TTT-Embed: Boosting Retrieval Up to 8.36 Without Weight Updates
_reachsumit · x · 2026-08-14
Google DeepMind introduces TTT-Embed, a framework that enhances dense retrievers via Test-Time Tuning of Embeddings.
Core Mechanism:
- Distills ranking rewards from a reranker or LLM judge into a lightweight vector.
- Adds this vector to frozen query embeddings without updating model weights or modifying the index.
Key Advantages:
- Applicable to closed-source models where parameter access is unavailable.
- Uses a single scope parameter to control reward reuse (global, task, or query level).
Evaluation:
Tested across 5 embedding models and 15 MTEB tasks, achieving up to +8.36 nDCG@10 improvement. The learned states also generalize effectively to unseen queries and tasks.
More from Research
- Researcher slams papers whose stacked changes beat even weak baselines barely — retr0jirachi · 2026-10-16
- FrankenSim builds a physics-based virtual orchestra, piano and drum demos live — doodlestein · 2026-10-16
- FoveaVLA replaces wrist cameras with learned foveation, hitting 70.8% success — Haoyu_Xiong_ · 2026-10-16
- ApolloPFN: a zero-shot forecasting network that natively handles exogenous covariates — burkov · 2026-10-16
- ICLR reviewer says 4/5 papers in review pile read like AI slop — berkustun · 2026-10-16
- Tsinghua's DBS learns content-adaptive frequency band splits for audio source separation — thu-ml · 2026-10-16