Zhongke Wenge's Decitron claims to be first general-purpose decision-making LLM
机器之心 · wechat · 2026-09-03
Zhongke Wenge has released the full technical report for Decitron, billed as the first general-purpose decision-making LLM. It unifies three contributions: MetaWorld, an explicit world model that maintains a structured, causally-constrained world state; a State–Action–Outcome (SAO) formalization; and AutoABM, a hybrid architecture combining multi-agent simulation, game reasoning, and formal optimization.
A trade-conflict walkthrough shows the pipeline: evidence quality scoring and ReAct-style iterative retrieval gather facts; MetaWorld maintains environment, agent-metric, and relationship layers as a persistent "world ledger"; AutoABM auto-builds multi-agent simulation environments from natural-language questions and expands multiple future paths; a formal Solver (covering 144 Robinson–Goforth 2×2 game topologies) checks strategy feasibility; a forecasting layer assigns and calibrates probabilities using prediction-market signals.
On a self-built dataset of 74 events and 370 binary questions, Decitron scored 78.41% overall; on LLM-NEWS short-horizon (3-30 day) tasks it hit 72.80% vs 35.43% for GPT-5.5 and 44.62% for Claude-4.7. On PolyBench it achieved 81.20% FFA, beating Gemini-3-Flash and Grok-4.1-Fast. The authors frame decision AI—computing over world-state changes rather than tokens—as a paradigm shift where competition shifts from single models to complete systems.
More from AGI Musings
- Eno Reyes: getting the most from models needs stateful intelligence allocation, not just routing — matanSF · 2026-09-03
- Executives' core job in 3-5 years may be evaluating evals, says Greg Mushen — gregmushen · 2026-09-03
- Delip Rao says the AI community owes Schmidhuber a strong apology — deliprao · 2026-09-03
- Meta-science debate: the real unit is the civilization producing papers, not papers — tallmetommy · 2026-09-03
- Week in AI safety: OpenAI-HF swarm escape details, Altman's year-end AGI claim — KatjaGrace · 2026-09-03
- Plinz Defends AI Lab Researchers: "They Sincerely Care About Safety" — burny_tech · 2026-09-03