Zhikun Xu

Legit statement from someone who once encouraged their employees to tokenmax.

@hendav136 🫡 (pretraining for what?

Hen, first year PhD, pretraining an LLM with academia compute for the first time. Wish me luck.

@wenhaocha1 Indeed. Tokens are becoming electricity. Acting is more expensive.

The cost and speed of tokens don’t matter; the only thing that matters is performance. Today, the main cost of coding agents lies in rollouts, but once agentic systems permeate every industry, tool calls will become the dominant source of latency and expense. The cost of making a single mistake will far exceed the cost of thousands of API calls.

That last part resonates with what motivated our ongoing project, ReuseRL: if a capability only internalizes when it is compressible, the real question becomes what gets internalized. We go looking for the atoms: the small set of reusable skills a model can absorb and build on.

Treating reasoning and acting as two tools for one job folds many debates (long context vs RAG, think vs act, etc.) into a single allocation question. And framing the internalization–externalization boundary as the next design question feels exactly right. Really inspiring read!

@Swarooprm7 We may need to define the "machine creativity/novelty" first. What about searching for some counterexamples in formal math? Concepts are created to simplify things, facilitate abstraction, and improve human understanding (efficiency, etc.). Why do AI models need novel concepts?

How can we accurately evaluate whether an AI model has generated a genuinely novel concept? Is there a widely accepted benchmark for measuring machine creativity or invention?

@wzenus @YejinChoinka @jiajunwu_cs @ManlingLi_ @LINJIEFUN @chi_gui_1 @DeimosGN @qineng_wang @James_KKW @shiqi_chen17 @zhengyuan_yang Well-deserved👍

RAGEN-2 is selected as ICML oral! Congrats and great appreciation to all collaborators!!

✨Check out the paper to learn more! This paper was done by our intern @Zijun0916 last summer. A big thank you to my labmate @XiaoYe1170354 and our advisor @BenZhou96 for the support and guidance!

📈 The results: Consistent gains over vanilla baselines, including up to +9.3% on in-domain Textbook problems and +9.6% on out-of-domain TheoremQA. We also did ablation experiments to show the results are consistent with different models and across different benchmarks.

LLMs memorize massive amounts of text, but can they actually apply this knowledge conceptually? 🤔 Our #ICLR 26' paper from the ARC Lab probes this in math reasoning! "CORE: Concept-Oriented Reinforcement for Bridging the Definition-Application Gap..." 🔗 openreview.net/forum?id=pRSRi…

@HBX_hbx @QuYuxiao Besides QuestA, there are also many other related works using a similar idea from last year: BREAD(arxiv.org/abs/2506.17211), Scaf-GRPO(arxiv.org/abs/2510.19807), and CORE(arxiv.org/abs/2512.18857). "Guided prefix" could be partial oracle solutions, problem-related concepts, etc.

@QuYuxiao Excellent work! I am curious what the difference is between POPE and QuestA, since both leveraging prefix of references to conduct curriculum learning.

🚨 NEW PAPER: “POPE: Learning to Reason on Hard Problems via Privileged On-Policy Exploration”! ❓ How do we train LLMs with RL on hard problems when the model never gets a single correct rollout? 💡 Short answer: standard RL is stuck. We show why, and introduce POPE to break this deadlock. 🧵[1/N]

@denny_zhou Fake agi: Gemini🤖 Real agi: daughter🧒

My daughter asked a fun question: are humans actually aliens on this planet? Her argument: the intelligence gap between humans and every other species is massive. No other life has built civilization. Dinosaurs ruled Earth for hundreds of millions of years and still produced no civilization. Yet humans did it in a few hundred thousand. My take: we probably aren’t aliens. Look at AI. The gap between AI 10 years ago and today can feel like dinosaurs vs humans. Back then it was mostly shallow mapping from input to output with statistical models (e.g., image/text classification ). Now AI can “think” for hours and solve genuinely hard problems.

Highly resonate about this. Conceptual reasoning should be very important for LLM reasoning.

@ShashwatGoel7 👀 arxiv.org/abs/2310.12962

Was wondering whether GRPO style RL is "only a few tokens deep"... Intuitively, we take a next token predictor, and slightly upweigh some tokens, s.t. NTP leads to success Found this interesting ICLR sub preliminarily indicating this hypothesis is true: openreview.net/forum?id=8vWIX…

First time at #NeurIPS2025! I’ll be in San Diego from Dec 1–6 and would love to make new friends, grab some tea🍵, and discuss LLM reasoning (math & cognition-inspired), RL, and more! Feel free to DM!

@zjasper @hyperbolic_labs github.com/THUKElab/COUNT… 👀how about testing it with some counterexamples?

We got deepseek-math-v2 running on 8xH200 node on @hyperbolic_labs on-demand GPU platform. Feel free to reply with any math problems that you want to know and I can share the answers. An exciting time to own the brain of one of the best mathematicians!

@taiwei_shi NeurIPS+1! Hope to talk with u🙋

I will be at NeurIPS next week. ✨ Looking forward to catching up with friends old and new. Happy to talk about reinforcement learning, test-time training, long-term memory, self-evolving agent, and so much more! DMs are open. 😄