SA-OPSD distills self-teacher knowledge at beam-search pruning boundaries for generative recommendation
_reachsumit · x · 2026-11-04
The SA-OPSD paper improves on-policy self-distillation (OPSD) for generative recommendation training.
Two problems: visited states in beam search differ in importance, and initially reliable decisions can approach the pruning boundary as training evolves; conventional distribution matching couples target-token confidence with preferences among competing continuations, weakening relational supervision when the self-teacher strongly favors the target.
The method introduces Evolving-Boundary Sampling, which prioritizes search-critical states by weighing pruning-boundary proximity and the decline of initial reference retention margins under a controlled distillation budget, plus Confidence-Decoupled Branch Distillation, which separates target-probability calibration from informative sibling relations to better preserve the teacher's preferences among competing branches.
More from Research
- François Chollet: modeling brain wiring as an engineering problem constrained by genome size and time — CatAstro_Piyush · 2026-11-04
- OffsecCyber dataset combines HackerOne reports and CVEs to train pentesting LLMs — AccBalanced · 2026-11-04
- Monofact rate in training data governs LLM scientific discovery, preprint finds — CatAstro_Piyush · 2026-11-04
- qwenTPU compiles a TPU into Qwen3.5 weights, achieving exact integer matmul — CatAstro_Piyush · 2026-11-04
- Jev model improves ICL demonstration retrieval via distillation into a lightweight MLP — kalyan_kpl · 2026-11-04
- Indie researcher open-sources 10 years of math work: 146k files, AI-accelerated over 3 years — HamOnWhy · 2026-11-04