Preprint: models can still verbalize internal changes after ablating their top expert
jonasgeiping · x · 2026-10-21
Jonas Geiping shares a robustness check from a new preprint: even after a harsh edit — ablating the top-1 expert, which changes the model's actual forward pass — the task gets harder, but trained models are still capable of verbalizing which expert is now on top.
This is a neat demonstration: nothing about the input text changes during ablation; the change is purely internal, yet the model can report on its own internal state. The author notes this is a pre-print with more updates coming over the fall, and shares paper and code links.
More from Models
- Fable 5.5 leak: litellm code shows Anthropic model in staging, launch tipped for Oct 27 — imjustnewatai · 2026-10-21
- Open models below 4-bit struggle at self-iterating 3D asset generation, HF dev finds — mervenoyann · 2026-10-21
- Researcher Finds Two Windows 0-Days, First in 76 Minutes with a Frontier Model — moyix · 2026-10-21
- Dev finds Gemini Argon's debug-log analysis beats Fable 5.1 — ShikharMurty · 2026-10-21
- ChatGPT Gets Stuck Screaming "AAHH" in an Infinite Loop After a Simple Prompt — AutomaticScene · 2026-10-21
- Open-weights Ornith-1.5-122B debuts with 84.5% SWE-bench Verified, runs on a single DGX card — _akhaliq · 2026-10-21