EMNLP paper: LLMs lack human-like knowledge structures that accuracy metrics can't see
rohanpaul_ai · x · 2026-09-12
The arXiv paper "Do LLMs Exhibit Coherent Knowledge Structures in Mathematical Reasoning? A Perspective from Knowledge Space Theory" (EMNLP 2026 findings, Peng Cui, Heejin Do, Mrinmaya Sachan) uses a KST-grounded framework to compare 8 open- and closed-source LLMs against real human learners.
Key findings:
- LLMs frequently violate knowledge dependencies and fail to leverage related knowledge in context for dependent questions.
- LLMs do not share a consistent knowledge structure among themselves (low overlap in knowledge distributions).
- These structural deficiencies are largely invisible to accuracy-based and LLM-as-judge evaluations.
Conclusion: current LLM knowledge does not follow a human-like structure; high accuracy can mask disconnected knowledge pockets, so reasoning models should be evaluated with connected easy-hard problem sets.
Related event: Study Finds LLMs Lack Human-Like Knowledge Structures in Math(2 posts)→
More from Research
- Neural Thickets Revisited: Frozen-Model Logit Perturbations Match Full-Weight Perturbation Ensembles — kastnerkyle · 2026-10-31
- CS312 course post explains sharpness and the edge of stability in LLM pretraining — tatsu_hashimoto · 2026-10-31
- AI decodes wolf howls: 500K hours of Yellowstone audio analyzed by models — aza · 2026-10-31
- NBER paper: training your AI replacement can actually raise workers' pay — joshgans · 2026-10-31
- Trata's Hedge-Bench: no model tops 50% on real-world financial forecasting — ycombinator · 2026-10-31
- Anthropic's internal Claude model claims 2^o(n) 3-SAT algorithm, purportedly refuting the ETH — ctjlewis · 2026-10-31