vec2vec translates any two text-embedding spaces without paired data, hits 0.96 cosine
maier_ak · x · 2026-09-10
Cornell's vec2vec learns a translation between any two text-embedding spaces without paired sentences, the original encoders, or a dictionary, using adversarial training, reconstruction, cycle-consistency and distance-preserving losses. On the hardest cross-backbone pairs it reaches up to 0.96 cosine similarity, 100% top-1 accuracy and mean rank 1; on out-of-distribution tweets and clinical notes it still exceeds 0.73 cosine, far above random guessing.
More from Research
- Goodfire commits $10M to launch a 'Human Genome Project for AI' interpretability effort — mathildepapillo · 2026-10-27
- Jon Barron: every softmax and cross-entropy implementation has two subtle numerical bugs — jon_barron · 2026-10-27
- New theory paper explains what Anthropic's Jacobian lens actually detects in LLMs — wgilpin0 · 2026-10-27
- Interpretability researcher joining Goodfire AI post-PhD, calls for speeding up the field — mathildepapillo · 2026-10-27
- EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- TRL adds Score Centering, a simple fix for RLHF training instability — lvwerra · 2026-10-27