finding the right context, and closing the gap between languages
Two problems that look separate but rhyme: giving a model the right context, and giving every language an equally good version of it.
Adaptive table retrieval
Fixed top-k retrieval is wrong in both directions — it drags in noise when one table would do, and truncates when the answer is spread across many. Adaptive Table Retrieval(Kim et al., 2026) adapts how much it retrieves to how much is actually relevant. Accepted to Findings of ACL 2026.
Cross-lingual RLVR data synthesis
RLVR needs verifiable tasks, and most domain text does not come with them. Code-switched task synthesis(Park* et al., 2026) turns unverifiable domain text into verifiable cross-lingual tasks, so capability earned in one language transfers rather than stopping at the language boundary. Accepted to EMNLP 2026.
This began as domain-specific data synthesis work during my NAVER Cloud residency, aimed at the multilingual performance gap, and fed into the HyperCLOVA X line (NAVER Cloud HyperCLOVA X Team, 2026).
The shared thread is sufficiency: retrieve enough but not too much, and make sure the training signal actually reaches the language you care about.
@inproceedings{kim2026retrieve,title={Retrieve Only Relevant Tables Whether Few or Many: Adaptive Table Retrieval Method},author={Kim, Taehee and Yang, Seungbin and Kim, Jihwan and Choo, Jaegul},booktitle={Findings of the Association for Computational Linguistics: ACL 2026},year={2026},}
@inproceedings{park2026unverifiable,title={From Unverifiable Domain Text to Cross-Lingual RLVR via Code-Switched Task Synthesis},author={Park, Cheonbok and Kim, Jihwan and Joo, Shinyoung and Lee, Joosung and Jo, Hwiyeol and Kim, Jeonghoon and Choo, Jaegul},booktitle={Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP)},year={2026},}
@article{hyperclovax2026think,title={HyperCLOVA X 32B Think},author={{NAVER Cloud HyperCLOVA X Team}},journal={arXiv preprint arXiv:2601.03286},year={2026},}