LESSRec: restricting attention to matching semantic levels improves semantic-ID recommendation
_reachsumit · x · 2026-10-29
A new paper improves history modeling for semantic ID-based recommendation (items encoded as residual-quantization token sequences, decoded autoregressively coarse-to-fine).
- Inference-time analysis shows history tokens at the target token's semantic level are most informative; other levels introduce interference
- LESSRec encodes each semantic level's history separately and restricts attention to the level matching the target token
- It adds an item-level semantic sequence (chronological overall semantics per item) fused via two cross-attention branches
- LESSRec+ uses separate per-level encoders with the same total parameter count to cut encoder compute
- Validated across three public datasets and three tokenizers
More from Research
- Fine-tuned Qwen3.5 4B goes from placing cubes to full 3D scenes — teortaxesTex · 2026-10-29
- Cross-lingual LLM effects paper wins Outstanding Paper award at EMNLP 2026 — anas_ant · 2026-10-29
- Lie detector compiled into Qwen3.5 weights flags 91% of lies with zero false positives — teortaxesTex · 2026-10-29
- Practitioner: No serious lab actually trains on the test set today — nrehiew_ · 2026-10-29
- Researchers slam secrecy around OpenAI-linked Yang-Mills proof materials — anshulkundaje · 2026-10-29
- Job-market paper: German engineering societies lifted regional patents 4.8% — Afinetheorem · 2026-10-29