DeepMind Paper Outlines LLM Limits in Scientific Discovery
In an ICML 2026 position paper, Google DeepMind researchers examine the capability boundaries of current Large Language Models (LLMs), arguing that data compression and deductive reasoning alone are insufficient for fundamental scientific breakthroughs. This has sparked broad industry debate over whether LLMs can truly lead to superintelligence.
已确认
- The paper, co-authored by DeepMind researcher Tom Zahavy, breaks down scientific discovery into three steps: encountering sensory data, proposing abstract axioms through illogical "intuitive leaps," and finally conducting rigorous logical deduction.
- It suggests that while current LLMs excel at combinatorial tasks and logical deduction, they cannot perform the crucial "intuitive leaps" required for scientific discovery.
- Using Einstein's theory of general relativity as an example, the paper illustrates that fundamental scientific breakthroughs require axiomatic leaps beyond existing data.
为什么重要
- The research directly addresses a core controversy in the AI field: whether scaling up models and increasing data volume is enough to achieve Artificial General Intelligence (AGI).
- Observers like Peter Berezin argue that if LLMs cannot bridge the gap of conceptual innovation, the current technological path might be a dead end to superintelligence (as illustrated by the caption "LLMs can’t jump").
- This study offers a new perspective for evaluating the true capabilities of large models, indicating that they still face structural limitations in domains requiring human conceptual priors and intuition.
2026-07-28 ~ 2026-07-29 · 5 related posts
Primary sources
- Google DeepMind says LLMs can prove theorems but still cannot make scientific “jumps” — kylekabasares · 2026-07-28
- What Can Current LLMs Solve? DeepMind Research Maps Capability Boundaries — burny_tech · 2026-07-28
- [source] Einstein’s three-step view of discovery says AI still lacks the intuition jump — TZahavy · 2026-07-28
- Peter Berezin says LLMs may be a dead end to superintelligence — GaryMarcus · 2026-07-29
- [source] DeepMind paper says LLMs need abductive reasoning, not just compression or proofs — burkov · 2026-07-29