Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs · HackerPapers discussions

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs — mentioned by stratos123

Tue, 28 Jul 2026 06:44:41 GMT

Several of your proposals would make sense if alignment was trivial, but, sorry, I think you're wrong about that. A base model is absolutely not aligned. Pretraining doesn't teach an LLM to mimic the average human - doing so would make an LLM perform quite badly at predicting most of the dataset. It teaches it to mimic all possible humans¹, inferring what sort of persona to take depending on the context. A base model can indeed convincingly act like an "average person in the developed world", including by making the same sort of moral decisions that such a person would do... but it can also…

Paper and discussion history. A mention is not an endorsement.

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs — HN submission

Mon, 02 Feb 2026 17:51:00 GMT

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs

3 points; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs — mentioned by pseudohadamard

Tue, 13 Jan 2026 06:25:02 GMT

It's already been done, without the model being aware of it, see https://arxiv.org/abs/2512.09742. They also made it think it was Hitler (not MechaHitler, the other guy), and other craziness. It's a relief to think that we're not trusting these things for stuff like financial advice, medical advice, mental health counselling, ...

Paper and discussion history. A mention is not an endorsement.

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs — HN submission

Wed, 07 Jan 2026 13:59:48 GMT

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs

3 points; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs — mentioned by FeepingCreature

Wed, 17 Dec 2025 16:56:39 GMT

At least with finetuning, yes: https://arxiv.org/abs/2512.09742 It's unclear if this technique could also work with in-prompt data.

Paper and discussion history. A mention is not an endorsement.

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs — HN submission

Wed, 17 Dec 2025 05:21:01 GMT

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs

7 points; 1 discussion reply on HN. Paper and source history. A submission is not an endorsement.

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs — HN submission

Sun, 14 Dec 2025 19:14:38 GMT

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs

3 points; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs — HN submission

Fri, 12 Dec 2025 02:47:46 GMT

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs

7 points; 2 discussion replies on HN. Paper and source history. A submission is not an endorsement.

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs — HN submission

Thu, 11 Dec 2025 05:07:40 GMT

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs

1 point; 0 discussion replies on HN. Paper and source history. A submission is not an endorsement.