Bing He

[New technical blog] 🚀From Llama-2 Replication to DeepSeek V3.2 & R1 When we wait for new model from DeepSeek, how about revisiting DeepSeek's 13 papers for the complete evolution of DeepSeek’s base + reasoning models?

Check our article on why harness matters and key finding of self-evolving agent

We separate the two phases and search over planning and execution individually with reward-guided pruning — achieving better accuracy with more efficient test-time compute allocation.

Our work on LLM Reasoning "A Reward-Guided Dual-Phase Framework for Adaptive Inference-Time Reasoning" (arxiv.org/pdf/2509.25420) has been accepted at ACL 2026 Findings! Key insight: LLM reasoning has a natural planning-execution structure, but existing tree-search methods ignore

@HenryL_AI @amazon Huge milestone for self-improving agents. A-Evolve makes it possible to turn a base agent into a continuously improving system with almost no manual overhead.

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

Take a look at our self-evolving agent! The self-evolution era.

Today I used Claude 4.6 Opus + its 1M context window to finally crack a long-standing LLM experiment environment config nightmare—and got my agentic RL training pipeline running (custom VERL, SGLang, vLLM, endless deps/configs). Big “aha” moment for Vibe/Agentic Coding #LLM #RL.

📌Blog link: binghe2727.github.io/Revisit-DeepSe… Authors: Bing He, zhan shi, Rui Sun, Hanqing Lu #LLM #DeepSeek #MoE #ReasoningModels #PostTraining #RLHF #AIResearch #GenAI

If you’re working on LLM training, inference efficiency, MoE architectures, or reasoning post-training, we hope this helps you connect the dots across the full trajectory.

[New technical blog] 🚀From Llama-2 Replication to DeepSeek V3.2 & R1 When we wait for new model from DeepSeek, how about revisiting DeepSeek's 13 papers for the complete evolution of DeepSeek’s base + reasoning models?

@YuZheng0404 Hi Yu, big congrats. Welcome to CA.

2024.6.14 🎓Dr. Zheng🎓 My big dream：Will still have the ability to lie down in a rocking chair “eventually”.

@KaiShu0327 Big congrats, Kai. Looking forward to seeing what is coming out from your research group.

I am thrilled to share that I received the NSF CAREER Award. We will explore the real world challenges faced when deploying trustworthy AI algorithms, especially in the age of LLMs. This project will focus on fairness AI under generalization, privacy and robustness challenges.

@eric_ma_23 @srijankedia @WebSciConf

we conduct a large-scale data-driven analysis and find that polite and evidenced correction works well. More can be found at our paper and shared code&data. Arxiv:arxiv.org/abs/2403.04852 Code and Data: github.com/claws-lab/resp…

Many people correct those who spread online misinformation. But how are users responding to it? Will correction have a backfire effect instead? In our new paper at ACM Web Science 2024 - “Corrective or Backfire: Characterizing and Predicting User Response to Social Correction",

@UsmanNaseem87 Big congrats, Usman. This is amazing. Looking forward to the research output.🎉

Excited to announce that my application to OpenAI's API Researcher Access Program has been accepted! Grateful for this opportunity to delve deeper into cutting-edge AI research #OpenAI #AIResearch 🚀🔍

@SRoyLee Big congrats! Relevant to my research area. Looking forward to reading it😄

📣 Our paper "RumourGraphXplainer: Do Structures Really Matter in Rumour Detection?" is accepted in IEEE TCSS! Congrats to Daniel (his first paper). We will put up the pre-prints soon! (ps: it does matter, but not as much as we thought) 🥳

@LiJundong @Yushun_Dong Big congrats! @LiJundong @Yushun_Dong

🌟 Proud Advisor Moment 🌟 Glad to share that my Ph.D. student Yushun Dong @Yushun_Dong will join the Florida State University, CS Department as a Tenure-track Assistant Professor this Fall! He will be the second Ph.D. student in my group to join academia after graduation.

@SRoyLee Big congrats, Roy! Looking forward to seeing you in Singapore at the conference!

📣 We have two memes papers accepted for #Web4Good #TheWebConf2024! 📷 "MemeCraft: Contextual and Stance-Driven Multimodal Meme Generation" and "Modularized Networks for Few-shot Hateful Meme Detection" We will put up the pre-prints soon! Congrats to the team! 🥳