Meta's MIMESIS: 9B user simulator beats GPT-5.5 for training interactive agents
facebook · hf · 2026-10-09
Meta introduced MIMESIS, a purpose-built user simulator for training interactive language agents, trained on human conversations with explicit reasoning supervision and 13 realistic behavioral patterns.
- The 9B model hits SOUL-Index 65.7, beating frontier models; vs Claude-Opus-5 it improves behavioral fidelity by 13.4 points and cuts Turing distance by 3.6.
- Agents trained with multi-turn RL against the frozen MIMESIS outperform GPT-5.5-trained agents across 8 environments and all 9 unseen user simulators.
- Coached On-Policy Self-Distillation (CSD) converts the simulator's private reasoning traces into coaching notes for dense token-level supervision, yielding further gains.
More from coding & agent
- xAI donates $1.5M in Grok tokens to Omarchy as founding corporate patron — SpaceXAI · 2026-10-09
- LightOn's Amélie Chatelain to teach free lesson on scaling late-interaction search, indexes 5-30x smaller — IgorCarron · 2026-10-09
- Researchers use Item Response Theory to predict coding agent performance on unseen tasks — dan_fried · 2026-10-09
- Nace launches NDI 1.0, a small doc-processing model claiming frontier-level parsing at 1/10th cost — nischay_twt · 2026-10-09
- Google unveils universal Gemini agent for work: one prompt box for Q&A, knowledge work and code — sundarpichai · 2026-10-09
- Long Shot Studio: an open-source ComfyUI UI for multi-shot long video with resumable rendering — R34vspec · 2026-10-09