Survey of 300+ papers says better reasoning does not make LLMs more self-aware
blaizedsouza · x · 2026-07-21
A new survey on **metacognition in LLMs** argues that stronger reasoning does **not** automatically make models better at knowing when they are wrong. - Across **300+ papers**, the authors find that models can reason better while still failing at self-monitoring. - Even **DeepSeek-R1** struggles with basic metacognitive tasks such as estimating the length of its own reasoning trace or predicting task success. - The paper’s central claim: **reasoning depth and self-knowledge depth are not the same thing**, and more reasoning can even correlate with worse metacognitive sensitivity. - It surveys methods, benchmarks, applications, open questions, and a GitHub reading list for the field.
More from Research
- OCT-Bench sets 10,076 questions to test whether multimodal models really understand retinal scans — Baochen Fu · 2026-07-21
- LTX-2.3 face-and-voice LoRA training can work on 12GB VRAM with heavy tradeoffs — __alpha_____ · 2026-07-21
- Follow-up paper argues digital twins could make clinical trials more adaptive — techhalla · 2026-07-21
- Nature npj Digital Medicine paper maps causal inference and digital twins for trials — techhalla · 2026-07-21
- Nature NPJ Digital Medicine Explores Causal Inference and Digital Twins in Clinical Trials — MihaelaVDS · 2026-07-21
- AI performance is increasingly limited by materials science, not just compute — nordicinst · 2026-07-21