IIT Delhi lands 4 NeurIPS 2026 main-track papers on multilingual interpretability and distillation
Tanmoy_Chak · x · 2026-09-27
Tanmoy Chakraborty's group at IIT Delhi announced 4 papers accepted to the NeurIPS 2026 main track, spanning multilingual LLM interpretability, knowledge distillation, and personalization:
- Interpretability: The Fractured Interlingua shows knowledge edits made in English with ROME/MEMIT/AlphaEdit miss 90% of the directions needed for Hindi/Urdu/Chinese, with a provable bound that 96% of target residual variance is language-private; aligning representations opens the key-space but not the value-space gate.
- Distillation: CaRE-KD uses a per-token gate switching between Forward and Reverse KL based on teacher-student confidence, plus a BALD-based Revival rule skipping updates when the teacher is uncertain and the student confident — modest but mostly consistent gains across 8 teacher-student pairs and 11 benchmarks, at notable compute cost.
- Personalization: recovering evolving user preference states via adaptive interaction-aware representation correction.
More from Research
- A 'new type of mind' trained without natural language writes kernels at up to 66x Triton speed — repligate · 2026-09-27
- ENGRAM and YOCO memory architectures could shift data center DRAM/HBM economics — AccBalanced · 2026-09-27
- From Pixels to Visual Tokens: How LLMs Actually 'See' Images — makaros622 · 2026-09-27
- Spatial-Interactor teaches VLMs spatial reasoning through physical interaction — arankomatsuzaki · 2026-09-27
- DeepMind researcher explains why AI hasn't transformed physics yet — DaniloJRezende · 2026-09-27
- WTF paper fine-tunes flow maps with optimal transport, cutting RL training compute up to 280x — yeewhye · 2026-09-27