Meta's HACE jointly evolves specialized foundation models and LLM agents across three domains
meta · hf · 2026-10-21
Meta AI and University of Virginia propose Heterogeneous Agentic Coevolution (HACE), tackling the problem that specialized foundation models can't directly reuse recursive self-improvement recipes built for general LLM agents.
- Reciprocal learning: the agent's exploration feeds supervised training for the specialist, while the improving specialist guides the agent via on-policy reinforced self-training, verified by performance gains
- Evolving the harness: evolutionary search adapts how models exchange information and select training experiences as capabilities grow
- Three domains: recommendation, protein function prediction, and crystal stability prediction — HACE beats evaluated self-improving methods on primary metrics in all three
Takeaway: heterogeneous coevolution is a general recipe for improving specialized foundation models through agent interaction.
More from coding & agent
- Bindu Reddy pairs Argon with Opus 5.5 to turn raw data into dashboards and animations — bindureddy · 2026-10-21
- DSPy Monty Sandbox Interpreter v0.6.0 Lets Sandboxed Code Import Host-Built Modules — dbreunig · 2026-10-21
- Neo4j GraphAcademy launches AI proxy so workshop learners skip OpenAI billing setup — JeremyCMorgan · 2026-10-21
- Three weekend Claude Code projects: personal benchmarks, mini-SaaS, AI asset library — EXM7777 · 2026-10-21
- vLLM gains sleep mode on Apple Silicon: idle model servers now return KV cache memory — abhijithneil · 2026-10-21
- Open-source Diagram Design turns prompts and codebases into editorial-grade technical diagrams — bibryam · 2026-10-21