Commercially permissive multilingual SFT and mechanistic work on low-resource translation pull in different directions: one expands what can be trained on, the other asks what the model is actually using when Hindi becomes Urdu or Sindhi. Multi-SFT adds 502,818 instruction-response pairs across five European languages, with human agreement at 0.87 or higher, while TITAN studies translation circuits in multilingual decoder-only LLMs for Hindi→Urdu and Hindi→Sindhi and uses causal intervention plus sparse feature decomposition to target the relevant components.

Multi-SFT: Permissively-Licensed Multilingual Instruction-Following

Multi-SFT gives you a commercially permissive five-language instruction-tuning set of 502,818 pairs, and it can match larger mixed releases on 13 benchmarks while improving safety.

Supervised fine-tuning datasets remain overwhelmingly English-centric and commercially restricted, as their content is frequently generated by proprietary models whose terms of service forbid training competing systems. We introduce Multi-SFT, a commercially permissive collection of 502,818 instruction-response pairs across five European languages, built via translation and regeneration pipelines for tasks where translation is insufficient. Human evaluation yields inter-annotator agreement of at least 0.87 across all datasets and languages. Fine-tuning nine models from three different families (EuroLLM, OLMo 2, Qwen3; 1B-32B) and evaluating them on 13 benchmarks shows that Multi-SFT alone is competitive with larger augmented mixes and official instruction-tuned releases, with decisive advantages on safety tasks.

Translation Circuit Discovery and Targeted Optimization for Low-Resource Multilingual Translation

TITAN updates under 5% of parameters by targeting sparse translation circuits, yet still lifts BLEU by 33.75% on Hindi-Urdu and 45.86% on Hindi-Sindhi.

Recent multilingual large language models achieve strong translation performance, yet the internal mechanisms responsible for low-resource multilingual translation remain insufficiently understood. In this paper, we study translation circuits in multilingual decoder-only LLMs for Hindi-to-Urdu and Hindi-to-Sindhi (Devanagari). We propose TITAN, a mechanistic framework that combines causal intervention, sparse feature decomposition, and targeted preference optimization to identify and optimize crucial translation components in the transformer. Unlike conventional full-parameter fine-tuning, TITAN restricts optimization to sparse, causally important circuits, enabling scalable and parameter-efficient adaptation while updating fewer than 5% of model parameters. Experimental results show that TITAN achieves BLEU improvements of 33.75% on Hindi-to-Urdu and 45.86% on Hindi-to-Sindhi over the strongest multilingual baselines. Human expert evaluation further confirms substantial gains in adequacy and fluency across both language pairs. Additional analyses reveal that multilingual translation is implemented through sparse, structured circuits operating over a shared semantic subspace rather than through an English-centric pivot representation. Overall, our findings demonstrate that mechanistic circuit discovery can directly improve low-resource multilingual translation while providing interpretable insights into multilingual transformer behavior.

Papers announced by arXiv on Thursday, 15 October 2026, covering submissions from 14 October to 15 October 2026.