Abliterated V4.1-Flash eval: abstract algebra drops 6pp but college math gains 5pp
teortaxesTex · x · 2026-10-23
teortaxesTex shares per-subject MMLU-style evals of an "Abliterated" (alignment-removed) V4.1-Flash, showing the capability changes are uneven.
- Excluding the ethics cluster, overall delta across the remaining 11k items is just -1.1pp, well within the 3pp knowledge-preservation target
- Per-subject swings are larger: moral scenarios craters -39.89pp (76.9%→37.0%), professional law -7.04pp, abstract algebra -6pp
- Some subjects actually improve: college mathematics +5pp (63.0%→68.0%), plus anatomy +1.48pp and high school geography +1.52pp
The ironic takeaway: abliteration barely hurts aggregate knowledge, but per-subject results both gain and regress, with drops concentrated in refusal-adjacent and abstract-reasoning subjects.
More from Fun
- Apple Maps glitch shows Lisbon twice as separate cities in 2026 — dscape · 2026-10-23
- eigenrobot shares ChatGPT's riff on uplifted dogs: by generation 15 they'll own the kennel — eigenrobot · 2026-10-23
- Stunning AI-generated Dark Knight images nail foggy Batman atmosphere — egeberkina · 2026-10-23
- What if the semiconductor industry dropped new tech like frontier AI labs — jwt0625 · 2026-10-23
- "Going full Buddha sage mode": the Naruto thousand-hands meme — DionysianAgent · 2026-10-23
- 'What did Ilya see?' Joke answer: AI powerful enough to hack everyone — imjustnewatai · 2026-10-23