Abliterlitics benchmarks GLM-4.7-Flash abliteration methods

The real story is not that safety disappears, but that different abliteration methods trade off where the damage shows up. Heretic looks like the cleanest cut; broader or router-heavy edits preserve ASR but increasingly distort reasoning efficiency and output reliability.

• –Heretic is the best balance here: strongest GSM8K adjusted score, lowest empty rate, and the smallest visible collateral on the rest of the benchmark suite.
• –HauhauCS does not look “lossless” in practice; its raw GSM8K drop is mostly an empty-response problem, but the higher empty rate still means worse usability on a reasoning model.
• –Abliterix is the most extreme case of preserving underlying reasoning while breaking delivery: adjusted GSM8K stays near base, but half of raw runs go empty.
• –The CoT forensics are the most interesting part: safety reasoning still appears in a large share of outputs even when the final refusal layer is gone, which suggests the edits reroute behavior more than erase it.
• –Cross-technique cosine similarity staying low supports the paper’s main conclusion: there is no single universal “abliteration subspace,” even on the same base model.