Continuous queries beat Semantic ID beam search in generative recommendation with 55% fewer params
_reachsumit · x · 2026-10-12
The paper identifies a flaw in generative recommendation (GR) as used by social media platforms: autoregressively decoding discrete Semantic IDs (SIDs) with finite-width beam search means held-out targets can be unreachable in the finite returned lists, and aligning a discrete SID vocabulary with the pretrained backbone is technically painful.
Two continuous-query paradigms are proposed: CQG-Single generates one retrieval query in a single pass, while CQG-AR autoregressively generates continuous residual updates — keeping step-wise decoding but removing the SID vocabulary, code lookup and discrete beam expansion. Both make every encoded item directly scoreable via dense or ANN retrieval.
Under a unified protocol with valid-prefix constrained decoding, CQG-Single beats the TIGER-style baseline in all seven runs with only 55.3% as many trainable parameters; CQG-AR is comparable on ML-1M and numerically stronger on Beauty and Sports. A controlled depth ablation further validates the designs.
More from Research
- Kling's DiffusionART cuts video RL fine-tuning cost by 12x while raising rewards up to 53% — KlingTeam · 2026-10-12
- Reflex teaches humanoid robots to catch thrown boxes from egocentric vision, 65% success on real Unitree G1 — UW · 2026-10-12
- SynthSelf enables continual post-training of small models with +22pp acquisition and only -1pp forgetting — LG-AI-Research · 2026-10-12
- Meta's ReHarness adds backward revision to agent harnesses, beating memory and reflection baselines — Yuxin Xiao · 2026-10-12
- Packing lab: Monte Carlo sandbox nudges within 0.04% of published 5D sphere packing — idanbeck · 2026-10-12
- willcb: every non-RL algorithm is a proxy — policy gradients are the minimum-variance unbiased estimator — willcb · 2026-10-12