Hi there, I’m currently a PhD student in computer science at Emory University.
United States
Joined August 2024
- Tweets7
- Following26
- Followers8
- Likes14
Zewen retweeted
Can we train language models to propose better ideas to improve LM—not just run them?
Introducing 𝐋𝐌𝟐𝐋𝐌: Language Models Learn to Improve Language Models. We train a 27B model with RL to propose improvements to language models. On held-out autoregressive and diffusion LM development tasks, it outperforms 𝐆𝐏𝐓-𝟔-𝐬𝐨𝐥 —at 33% lower data-generation cost. It also generalizes to a model family unseen during training.
𝐓𝐡𝐞 𝐛𝐨𝐭𝐭𝐥𝐞𝐧𝐞𝐜𝐤: evaluating every proposal with a real training run makes RL prohibitively expensive.
𝐎𝐮𝐫 𝐚𝐩𝐩𝐫𝐨𝐚𝐜𝐡: a Research World Model (RWM) predicts experimental gains as a cheap reward signal. We periodically execute the current policy’s top proposals and add the measured outcomes to the RWM’s context, keeping its predictions calibrated as the policy evolves.
This is a step toward cheaper, more reliable feedback loops for recursive self-improvement (RSI).
Zewen retweeted
Current Recursive Self-Improving AI misses a 𝐩𝐮𝐛𝐥𝐢𝐜 datapoint: AI trains a frontier-class model. Today we share 𝐀-𝐄𝐯𝐨𝐥𝐯𝐞-𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 that 𝐚𝐮𝐭𝐨-𝐩𝐨𝐬𝐭-𝐭𝐫𝐚𝐢𝐧𝐞𝐝 𝐚 𝟑𝟎𝐁 𝐍𝐞𝐦𝐨𝐭𝐫𝐨𝐧 model with no human-in-the-loop.
~10²× larger than prior public autoresearch (GPT2, ~124M) demos.
It achieved 0.86 on NVIDIA Nemotron Reasoning Challenge (snapshot 6/1 as of writing the tech report) vs human #1's 0.87.
Same system has been run at 𝟏𝟐𝟎𝐁 𝐚𝐧𝐝 𝟓𝟓𝟎𝐁 scale; effectiveness claim deferred at those scales pending human anchor.
But the important part is not the score: the loop didn't just execute — it discovered a flaw in its own optimization, and inverted its own objective mid-campaign.
👇Tech report and website
Our internal data shows Claude is accelerating AI development—a possible path to recursive self-improvement, or AI autonomously building a more capable successor.
It’s happening faster than we thought, and the implications deserve greater attention. anthropic.com/institute/recu…
Zewen retweeted
When you building a self-improving agent, two natural questions emerge: 𝐐𝟏: 𝐖𝐡𝐢𝐜𝐡 𝐦𝐨𝐝𝐞𝐥𝐬 𝐩𝐫𝐨𝐝𝐮𝐜𝐞 𝐭𝐡𝐞 𝐛𝐞𝐬𝐭 𝐡𝐚𝐫𝐧𝐞𝐬𝐬 𝐮𝐩𝐝𝐚𝐭𝐞𝐬? 𝐐𝟐: 𝐖𝐡𝐢𝐜𝐡 𝐦𝐨𝐝𝐞𝐥𝐬 𝐛𝐞𝐧𝐞𝐟𝐢𝐭 𝐦𝐨𝐬𝐭 𝐟𝐫𝐨𝐦 𝐡𝐚𝐫𝐧𝐞𝐬𝐬 𝐮𝐩𝐝𝐚𝐭𝐞𝐬?
Our new paper has counter-intuitive answers to both: they decouple from model capability, in opposite ways.
Q1 (who produces good updates): the updater's base capability barely matters.A 9B model (Qwen3.5) produces harness updates that match Claude Opus 4.6's. Best vs worst evolver gap ≤3.1pp.
Q2 (who benefits most): non-monotonic.Mid-tier solvers benefit the most. Strong-tier hits ceiling. Weak-tier benefits LEAST despite the most headroom — failing at two layers: skill activation (25% vs ~96% for strong) and adherence drift across trajectory (~4x steeper).
Tested across 7 evolver models × 6 solver agents × 3 agentic benchmarks (SWE-bench Verified, MCP-Atlas, SkillsBench).
Implication: don't pay frontier prices for both halves of the loop. Put capability budget on the agent (solver), not the evolver.
Launch Post🧬 A-Evolve: The PyTorch Moment for Self-evolving AI
Today we at @amazon launch the universal infrastructure that turns any agent into a self-improving SOTA agent — zero human intervention.
You give it a base agent → it returns a continuously evolving Top-10 agent.
3 lines of code. 0 hours of manual harness engineering:
🟢 MCP-Atlas → 79.4% (#1) +3.4pp
🔵 SWE-bench Verified → 76.8% (~#5) +2.6pp
🟣 Terminal-Bench 2.0 → 76.5% (~#7) +13.0pp
🟡 SkillsBench → 34.9% (#2) +15.2pp
Thanks @binghe2727 @YisiSang @sammyershi @linminhua16 for the contribution!
#AgenticAI #AEvolve #SelfImprovingAgents
Zewen retweeted
Very good advice on self-improving agents.
(bookmark it)
This is something I am seeing in my own experiments with coding agents and harnesses for long-horizon tasks.
What I have found is that stronger models do not always evolve better agents.
The current believe in self-evolving agents is that a bigger model writes better prompt and skill edits, so devs put their best model in the evolver seat.
New research shows that intuition is mostly wrong.
The work separates two abilities that usually get conflated. Producing harness updates stays flat across model capability, so Qwen3.5-9B writes edits roughly as good as Claude Opus 4.6. Benefiting from those updates follows an inverted-U that peaks at mid-tier models, while weak models fail to even activate the edits and strong models have little headroom left.
This is important to understand as it tells you where to spend. Put a cheap model on the evolver and your expensive model on the solver, because the gains land solver-side, not evolver-side.
Paper: arxiv.org/abs/2605.30621
Learn to build effective AI agents in our academy: academy.dair.ai/
Zewen retweeted
Thrilled to share that this survey on Graph in Epidemiology has been accepted by #KDD2024. Do not miss this work if you want to learn more about the applications of graphs and GNNs in epidemic modeling! #Epidemiology #GNN #MachineLearning
We recently released a review on Graph Neural Networks (GNNs) for Epidemiology (arxiv.org/pdf/2403.19852.pdf). Epidemiological tasks involve lots of graph data, where GNNs have shown strong predictive power, aiding in the monitoring and prevention of disease spread #Epidemiology #GNN
Zewen retweeted
We’ll also be showcasing EpiLearn (github.com/Emory-Melody/EpiL…), our Python Library for Machine Learning in Epidemic Modeling, which has been accepted to the KDD epiDAMIK Workshop @EpidamikW on August 26.
Zewen retweeted
We recently released a review on Graph Neural Networks (GNNs) for Epidemiology (arxiv.org/pdf/2403.19852.pdf). Epidemiological tasks involve lots of graph data, where GNNs have shown strong predictive power, aiding in the monitoring and prevention of disease spread #Epidemiology #GNN