TCBSA: training-free test-time method injects collaborative scores into generative recommender beam search
_reachsumit · x · 2026-10-26
The arXiv paper TCBSA (Test-Time Collaborative Beam Search Adaptation) addresses premature beam pruning in generative recommenders using semantic identifiers. It converts score margins from an independently trained discriminative recommender into potentials over shared semantic edges, with a confidence gate regulating their contribution to cumulative path scores before beam selection—no retraining of either model. The paper also establishes a sufficient condition for target prefixes to survive pruning.
More from Research
- HKU/ZJU/Fudan with Lightspeed launch HiSTEVE: hierarchical VLA agent playing Minecraft via native mouse and keyboard — tencent · 2026-10-26
- Sony researchers apply flow matching to four-stem stereo music source separation in FLOMSS — Shuyang Cui · 2026-10-26
- Zhejiang University and Tencent propose Easy-3Dfy: decoupling object generation and layout for single-image 3D scene construction — Ziang Zhang · 2026-10-26
- One week: 80-year-old cells reset to under 20, AI designing working proteins — rand_longevity · 2026-10-26
- Anthropic interpretability demo shows neurons combining into tiny pirate/wizard personas — Sauers_ · 2026-10-26
- 17-page HMM chapter PDF from Jurafsky & Martin's NLP textbook shared — burny_tech · 2026-10-26