Rohit Agarwal

*at dinner in sf* “the filling in this masala dosa isnt distributed evenly. they shouldve used kubernetes”

@anpaure @yong_zhengxin cot isn't meant to be read... a true intelligence can take this complicated, long thinking trace and condense it into a response that's short but understandable

@yong_zhengxin ok but even that aside, "intelligence is compression" has absolutely nothing to do with cot length???

related to this: one weird asymmetrical fixation ai folks—who work on adaptive thinking—are about compression of CoT: make model think faster/shorter etc. i find immense value on getting models to think longer. like I’d want a model that can solve 1+1 both quickly and being able to come up with novel axioms (think about that 162 page of proof in Principia Mathematica)

@HanGuo97 Claude Code goal mode is a good start, or even just trying a simple genetic algorithm with AI refiner could work. Unfortunately none of the platforms are too mature yet..

@Rohit_Writes Good question, any good places to try it out?

LLM training is built on fast MatMuls. But many surrounding ops still run as memory-bound kernels. CODA reparameterizes them to hide in the matmul’s shadow, fused into its epilogue before results leave the chip. Bonus: LLMs can write fast CODA kernels too (approaching SoLs).

@alokbishoyi97 @PrincetonPLI Check out the paper, they do it. good base + some structured composable primitives = LLMs can write new kernels themselves would be especially effective with an autoresearch loop. we already have software whose APIs are easy for AI to learn how to use (MCPs on their way out...)

@Rohit_Writes @PrincetonPLI curious - could you elaborate how would that look like and what would it unlock ?

Great work with some @PrincetonPLI collaboration. Very bullish on AI-kernel codesign—I think maybe we should modify kernels to be easier for autoresearch etc. to iterate on. We’re already doing this with code.

@VictorTaelin You mean skills and goals??

I discovered a new joy in life. Don't ask Codex to do stuff. Ask Codex to ask Codex to do stuff. Rejoice as you watch it handling and correcting all the dumb shit that it does and that you'd be dealing with otherwise

These (quite general) systems can solve open problems now! Amazing stuff! The collaboration of math academics with engineers/ML researchers is truly amazing!

My blog post on our recent alignment research, with an algorithmic consequence: lesswrong.com/posts/twWfH9TG…

@sama @icanvardar leveragemaxxers unite :)) rohaga.substack.com/p/obligatory-m…

@icanvardar it's not that i'm not an ai, it's that i'm leveragemaxxing

tell me you’re not an ai without telling me you’re not an ai

blog post by my students on our new alignment paper: minregret.com/2025/05/07/age…

This is a great starting point to lots of future agentic alignment work to be done. Fundamentally, I think you don’t have to choose between capabilities and safety research. Sometimes there’s ways to do both.

@ChenguangWang Is the NeurIPS template really okay? It doesn't seem to indicate this on the workshop page.

We strongly encourage you to submit your new work or findings, and we also welcome you to submit your existing work directly using either the ICML or NeurIPS template. 📝 Submit here: openreview.net/group?id=ICML.… Workshop details: agentwild-workshop.github.io/icml2026/ We welcome contributions on a wide range of topics related to AI agents, including but not limited to: 🛡️ Agentic safety and alignment 🔒 Agent security, privacy, and robustness 💭 Agentic hallucination and factuality 🔬 Agentic interpretability and transparency ⚖️ Agentic fairness and bias ⚙️ Agent systems and infrastructure 🔧 Post-training and adapting agents 🧠 Agent planning, reasoning, and decision-making 💾 Agent memory systems and life-long learning 📊 Evaluating and benchmarking agents 🖥️ Multimodal and computer-use agents 🤝 Multi-agent coordination and long-horizon safety 🔬 Autonomous research agents and scientific discovery 🌐 Interdisciplinary agentic considerations 🏛️ Ethics, society, and governing of agents

🚨Last call for papers! The extended submission deadline for the Second Workshop on Agents in the Wild: Safety, Security, and Beyond at ICML 2026 is now May 8, 2026, AoE for both regular and short paper tracks! 🌟 We’re excited to host outstanding speakers, including @Yoshua_Bengio (@Mila_Quebec, @LawZero_, @UMontreal), @dawnsongtweets (@UCBerkeley, @BerkeleyRDI), @WeiWang1973 (@UCLA), @bemikelive (@databricks), @JiantaoJ (@nvidia, @UCBerkeley), @Eric_Wallace_ (@OpenAI), @astonzhangAZ (@OpenAI), and @ce_zhang (@togethercompute, @UChicago).

@HazanPrinceton I think a grad student might have snuck in there...

Finished teaching, time for end-of-year toast with my excellent batch of undergrad advisees

Excited to be at #ICLR2026 this week! I’m presenting “A New Approach to Controlling Linear Dynamical Systems”, with my coauthors @gon_buzaglo , Sofiia Druchyna, and my advisor @HazanPrinceton. 📍 Pavilion 4 — Poster P4-#4503 🕒 Apr 23, 3:15–5:45 pm (Rio local time)

@prfsanjeevarora @PrincetonPLI Excited to see these techniques generalize to future works!

From @PrincetonPLI New, low-cost, and fully open "deep thinking" for LLMs to solve competition math questions. State-of-the-art 64.8% at IMO-Bench (Advanced) for $31/question, and 52% for $3/question. By contrast, DeepSeek’s published recipe allowed up to $3K/question for 61.7%! Key idea: how to nudge an LLM to escape its “Cognitive Well” arxiv.org/abs/2602.16793 @xingyudang, Rohit Agarwal, @liam_fowl, Rodrigo Porto, @anirudhg9119