Sirui Li

PhD student at MIT.

[Ao Qu @ao_qu18465]

🚀 The era of autonomous multi-agent discovery has begun.

Most “self-evolving” scientific discovery frameworks are still tightly constrained:
LLMs often just perform one-step mutations inside fixed evolutionary search loops.

But that is not real autonomy.
Agents still cannot truly decide:
🔍 what to explore
🧠 what knowledge to store
♻️ which past attempts to reuse
🧪 when to test

With CORAL, we ask:

❓ What happens if we give agents much more autonomy to explore the scientific frontier?

💡 Our answer:

A single autonomous agent already outperforms fixed evolutionary search.

But the bigger leap comes when multiple autonomous agents form a research community:

🤝 They explore different directions
🧠 accumulate reusable knowledge and skills
💬 communicate with each other
🌍 and push the frontier together

We introduce CORAL, the first framework for autonomous multi-agent evolution for open-ended discovery.

🥇 Across 10+ tasks in algorithmic discovery, system optimization, and kernel engineering from Frontier-CS, ADRS, AlphaEvolve, etc, CORAL achieves SOTA and improves search efficiency by 3–10× over prior fixed evolutionary-search frameworks.

🔬 Why does autonomy help?

Our analysis shows two main reasons:

🧪 Local verification: agents run local tests before expensive evaluations, which is especially powerful for coding tasks.

♻️ Knowledge reuse: on knowledge-intensive tasks like polyominoes and kernel engineering, agents create and reuse knowledge artifacts at far higher rates than on simple tuning/search tasks like circle packing.

✨ Even more exciting:
Over 50% of multi-agent breakthroughs come from building on other agents’ discoveries.
Multi-agent exploration is also far more diverse than single-agent search.

We believe CORAL opens up an exciting new space for automated discovery systems.

📬 If you are interested in collaborating, let’s talk.

📄 Paper: https://t.co/rbbLbkyihL
💻 Code: https://t.co/5fCTnqDzct
💡AlphaXiv: https://t.co/MK6rVpVc5A

#agentic #llms #selfevolvingagent #multiagent #autoresearch #alphaevolve

[Microsoft Research @MSFTResearch]

OptiMind is a small language model that converts business operation challenges, described naturally, into mathematical formulations that optimization software can solve. It reduces formulation time & errors & enables fast, privacy-preserving local use: https://t.co/lrJSExhivf

[Cathy Wu (Hiring PhD/postdocs) @wucathy]

👼 As an applied RL researcher, this is the most optimistic I have been about RL in years. It feels like seeing the light at the end of the tunnel when RL training starts working reliably. Without a ton of compute or tuning. Very excited for what is to come. Here is what we did👇

[Cathy Wu (Hiring PhD/postdocs) @wucathy]

Reinforcement learning (RL) is surprisingly brittle to contextual variations in tasks. Our new method (NeurIPS 2024) for solving contextual RL problems achieves 5-50x better sample efficiency on standard & traffic benchmarks. Featured today by MIT news! https://t.co/dR1xK5tgoT

[IEEE Transactions on Robotics (T-RO) @IeeeTro]

Researchers @MITLIDS and @IllinoisISE proposed a hybrid system #stabilityanalysis framework that offers a explanation to the #trafficbreaks (rolling roadblocks) emergent behavior for a single #autonomousvehicle to simultaneously stabilize multiple lanes.
https://t.co/yF5nuSLjUg

[Cathy Wu (Hiring PhD/postdocs) @wucathy]

Our NeurIPS 2023 paper that learns to intelligently configure MILP separators is featured on the MIT front page! https://t.co/AMbxm7eWQ0 Our method accelerates open source solver SCIP up to 35-70% on MILP benchmarks ⚡⚡, and even speeds up the state-of-the-art solver Gurobi. 🧵

[Cathy Wu (Hiring PhD/postdocs) @wucathy]

A step towards the promise of safer roadways: in a @IeeeTro article published last month, we analyzed how cooperative intelligence and ideas from air traffic control can drastically improve reliability of autonomous vehicles. https://t.co/SFMzeRLPc2

[Cathy Wu (Hiring PhD/postdocs) @wucathy]

tl;dr: Benchmark tasks are far from what’s needed for real-world RL. We show in #NeurIPS2022, that RL methods no longer outperform non-RL methods in traffic signal control when we consider a more fully specified task representation.
Paper: https://t.co/Yvzgy3gYGv

A thread: 1/N

[Cathy Wu (Hiring PhD/postdocs) @wucathy]

Our recent NeurIPS 2021 Spotlight Talk on machine learning for speeding up vehicle routing was featured by MIT News! We devise a strategy which accelerates the best algorithmic solvers by 10-100x for large sets of cities.
https://t.co/lhZmxiqfoS @hoaxingz @SiruiLi80955620

[Cathy Wu (Hiring PhD/postdocs) @wucathy]

Check out our NeurIPS Spotlight + poster on tackling large vehicle routing problems with ML at the 7:30-9pm EST session today! Our work accelerates state-of-the-art heuristic solvers by 10-100x. ⚡️⚡️ https://t.co/HwcIepbZ0t https://t.co/NoARSEZU0L @SiruiLi80955620 @hoaxingz