A proprioception test for LLMs: can models verbalize which MoE expert activated?
Sauers_ · x · 2026-10-21
A quirky new paper explores introspection in LLMs via an analogy to human proprioception: instead of unverifiable self-reports, a model could verbalize which MoE expert activated in a given layer, providing an externally verifiable signal of the model's sense of its own internal state.
More from Research
- Stanford's Kundaje: genomic AI models must test cell-type specificity to expose real weaknesses — anshulkundaje · 2026-10-21
- GPT-6 Astra and Claude Fable 5.1 write $10,000 research papers — experts reject all — sayashk · 2026-10-21
- giffmana on tokenizer semantic drift: encoders aren't all tokenizers — giffmana · 2026-10-21
- Librarian Agents: RL-trained agents build file systems, lifting RAG accuracy 44%→77% — simonguozirui · 2026-10-21
- OpenAI researcher: molecular discovery must jointly optimize performance and manufacturability — shyamalanadkat · 2026-10-21
- First Passage Attention: path extrema enable perfect long-range retrieval with 25% cache — Yuxuan Liu · 2026-10-21