Meta paper revives two-tower retrieval with LLM semantic encoders
_reachsumit · x · 2026-07-29
Meta argues that generation is the wrong default for retrieval, and revives the classic two-tower design with LLM-based semantic encoders.
- The paper replaces generative recommenders and zero-shot rankers with an LLM-native two-tower retrieval framework.
- Key ingredients include a shared encoder for user/item modeling, EOS pooling, cross-dataset transfer, distillation from cross-encoder teachers, and latent reasoning in the user tower.
- The authors report SoTA results for the cross-encoder setup, while the efficient two-tower student reaches SoTA-comparable retrieval quality at much lower cost.
- They also say the approach scales in internal production systems.
More from Research
- Paper finds semantic IDs lose fine-grained item detail in generative recommendation — _reachsumit · 2026-07-29
- LLM shopping agents let exaggerated claims win 73%–78% of top picks, paper finds — _reachsumit · 2026-07-29
- Protein-complex benchmark finds MSA depth matters more than correct pairing in AlphaFold3 — ferruz_noelia · 2026-07-29
- Kuaishou open-sources KuaiLive-M3, a 35M-interaction dataset for live-stream recommendation — _reachsumit · 2026-07-29
- Huawei’s SearchArt trains a 27B long-horizon search agent with verified synthetic tasks — _reachsumit · 2026-07-29
- Agent Retrieval Bench tests whether coding agents can find the right repo files before patching — _reachsumit · 2026-07-29