Voyage AI: embedding + sklearn classifier beats Jev decision models with 37% error reduction
csinva · reddit · 2026-10-29
Researchers at Voyage AI push back on the hype around the decision model Jev: in most real use cases you already have labeled data, so you can personalize your model. They tested the old-school pipeline:
- Embed inputs with an open-source or cheap embedding model
- Fit a simple sklearn classifier/regressor on top using your own labels
The result was cheaper and almost always significantly more accurate — 37% error reduction vs Jev. The author (disclosed Voyage employee) published a write-up and code, and asks whether others have found cases where zero-shot wins.
More from Research
- RL can train design taste into models — but the taste belongs to the judge, not you — simonguozirui · 2026-10-30
- BellowsKV shares one latent across layers, keeping 97-98% accuracy at 13-14x KV cache compression — burkov · 2026-10-30
- Ryan Williams: OpenAI's ETH-violating solver fits the "split and list" algorithm class — rrwilliams · 2026-10-30
- Robot contact prediction idea: mask body patches by imminent-contact likelihood — breadli428 · 2026-10-30
- New paper benchmarks how far written rubrics go vs. humans' tacit knowledge in judging AI — stanfordnlp · 2026-10-30
- What concepts can't words describe? Stanford team frames it as channel coding — stanfordnlp · 2026-10-30