Probing Qwen3-8B: a simple guesser decodes fictional wing anatomy from hidden states at 100%
Sauers_ · x · 2026-11-11
An interpretability demo: two parallel chats have Qwen3-8B role-play an avatar with and without wings. A simple probe is trained on 2,400 sentences naming wing parts (front edge/membrane/back edge), then tested on new sentences that never say "wing". With wings in the chat the probe picks the right part 100% of the time vs 50% without (random: 33%), showing the model tracks fictional anatomy in its hidden states even when unmentioned.
More from Research
- Cheap builder + frontier reviewer loop matches flagship on 66 tasks at ~1/16 the cost ($0.05 vs $0.72) — KrystalUnity · 2026-11-11
- UMAP 0.6 now installable via PyPI and conda-forge — leland_mcinnes · 2026-11-11
- UMAP 0.6 optimizers lift attraction-repulsion restrictions, boost kNN recall — leland_mcinnes · 2026-11-11
- UMAP 0.6 fidelity gains at scale driven by efficient hard-negative mining — leland_mcinnes · 2026-11-11
- UMAP 0.6 recursive init converges to within 3% of final layout in far fewer epochs — leland_mcinnes · 2026-11-11
- UMAP 0.6 Released: 3M arXiv Papers Embedded in Under 5 Minutes — leland_mcinnes · 2026-11-11