HERO uses dual-teacher on-policy distillation to cut popularity bias in generative recommenders
_reachsumit · x · 2026-11-04
The paper proposes HERO, a diversity-aware on-policy distillation (OPD) framework for generative recommendation, deployed in a production "Guess What You Like" module.
- Problem: OPD gives denser supervision but aggravates popularity bias — teacher feedback favors popular-item tokens while student rollouts rarely generate unpopular-item SID paths, and the two biases reinforce each other
- Method: trains heterogeneous teachers specialized in popular vs. unpopular items, routes each sample by ground-truth popularity group; diversity-aware rollout injects ground-truth SIDs of unpopular items; only the student is used at inference
- Validation: offline experiments on an industrial dataset and three public datasets show effectiveness, diversity, and robustness
More from Research
- Noetra engineer's first project: building an image encoder from scratch for multimodal LLMs — HirokatuKataoka · 2026-11-04
- Japanese team trains image encoder from scratch to match SigLIP2, with better CJK performance — HirokatuKataoka · 2026-11-04
- SpecJudge author argues LLM-as-judge tools should never ask models for self-reported confidence — jokiruiz · 2026-11-04
- Tsinghua's BEMULA benchmark: no LLM agent memory system passes over 1.5% of forget requests — rohanpaul_ai · 2026-11-04
- KAIST AI's FactEmbed decouples query understanding to improve composed image retrieval — kaist-ai · 2026-11-04
- Intology's Locus auto-designs molecular models, beating GPT-6 and Claude agents — rohanpaul_ai · 2026-11-04