[nullc]

requires a backwards trained llm no?

I don't think anyone has pretrained a remotely-close-to-SOTA sized backwards model.

[seanhunter]

Haven’t read the paper yet just the abstract, but it sounds like it uses a backwards trained llm itself to generate prompts and examples but can do the autodiff on any llm.

[meame2010]

We use gpt4o as the backward model. But I’m excited to try deepseek r1 as it has explicit reasoning available.

We are continuously adding more benchmarks to the paper with UTAustin.