Stable AI proposes DPDM: frozen LLM representations plus shared predictor boost agent decisions
stable-ai · hf · 2026-11-06
A paper from Stable AI and Tsinghua University introduces Decoupled Predictive Decision Making (DPDM), an architecture that frames agent decision-making as general-purpose conditional inference.
A frozen language model provides semantic representations of current and historical records, while a shared pretrained conditional predictor learns task-relevant empirical relationships from deployment evidence. Pretraining supplies reusable inference capability, and problem-specific predictions are instantiated without updating any neural parameters, improving agent decision accuracy.
More from coding & agent
- Formal proofs of tens of millions of lines of code are now feasible, says Anthropic engineer — morqon · 2026-11-06
- Terminal-Bench's two broken tasks traced to Debian 11 archive repo migration — PawelHuryn · 2026-11-06
- One prompt ran Codex for 17+ hours with 109 sub-agents and 2B tokens — andimarafioti · 2026-11-06
- Reddit devs compare agent memory approaches: summaries, MEMORY.md, wikis and vector DBs all fail differently — ombre-ai · 2026-11-06
- MCP vs API vs CLI: build both on one operations layer, prioritize MCP — hottown · 2026-11-06
- Web Search MCP Server adds Google-powered search and scraping to agents — modelcontextprotocol · 2026-11-06