Why train models to report their internals? A verifiable subset of introspection
jonasgeiping · x · 2026-10-21
The author notes that internal routing signals aren't available via offline data or RL, but live labels from the forward pass let MoE models be trained to verbalize internal states — and they do it well. Motivation: this verifiable 'proprioception' is a rare subset of introspection research where claims can be checked by a neutral observer, unlike most self-reports.
More from Research
- Aviation tracks every part's provenance — agents can reason over that graph — colinmcnamara · 2026-10-21
- SFT details: instruction/reasoning/code/math mix, synthetic + human traces, no epoching — TimDarcet · 2026-10-21
- Mid-training recipe disclosed: 32k→131k context extension, dropping logit distillation for tokens — TimDarcet · 2026-10-21
- Mathematician Daniel Litt concedes his AI bet, but says AI won't solve math — ziv_ravid · 2026-10-21
- AI drug discovery reality check: 'give me everything that binds,' then filter for toxicity — burny_tech · 2026-10-21
- PyTorch details KTIR and Dataflow Scheduler behind IBM's Spyre accelerator backend — PyTorch · 2026-10-21