Alibaba proposes UniSGR-V2, anticipating continuous target-side features in generative recommendation
_reachsumit · x · 2026-10-14
Alibaba researchers propose TacGR (UniSGR-V2), tackling the "target-side feature vacuum" in Semantic-ID-based generative recommendation.
- Problem: generative recommenders unify retrieval and ranking via autoregressive decoding, but quantization compresses item attributes into a finite codebook and the target item's features can't be looked up during decoding — causing high conditional entropy, cascading errors per SID level, and missing feature crossing for final ranking
- Method: before decoding each codeword, anticipate a continuous target-side feature vector and cross it with the semantic prefix and user behavior sequence inside the decoder; the crossed representation conditions generation and is directly used for ranking
- Challenge: continuous-vector regression risks target-prediction collapse and noise propagating into generation via attention; the paper proposes mitigations
Abstract truncated; see paper for full details.
More from Research
- PyVBBC 1.5 Released After 3 Years, Bringing Major Upgrades to Bayesian Model Inference — twiecki · 2026-10-14
- AgentHorizon: 1,373-task benchmark exposes unreliable LLM judges for computer-use agents — rohanpaul_ai · 2026-10-14
- ServiceNow and McGill's AgentHorizon benchmark shows LLM judges miss subtle agent failures — rohanpaul_ai · 2026-10-14
- Cambridge's van der Schaar to keynote on AI and oncology at Momentum 2026 — MihaelaVDS · 2026-10-14
- Falcon engineer asks: what does RL add over SFT in a single-step decision process? — antoine_chaffin · 2026-10-14
- Visual SLAM Match Moving: Reconstructing Camera Pose to Composite CG Into Footage — rsasaki0109 · 2026-10-14