Join the discussion
Write your take first — we'll ask for email only when you're ready to publish.
- Hacker News
- Author here. Quick version: “abliteration” (basically removing the direction in the model that causes it to refuse) is the go-to method people use to make open models uncensored.
Most people treat it like a clean surgical cut - it just kills the refusals and leaves everything else untouched. I tested that assumption on Gemma and Qwen with 21,600 pre-registered decisions under uncertainty, using identical frozen inputs for the base vs. abliterated versions.
Turns out it’s not surgical at all.
The abliterated models systematically become more optimistic, hedge less, show no improvement in actual task performance, and the same edit even moves their expressed confidence in opposite directions depending on the model family.
Preregistration, dataset, and analysis code are all public. Happy to answer any methodology questions or hear where you think this falls apart.
by oleczek - I vouched for this post, because it's the author's post and certainly contributes to the thread.
I'm sure you're being flagged because you use LLMs to clean up your English grammar. Probably best not to do this here, because HN has a policy against LLM generated posts. There should be some leeway when a non-native speaker is using it for assistance, but there are a lot of people who do not care and will eagerly flag you for it.
by kbelder - > Quick version: “abliteration” (basically removing the direction in the model that causes it to refuse) is the go-to method people use to make open models uncensored.
Tru-ish (lots of people distinguish between abliteration and uncensoring, though.)
> Most people treat it like a clean surgical cut - it just kills the refusals and leaves everything else untouched.
Basically no one does this, its widely recognized that this isn’t how it works and it has for quite some time been common for makers of anliterated model versions to publish metrics for how far a particular abliteration (1) removes refusals (typical before/after refusal rate on a standard test set), and (2) diverges to the output of the base model (KL divergence), and it is widely understood that there is generally, in practice, a tradeoff between these two metrics, where more refusal reduction tends to come at the expense of higher KL divergence.
That’s not saying that it isn’t interesting and new to characterize the kind of divergence that occurs with abliteration in different model families, but there is no reason for a late-night informercial level of misrepresentation of the existing understanding to come along with that.
by dragonwriter - Obviously Claude written paper.by invictati
- yeah English isn't my first language so i used AI to clean up the writing. the research, the data and the analysis are all mine and all open (in the linked repo)by oleczek
- Color me unsurprised that caution is based in shame and anxiety.by 60secs
- Yeah, the optimism/hedging part lines up nicely with that framing. The bit that still puzzles me is that the same edit moved expressed confidence in opposite directions — down for Gemma, up for Qwen. If it were simply removing one shared "anxiety" factor, I’d have expected the sign to stay consistent. So whatever caution is doing here seems a bit more tangled up with each model’s own disposition than a single clean knob.by oleczek
- Should have kept the original title.by gweinberg
- Yeah, original is clearer. Just went with the shorter one so more people would actually click.by oleczek
- Censorship has no place in policy of any frontier company hoping to be competitive. That onus should be on the user, the law, and society. Anything less is not a competitive product, as proving out here.
Treating society with gloves reduces progress and creativity alike.
by 4d4m