Tao Feng (Attending NeurIPS 2025 in San Diego)
@taofeng_uiuc
Followers
181
Following
118
Media
38
Statuses
139
PhD student of UIUC, Looking for Summer 2026 Research Intern positions in ML/AI
Champaign
Joined April 2025
🔥 We're open-sourcing the First unified LLM routing library! 🧭 LLMRouter: 16+ routers in ONE framework ✅ Single-round routers ✅ Multi-round routers ✅ Agentic routers ✅ Personalized routers Code: https://t.co/vsJ1yF5X5I Project Page: https://t.co/QroCqIUguu
8
22
112
Thread (8/8) 🙌 Huge thanks to all co-authors: Tao Feng, @haozhen_ntu , @lei_zijie , Haodong Yue, Chongshan Lin And our advisor @youjiaxuan for invaluable guidance! ⭐
github.com
LLMRouter: An Open-Source Library for LLM Routing. Contribute to ulab-uiuc/LLMRouter development by creating an account on GitHub.
0
0
2
Thread (7/8) 👤 Personalized Routers: Your preferences matter! Learn from YOUR interaction history to route queries based on YOUR unique needs. One size doesn't fit all—routing shouldn't either! 🎨
1
0
1
Thread (6/8) 🤖 Agentic Routers: For complex, multi-step tasks! Break down hard problems, route each step to the best model, and combine results. LLM routing meets agent workflows! 🔗
1
0
1
Thread (5/8) 🔄 Multi-Round Routers: Think before you route! RL-powered reasoning that aggregates multiple LLM responses iteratively. Not just routing—reasoning about routing! 🧠
1
0
1
Thread (4/8) 🎯 Single-Round Routers: One query, one decision! From classic ML (KNN, SVM, MLP) to neural approaches (contrastive learning, graph neural networks, fine-tuned LLMs) 10+ methods ready to use! ⚔️
1
0
1
Thread (3/8) 💸 Why pay GPT-4 prices for "What's the weather?" Smart routing = Right model for the right query Simple query → cheap model Complex query → powerful model Save 30-50% cost without sacrificing quality! 🎯
1
0
2
Thread (2/8) 🛠️ Complete toolchain out of the box! Unified CLI for train / infer / chat Gradio-based interactive UI 11 benchmark datasets ready to use Full data generation pipeline included Research to production in minutes! 🚀
1
0
1
Thread (1/8) ⚡ One command. 16+ routers. Zero hassle. Train, infer, or chat with ANY router using a unified CLI. Switch between KNN, GraphRouter, Router-R1... with just one flag change! No more reimplementing from papers. Just plug and play! 🔌
0
0
2
Reinforcement Learning for Self-Improving Agent with Skill Library SAGE framework uses Sequential Rollout: agents learn skills across task chains and reuse them. Results: 8.9% higher completion, 26% fewer steps, 59% fewer tokens on AppWorld.
8
27
145
My default workflow! This setup achieves best possible task accuracy after fixing constraints on the inference process. PDR achieves pareto frontier: "draft in parallel → distill to a compact workspace → refine" https://t.co/4Sca6dG1Uo
One workflow that seems to work really well on claude code is: Parallel Design > Distill/Summarize > Parallel Refine > implement Spawn 16 agents to design / plan topic X with detailed plan on integration Ask to spawn 16 agents to summarize each others work and check for
6
49
331
How can an imperfect agent learn to leverage help from an expert effectively and efficiently? We build a comprehensive benchmark and gain interesting insights into existing methods in a super challenging setting of this problem, where no supervision is given and test tasks are
✨ Ending the year with great news! Work from my internship at @UCBerkeley's @CHAI_Berkeley is accepted at @TmlrOrg 🥳 We study how to learn when to rely on a strong vs. weak agent, the core idea behind the YRC-Bench.
1
2
5
This is exactly the direction: Turning past experiences into reusable *skills* (externalised meta-learning). Our paper mines “how-to” reasoning into a skills library + shared workspace; at test time the model reads that memory in-context to improve. https://t.co/xLcwezJAEm
Learning Skills with Deepagents Reflection over past trajectories is one general approach for agent learning. Here, we show how Deepagents-CLI can use reflection to create new skills. 📹: https://t.co/cQc4KCIxgw
15
48
508
🥈 Silver Medal at IOI 2025 & Outperforms DeepSeek-R1-0528 on LiveCodeBench. Instead of mixing different tasks together, we scale *Cascade RL* to develop general LLMs in curriculum (RLFH -> Instruct -> Math -> Code -> SWE). So many learnings, check out our report!👇
🚀 Introducing Nemotron-Cascade! 🚀 We’re thrilled to release Nemotron-Cascade, a family of general-purpose reasoning models trained with cascaded, domain-wise reinforcement learning (Cascade RL), delivering best-in-class performance across a wide range of benchmarks. 💻 Coding
5
43
225
Think of orchestration as search over thoughts. Then train the model to match that orchestration. This allows shifting the accuracy↔latency/context tradeoff by compressing search into weights.
I'm glad this paper of ours is getting attention. It shows that there are more efficient and effective ways for models to use their thinking tokens than generating a long uninterrupted thinking trace. Our PDR (parallel/distill/refine) orchestration gives much better final
7
32
260
Manual tuning of multi-agent systems is a massive bottleneck. 🛑 We’re fixing that. In our new ADRS blog, we show how OpenEvolve uses MAST-guided evolutionary search to autonomously change agent system architecture, boosting reliability by nearly 7x 📷🚀
🎯 AI evolves better AI agents, boosting reliability scores by 7x [ADRS Blog #7] We explore automating Multi-Agent System (MAS) design. By combining OpenEvolve with MAST, we let the code rewrite itself—turning a fragile baseline into a robust architecture in just 46 iterations!
5
16
82
What if video models could understand what’s actually happening step by step, instead of just matching high level text? This paper introduces a Task Step State framework that grounds procedures in observable object states, not just abstract “task” or “step” labels. By
2
11
77
What if AI models could collaborate in their own "thought language" instead of just trading text? This research introduces LatentMAS, a framework where LLM agents share hidden embeddings directly, cutting token use by up to 83% and boosting accuracy by 14.6%—all without extra
5
32
151
Action chunking is an essential part of modern imitation learning. But it's still unclear how to use it with RL. In our new work, Decoupled Q-Chunking, we show how using larger chunks for the critic and short chunks for a reactive actor can produce state-of-the-art RL results!
15
61
644
Thanks @_akhaliq for sharing our work!
0
2
53