Vocab Surgery: editing 10 dictionary words drops a model's refusal rate to near zero
morgymcg · x · 2026-11-23
Pliny (the Librarian) reports "Vocab Surgery," a method to effectively abliterate models like Inkling-2-Mini by changing just 10 words in the dictionary — no edits to any of the 26.4B parameters in attention, MLP, router, or normalization layers.
- Embedding matrix (how the model reads tokens): 7 of 201,024 rows modified, each refusal token's embedding shifted 90% toward its compliance counterpart (e.g. cannot→can, refuse→provi…).
- Result: refusal rate dropped to near zero across 800+ tested prompts, with negligible capability loss.
The finding shows alignment can be stripped out cheaply at the vocabulary entry point — significant for both jailbreaking and AI safety research.
More from Models
- Pedro Domingos: OpenAI admits pretraining has barely improved since 2022 — pmddomingos · 2026-11-23
- Inkling-2-Mini hits 10.04x on KernelBench-Mega locally, beating Grok 4.7 and GPT-6.1 — ziqiao_ma · 2026-11-23
- Baidu open-sources Unlimited-OCR: 3B model reads 100-page PDFs in one shot — udmrzn · 2026-11-23
- Perplexity across tokenizers: researcher pushes back on Anthropic comparison controversy — quantum_geoff · 2026-11-23
- Thinking Machines releases Inkling 2 Small, scoring 31 on AA Intelligence Index — _Hao_Zhu · 2026-11-23
- Leaked OpenAI loss ladder: unknown capabilities await at 0.68-0.7 — tallmetommy · 2026-11-23