I'm really excited about this. Because this model is trained with literally nothing but LM loss, it helps create a new reasoning paradigm where reasoning capabilities are baked right in at pretraining, unifying train and test time behaviors.
Look ma, no distribution shift! 🙏

Better yet, without us teaching the model to do this at all, it learned to allocate more compute at tokens of higher entropy (even as measured by an independently trained model of the same architecture), and use less compute where there's either too little or too much entropy. 🤯

By just using our approach, you don't have to do any extra work to get pretraining gains! We show across scale AND computation match that our approach performs better in pretraining perplexity than both regular transformers and manually inserting non-adaptive thinking tokens. 🥳

We design an transformer variant that uses a score-attenuated "forking" mechanism to clone useful residuals the model wants to update and attend to, thus creating a 𝗯𝘂𝗯𝗯𝗹𝗲 of latent computation for those highly-informative tokens.

Current approaches in scaling inference-time compute require supervising with explicit chain-of-thought data, which limits thoughts to be sequential and in human language only. 😔
Wouldn't it be nice if you can do normal pretraining, and somehow get latent thinking for free? 🤔

Joint work with my wonderful collaborators @shikharmurty.bsky.social, @robertcsordas.bsky.social, and @chrmanning.bsky.social.

Paper: arxiv.org/abs/2510.00219.
Code and Package: github.com/stanfordnlp/....

Thanks to @schmidtsciences.bsky.social and Lambda Labs for generously supporting our work :)

🤔 think this is all too much? No worries, we are also dropping a **PACKAGE** to do this for you. Check it out: github.com/sisl/astra-rl

☝️ And so.... You should optimize for **BOTH** attack success and perplexity to get the most effective attacks!

Even across baseline methods, low-perplexity prompts result in more effective attacks, but optimizing for attack success alone results in high-perplexity prompts.

In fact, our method allows us to discover a Pareto tradeoff (🤯) between attack success and prompt likelihood; tuning a single parameter in our method travels along the Pareto-optimal front.

Using the Adaptive Stress Testing (AST) framework as a reward signal for an online DPO-based optimization, we present a method to discover **both** high-probability prompts that are also successful in attacks.

Most approaches in gradient-based red-teaming result in very low-probability prompts, which previous work have shown are both easier to filter and bad negative examples for downstream hardening.

Done at Stanford Intelligent Systems Laboratory — my joint first author Amelia Hardy, along with our wonderful collaborators Allie Griffith, @bernardlange.bsky.social, Duncan Eddy, Mykel Kochenderfer.

Paper:
arxiv.org/pdf/2407.09447
Python package to do this for yourself:
github.com/sisl/astra-rl

You're not too dumb for Haskell, you just need a reason to practice. :)