Evidence: Calibrating Token-Level Ordinal Visual Evidence to Mitigate Hallucinations in Large Vision-Language Models  (2608.19075 - Jeong et al., 19 Aug 2026) in Limitations section

An existing English reference-free metric such as REFeREE could therefore be used as a reward on English data, and whether the resulting gains transfer to Catalan and Spanish without curated resources in those languages is a question these experiments motivate rather than answer.

— Reinforcement Learning for improving Large Language Models' Catalan text simplification capabilities  (2609.04823 - Domingo et al., 4 Sep 2026) in Section 5, Conclusions