Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Tsinghua team introduces AWM for formal reasoning in unknown environments

AI summary

A research team led by Tsinghua University’s Yao Zhi and Yuan Yang introduced the Algebraic World Model (AWM), which explores unfamiliar environments, revises hypotheses about their rules, and converts them into mathematical models with proofs checked by the Lean 4 kernel. In tests on ARC-AGI-3, the team reports that AWM produced kernel-checked proofs of unreachable goals in all six runs, while programmed Claude Opus 5.5 and GPT-6 Astra baselines produced no such proofs in twelve runs. The researchers say these tests were confined to discrete-action ARC-style environments, and that a valid proof still depends on whether the model’s assumptions accurately describe the real environment.

In-depth insight

AWM offers a way for AI agents to support conclusions about unexplored actions with explicit, machine-checkable proofs and traceable modeling assumptions.

假设有一个游戏关卡,AI 尝试了很多次都没能通关,那么究竟是它还没找到解法,还是给定的行动次数根本不够?

前一个问题,只要找到任意一条获胜路线就能回答,但要证明无解,就必须排除所有合法路线,包括那些尚未尝试过的路线。如果游戏连规则都不公开,AI 还需要先弄清自己究竟在证明什么。通关和证明无解,难度完全不同。

为此,图灵奖得主、中国科学院院士、清华大学交叉信息研究院院长姚期智团队提出了“代数世界模型”(Algebraic World Model,AWM),一个将与陌生环境的交互转化为支持形式化推理的显式数学模型的框架。

AWM 首先通过交互探索环境,提出并修正规则假设,再将这些规则组织成数学模型,最终写入 Lean 4,构造可由内核检查的形式化证明。借助这套方法,AI 不仅能够回答怎样获胜,还能证明在给定约束下为什么无法获胜。

#Algebraic World Model#Tsinghua University#Lean 4#ARC-AGI-3#Claude Opus 5.5#GPT-6 Astra

Original · 6 min readRead the original
Group chat

No comments yet. Start the conversation.