Moksh Jain

Check out our new work on action abstractions for amortized samplers led by @boussifo.bsky.social! Simple tokenization schemes like BPE result in meaningful action abstractions with several empirical implications for amortized samplers. Come chat with us @iclr-conf.bsky.social!

Amortizing intractable inference in large language models Autoregressive large language models (LLMs) compress knowledge from their training data through next-token conditional distributions. This limits tractable querying of this knowledge to start-to-end a...

It's nice to see the elicitation perspective getting discussed! RL on CoT is really just a more reliable way of eliciting latent capabilities of the model than simple prompting. We took this perspective in our work (arxiv.org/abs/2310.04363), which was also one of the first to use RL on CoT.