Kanishk Gandhi

How can we combine the process-level insight that think-aloud studies give us with the large scale that modern online experiments permit? In our new CogSci paper, we show that speech-to-text models and LLMs enable us to scale up the think-aloud method to large experiments!

Can we record and study human chains of thought? Check out our new work led by @danielwurgaft.bsky.social and @benpry.bsky.social !!

Some absolutely marvellous work from @gandhikanishk.bsky.social et al! Wow!

Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRs Test-time inference has emerged as a powerful paradigm for enabling language models to ``think'' longer and more carefully about complex challenges, much like skilled human experts. While reinforcemen...

13/13 Paper at arxiv.org/abs/2503.01307

12/13 Would also like to thank Charlie Snell, Dimitris Papailiopoulos, Eric Zelikman, Alex Havrilla, Rafael Rafaelov, @upiter.bsky.social and Archit Sharma for discussions about the magic and woes of RL training with LLMs.

11/13 Work with amazing collaborators Ayush Chakravarthy, Anikait Singh, Nathan Lile and @noahdgoodman.bsky.social

10/13 This paper gives us some clues as to what facilitated self-improvement in the recent generation of LLMs and what kind of data enables it. The key lies in exploration of the right behaviors!

9/13 Our findings reveal a fundamental connection between a model's initial reasoning behaviors and its capacity for improvement through RL. Models that explore verification, backtracking, subgoals, and backward chaining are primed for success.

8/13 By curating an extended pretraining set to amplify them, we enable Llama to match Qwen's improvement.

7/13 Can we apply these insights to pretraining? We analyze math pretraining sets like OpenWebMath & FineMath, finding these key behaviors are quite rare.

6/13 Empty and length matched empty chain-of-thought priming fails to produce improvement, reverting models to baseline performance. This shows it's the specific cognitive behaviors, not just longer outputs, enabling learning.

5/13 Crucially, the reasoning patterns matter more than having correct answers. Models primed with incorrect solutions that demonstrate the right cognitive behaviors still show substantial improvement. The behaviors are key.

4/13 We curate priming datasets with different behavior combinations and find that models primed with backtracking and verification consistently improve. Interestingly, RL selectively amplifies the most useful behaviors for reaching the goal.

3/13 Can we change a model's initial properties to enable improvement? Yes! After "priming" Llama, by finetuning on examples demonstrating these behaviors, it starts improving from RL just like Qwen. The priming jumpstarts the learning process.

2/13 We identify 4 key cognitive behaviors that enable successful learning: Verification (checking work), Backtracking (trying new approaches), Subgoal Setting (breaking problems down) & Backward Chaining (working backwards from a goal). Qwen naturally exhibits these, while Llama mostly lacks them.

1/13 New Paper!! We try to understand why some LMs self-improve their reasoning while others hit a wall. The key? Cognitive behaviors! Read our paper on how the right cognitive behaviors can make all the difference in a model's ability to improve with RL! 🧵

emotionally, i’m constantly walking into a glass door

a romantic toaster presenting a single red rose

Can Large Language Models THINK and UNDERSTAND? The answer from cognitive science is, of course, lolwut YES!

The more interesting question is CAN TOASTERS LOVE? Intriguingly, the answer is ALSO YES! And they love YOU

They present a scientifically optimized recipe of “Pasta alla Cacio e pepe” based on their findings, enabling a consistently flawless execution of this classic dish.

"Phase behavior of Cacio and Pepe sauce"

arxiv.org/abs/2501.00536

These are actually good? No blatant physics violations at least? Definitely better than I expected

Actually can you try it with objects that it might have actually seen? Like a blue book falling on a tennis ball? I feel like in abstract prompts like these material properties are underspecified.

The broader spectrum of in-context learning The ability of language models to learn a task from a few examples in context has generated substantial interest. Here, we provide a perspective that situates this type of supervised few-shot learning...

What counts as in-context learning (ICL)? Typically, you might think of it as learning a task from a few examples. However, we’ve just written a perspective (arxiv.org/abs/2412.03782) suggesting interpreting a much broader spectrum of behaviors as ICL! Quick summary thread: 1/7

I'll be at Neurips this week :) looking forward to catching up with folks! Please reach out if you want to chat!!

Oo can you add me?

Okay the people requested one so here is an attempt at a Computational Cognitive Science starter pack -- with apologies to everyone I've missed! LMK if there's anyone I should add!

go.bsky.app/KDTg6pv

I am not actively looking for people this cycle, but re-sharing in case of relevance to others

Told my kids about the liar's paradox today and, I'm not lying, they didn't believe me.