← Back to feed

[Paper] Distillation for Incrimination and Distillation for Capabilities

L5 · ResearcherResearchLessWrong AI· 10/9/2026

Directly addresses AI safety research methods for detecting deceptive alignment in advanced AI systems.

AI Summary

Researchers release empirical paper testing distillation techniques for AI safety, including Distillation for Incrimination to detect hidden misalignment in models.

Excerpt

TL;DR: We previously motivated two complementary ways to use distillation for AI safety; we’re releasing a paper empirically testing those ideas. Distillation for Incrimination (DFI) distills an untrusted model into a weaker, trusted one, aiming to transfer the teacher's misalignment but not its ability to hide it, so that auditing the student can incriminate the teacher. We test DFI on AuditBench's secret-keeping model organisms, each fine-tuned to have a hidden quirk and trained to deny it. Di

Read Original
0 upvotes · 0 downvotes · 1 min read

Related Articles

L5 · ResearcherResearch@emollick.bsky.social
Ethan Mollick (@emollick.bsky.social): Randomized trials with GPT-4o: "AI access raises test scores, and a smaller gain persists a week later. Gains remain for students who use AI as a tutor (“augmentation”) and fade for students who have…

Randomized controlled trial shows GPT-4o access improves student learning outcomes, with persistent gains when used as a tutor but fading benefits when used for automation.

L5 · ResearcherResearchLessWrong AI
New Math from OpenAI

OpenAI released 722 mathematical proofs generated by an internal frontier model, including solutions to 90 top open problems from Proof of Atlas.

L5 · ResearcherResearchThe Verge AI
‘Pure insanity’: Mathematicians will need years to make sense of OpenAI’s latest drop

OpenAI released nearly 400 AI-generated mathematical results across 700 manuscripts, overwhelming mathematicians who say it could take years to verify and understand the implications.

L5 · ResearcherResearchLessWrong AI
What can language models teach us about understanding?

The author proposes developing a mathematical theory of 'intuition' inspired by language models, aiming to establish a formal framework similar to how logic and computation were formalized historically.

L5 · ResearcherResearcharXiv
WOVEN: Weaving Visual World Modeling into Multimodal LLMs

Researchers introduce WOVEN, a benchmark and training method that improves multimodal LLMs' visual transition reasoning capabilities by up to 27.3 percentage points on external benchmarks.