Teaching LLMs digital proprioception: models learn to report their own internals
jonasgeiping · x · 2026-10-21
Jonas Geiping presents a 'quirky' new paper. Introspection is hard to study in both humans and LLMs, but some internal senses are verifiable — like proprioception, the sense of where your arm is. A digital analogue for LLMs: verbalizing which expert activated in a given layer for a previous token.
Key findings
- LLMs can learn a whole zoo of such proprioception skills, understanding and reporting on their internals
- These reports can be calibrated and scored, and the team evaluates whether proprioception training changes model behavior
- Motivations: potential safety gains (reasoning over failures, injections, spiraling states), more probe-legible models, and echoing Perez & Long (2023) — calibration on verifiable introspection may improve reliability on unverifiable self-reports
Related event: Geiping team trains LLMs to report internal MoE routing(6 posts)→
More from Safety
- RedMonk: MCP and A2A Verify Publishers, Not Safety — Registries Become the Trust Layer — rseroter · 2026-10-21
- Dev complains Anthropic's bio safeguards trigger just from reading a bio paper — robleclerc · 2026-10-21
- Redditor claims uncensored local DeepSeek build answers virus-writing prompt with zero hesitation — Only_Light_9817 · 2026-10-21
- Pristine Models: training image generators on unsafe data without reproducing unsafe concepts — rzhang88 · 2026-10-21
- Interconnects Artifacts 25: 28 new models, DeepSeek V4.1 Flash, and US-China regulatory controls — natolambert · 2026-10-21
- Prompt injection via state input works on decision models too, security researcher demos — wunderwuzzi23 · 2026-10-21