Unverified50% confidenceFactExact time
GLUE和SuperGLUE基准曾经历快速饱和，模型分数超越人类标注者但不意味着语言理解的真正突破
1
Sources
50%
Confidence
Long-term
Relevance
8/18/2026
First Seen
Sources
Related Claims
UnverifiedRecent research from Google DeepMind indicates that models' capability growth in 'understanding and fixing existing code' has outpaced their growth in 'generating correct code from scratch'.57% similarPartially VerifiedDeepMind's AlphaProof and AlphaGeometry have demonstrated strong capabilities on International Mathematical Olympiad (IMO)-level problems, but are essentially specialized systems.57% similarUnverifiedGemini's Deep Research output is more humanized, closer to a natural expression style; ChatGPT's Deep Research is more engineering-oriented, suitable for structured work tasks.56% similarUnverifiedAfter 2022, DeepMind's AlphaProof and various large language models began combining neural networks with symbolic reasoning.56% similarVerifiedDeepMind's AlphaProof combined large language models with Lean 4 in 2024 and used reinforcement learning to self-play in a formal environment to generate and verify proof steps.55% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/771151API
curl https://kongchang.com/api/v1/knowledge/claims/771151MCP
get_claim(id=771151)