Using LLMs to build AI scientists is all the rage now (e.g., Google’s AI co-scientist [1] and Sakana’s Fully Automated Scientist [2]), but how much do we understand about their core scientific abilities?.We know how LLMs can be vastly useful (solving complex math problems) yet

RT @xiangyue96: People are racing to push math reasoning performance in #LLMs—but have we really asked why? The common assumption is that i….

RT @gneubig: What will software development look like in 2026?. With coding agents rapidly improving, dev roles may look quite different. M….

RT @ChengleiSi: Are AI scientists already better than human researchers?. We recruited 43 PhD students to spend 3 months executing research….

I'm thrilled to share that I've moved to Pittsburgh and joined NeuLab at CMU as a research intern this summer, advised by @gneubig! I'll also start my PhD @LTIatCMU this fall. Feel free to reach out if you're interested in chatting about multi-agent systems, LLMs for scientific.

RT @EchoShao8899: 🚨 70 million US workers are about to face their biggest workplace transmission due to AI agents. But nobody asks them wha….

RT @AnthropicAI: New on the Anthropic Engineering blog: how we built Claude’s research capabilities using multiple agents working in parall….

RT @yikewang_: LLMs are helpful for scientific research — but will they continuously be helpful?. Introducing 🔍ScienceMeter: current knowle….

RT @ChengleiSi: This year, there have been various pieces of evidence that AI agents are starting to be able to conduct scientific research….

RT @JiahaoQiu99: The GAIA game is over, and Alita is the final answer. Alita takes the top spot in GAIA, outperforming OpenAI Deep Researc….

[1] [2] [3] (10/10).

Check out more details from our paper: Work done with amazing collaborators Howard Chen (@__howardchen), Dilip Arumugam (@Dilip_Arumugam), and Tom Griffiths (@cocosci_lab)! 🎉. Special thanks Danqi Chen (@danqi_chen) and people from @PrincetonPLI.

In summary, we show that LLMs still struggle to fully utilize passive observations to form hypotheses and allowing active intervention mitigates some of the failure modes. By evaluating their ability to reverse-engineer black-box systems, we offer a principled and controlled.

So, what makes active intervention so effective? It allows the LLM to strategically design its queries to test and iteratively refine its hypotheses about the black-box system. (7/n)

Active intervention effectively mitigates two common failure modes across all three black-box systems we studied:.1) Overcomplication – LLM tends to construct overly-complex hypotheses;.2) Overlooking – LLM neglects observations and draws overly-generic conclusions without

🤔Can LLMs successfully transfer their experiment data and findings to other LLMs?.The answer is not quite! Transferred interventions consistently underperform compared to LLMs conducting their own active interventions. (5/n)

🤔Why do interventions help? Do they result in more informative observations, or is it the process of generating the interventions itself that matters?-- Inspired by the passive-yoked design from human learning studies [3], we find that active learning outperforms passive

🤔How well can LLMs infer the internal mechanisms of black box systems from passive observations? . We looked at three kinds of black box systems inspired by cognitive studies:.1) list-mapping programs.2) rules of formal languages.3) parameters of math equations . Our findings

Doing science requires several skills:.1) Performing inductive reasoning based on passively observed data;.2) Actively interacting with a system to collect informative data and reduce uncertainty about its internal mechanisms;.3) Communicating the results. Based on these

RT @yueqi_song: Humans can perform complex reasoning without relying on specific domain knowledge, but can multimodal models truly do that….