Perplexity's pplx-embed-v2-context ditches gold-chunk labels, sets SOTA on retrieval benchmarks
_reachsumit · x · 2026-10-30
Perplexity researchers train contextual embeddings (pplx-embed-v2-context) by distilling graded chunk-level relevance from a query-aware context compression model, removing the need for gold-chunk labels. The vision-language-based family ships in 9B and 0.6B sizes with compatible embedding spaces; the 9B model sets new SOTA on their two new benchmarks (Context-bench, Q2D-Chunk) and ConTEB. They also release Q2D-Chunk as a public leaderboard and show visual retrieval transfers from the base model.
More from Research
- Interactive Visual Guide: Byte-Level vs Token-Level LLMs (and Why Opus Misspells 'Flowerbed') — dwmkerr · 2026-10-30
- First Diffusion Model Running on a Game Boy: 1.08M Params, 85 Seconds Per Doodle — ready-eddy · 2026-10-30
- Amazon's DreamFAR world model nearly doubles π₀ real-world performance for robot training — CyberRobooo · 2026-10-30
- Investor's Last Exam Benchmark: Top AI Models Ace Research but Fail at Investment Judgment — Huangfei Jiang · 2026-10-30
- SpecSlot Boosts Pixel-Space Image Generation: +2.84 FID on ImageNet-512 at Lower Cost — Jiongfan Yang · 2026-10-30
- Apple Trains UI Navigation Agents with Online RL Directly on Live iOS Apps — apple · 2026-10-30