Meta's MetaEncoder: bi-encoders keep constant inference latency while cross-encoders win on accuracy
_reachsumit · x · 2026-11-13
Meta released MetaEncoder, a paper exploring the limits of bi-encoders for multimodal System One decision-making — fast, low-latency decisions with calibrated probabilities instead of token generation.
Key findings
- Compares three scoring architectures adapted from a pre-trained decoder: bi-encoders, cross-encoders and constrained decoders.
- Cross-encoders achieve the highest accuracy across model sizes; bi-encoders and constrained decoders perform similarly as training tokens scale.
- Bi-encoders keep near-constant inference latency as candidate pools grow (embeddings cacheable offline), while the other two scale quadratically — making bi-encoders uniquely suited to bridge small closed-set decisions and massive open-set retrieval.
- All architectures scale clearly with model size.
Related event: Meta Unveils MetaEncoder: One Encoder for Retrieval and Decision-Making(2 posts)→
More from Models
- METR data on Anthropic employee speedups validates vibe-testing, says Mollick — emollick · 2026-11-13
- Google's Gemini 4 Hydrogen outscores bigger Argon just four weeks after launch — The Decoder · 2026-11-13
- Kaggle launches Arenas: community-voted head-to-head AI model leaderboard — amankhan · 2026-11-13
- Screenshot shows Claude Code account disabled mid-task for 'Terms of Service violation' — cryptodunck · 2026-11-13
- Dev begs AI industry to bring back small/medium/large model naming — andpoul · 2026-11-13
- Same Ghibli 3D prompt: Gemini 4 Hydrogen clearly out-builds Opus 5 — LightAppropriate624 · 2026-11-13