Multilingual retrievers bury correct foreign docs in shared index; training-free steering fixes it
_reachsumit · x · 2026-10-14
A paper shows multilingual dense retrievers retrieve substantially different documents for the same question in different languages—not because of weak language understanding, but because in a shared index each language's correct answers get buried under non-relevant same-language or English documents.
- Introduces disagreement-guided steering: compute a language-specific direction from the difference between cross-language consensus documents and that language's idiosyncratic retrievals, added to queries at inference
- Training-free, requiring no fine-tuning
- Most strongly reduces same-language crowding versus mean-removal, subspace-removal, and per-language-rotation baselines while staying competitive on recall
- Trades off against mean removal, which better preserves ranking quality
More from Research
- FrontierML benchmark tests whether agents can spend money to scale model performance — JayAlammar · 2026-10-14
- AIPOCH's Open-Science ships portable .science packages that share full research context, not just outputs — ahuja_priyank · 2026-10-14
- Video-Arena: Game-Based Video Benchmark Shows Native Models Score Just 13.2 F1 — Kuan Zhang · 2026-10-14
- LSCBench: Vibe Coding Agents Write Functional Code, But Only 25% Is Also Secure — Qwen · 2026-10-14
- Omni-LLMs Know More Than They Say: NAVER's OmniCatalyst Adds 7.29% via Self-Generated Text — naver-ai · 2026-10-14
- Repeated-Data Pretraining Study: Smaller Batches Act as Regularizers and Beat More GPUs — Sewon Min · 2026-10-14