MA-BC paper shows provably efficient multi-objective imitation by pooling non-conflicting expert data
Yossarian_1234 · reddit · 2026-10-08
The paper Split the Differences, Pool the Rest tackles learning from experts with different objectives. Naively pooling all demonstrations loses each expert's trade-offs, while per-expert learning forfeits data sharing. MA-BC pools demonstrations where observed actions don't disagree, with proven upper and lower bounds on sample complexity. By Ziyad Sheebaelhamd, Luca Viano, Volkan Cevher, and Claire Vernade; arXiv paper and code are public.
More from Research
- Psych researchers predict AI will break peer review and reshape social science publishing — steverathje2 · 2026-10-08
- Yaroslav Bulatov Traces the Origins of the Memory Wall in New Research Writeup — yaroslavvb · 2026-10-08
- Scale AI partners on HSS, a new benchmark for 'seeing what's in front of them' — AndrewDai · 2026-10-08
- Xiaomi's open-sourced MiMo RL environments leak answers in Git history for 2/3 of coding tasks — teortaxesTex · 2026-10-08
- BeFOND: encoder-free sparse coding matches Gemma Scope with 1000x fewer training tokens — gottapatchemall · 2026-10-08
- Rust+Vulkan Inference Backend Goes Cross-Vendor: Pick Kernels by CPU ISA, Not GPU Vendor — PhysicsDisastrous462 · 2026-10-08