Agentic Post-Training, Self-Improvement | Research Scientist @GoogleDeepMind | Previously PhD @nlp_usc, B.Eng @TsinghuaNLP
Joined May 2021
- Tweets110
- Following955
- Followers1K
- Likes480
Jun Yan retweeted
On behalf of Gemini 4โs modeling team to improve VibeCoding capabilities (spanning WebDev, 3D, games, full-stack, Android dev, and more), we just landed 91.9% on VibeCodeBench!
As the AI race heated up, our release cadence accelerated dramatically. The pressure was immense - we worked day and night, non-stop, to land critical recipe improvements and push our model performance forward!
Today, with Gemini 4, we came back strong!
Huge thank you to my teammates who held the line through countless days and nights without ever complaining, and to our horizontal partner teams and leadership who believed in us and gave us unwavering support. Onward and upward!! ๐
Jun Yan retweeted
Introducing Gemini 4 Argon โ our new frontier model.
Itโs built for complex workflows across coding, enterprise knowledge work, and cybersecurity defense โ rolling out today to a set of trusted testers through our Fairwind Program.
Jun Yan retweeted
We revolutionized environment design by building the "Harness" for environments, not just agents! Static environments bottleneck the growth of LLM agents. When environments can't evolve, agent progress stalls. Enter EnvHarness: a powerful, unified wrapper framework engineered to dynamically adapt environments and unlock rich training signals.
๐ Non-Intrusive Interception
๐ Dynamic Environments
๐งฉ Plug-and-Play Integration
Title: EnvHarness: Awakening Static Worlds for Agent Learning
Paper: arxiv.org/abs/2608.19880
Code: github.com/google-research/eโฆ
Webpage: envharness.com/
Jun Yan retweeted
๐งฉ How can we assign fine-grained credit over long tool-use trajectories and let agents learn from past attempts in agentic reinforcement learning when rewards are no longer verifiable?
Excited to share RubricEM, an RL framework for long-form deep research agents that plan, search, use tools, and write reports without exact answer checks.
๐ Paper: arxiv.org/abs/2605.10899
(1/n)
How do we train LLM agents to be better continual learners that self-evolve after deployment? ๐ง
๐ ๏ธ Introducing SkillOS. We propose an RL recipe for training skill curation that distills wisdom from past experiences via indirect and delayed feedbackโa core piece of building self-evolving agents.
โญ First author Siru is on the industry job market! It has been an incredible experience working with her (both on this and ReasoningBank). She brings deep expertise in agentic post-training, continual learning, and agent memory. Don't miss out on hiring her! ๐
Thanks @_akhaliq for sharing our paper!
We study skill curation as a trainable capability for self-evolving agents. This capability actually is โผ๏ธhard to trainโผ๏ธdue to the long-term evolving setting that only has indirect and sparse learning signals.
In SkillOS, a skill curator operates over evolving Markdown skills like an operating system โ continuously inserting, refining, and organizing reusable experience over time.
With our training recipe, a trained 8B curator can outperform zero-shot frontier models (Gemini-2.5-Pro) for skill curation in self-evolving agent systems.
Jun Yan retweeted
See yall in Seoul!
First ICML thanks to my brilliant intern @MingZhong_ and the team!
Ming is on job market literally right now, and heโs literally the best. If youโre hiring people for coding agents, canโt think of a reason why Mingโs DM is not full of messages. Please reach out!
Arrived in Rio de Janeiro to attend #ICLR2026 ๐ง๐ท I will present three posters:
1. Incentivizing Agentic Reasoning in LLM Judges via Tool-Integrated Reinforcement Learning
Thu, Apr 23, 3:15PM โ 5:45PM, Pavillion 3 (Poster Session 2, #1511)
2. ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory
Fri, Apr 24, 3:15PM โ 5:45PM, Pavillion 3 (Poster Session 4, #1410)
3. Supervised Reinforcement Learning: From Expert Trajectories to Step-wise Reasoning Sat
Apr 25, 3:15PM โ 5:45PM, Pavillion 3 (Poster Session 6, #1315)
Happy to meet and chat about agents, memory, RL, and more!
Jun Yan retweeted
ReasoningBank, a novel agent memory framework, enables LLM agents to continuously learn from both successful & failed experiences. Our evaluation shows that it enhances agent effectiveness, boosting success rates and efficiency. Learn more: goo.gle/4dWrPGb
ALT Workflow of ReasoningBank integrated with an agent during test time.
Jun Yan retweeted
Introducing @NeoCognition, the agent lab for specialized intelligence.
Everyone needs experts, but human expertise does not scale.
Backed by $40M seed funding, we build self-learning agents that specialize across domains to make expertise abundant.
Jun Yan retweeted
๐งGiving your agent unlimited tool calls doesn't make it smarter.
๐กWhy? It lacks 'Budget Awareness'!
Introducing Budget Tracker, a simple plug-in that enables more effective scaling behaviors: higher performance, lower cost.
Paper: arxiv.org/pdf/2511.17006
Jun Yan retweeted
Happy to introduce my internship work at @Google and @GoogleDeepMind, collab w/ @googlecloud.
We introduce TIR-Judge, an end-to-end agentic RL framework that trains LLM judges with tool-integrated reasoning ๐ง ๐ ๏ธ
๐arxiv.org/pdf/2510.23038
#Agents #LLMs #Judges #RL #reasoning
Jun Yan retweeted
๐ง ๐ Excited to introduce Supervised Reinforcement Learningโa framework that leverages expert trajectories to teach small LMs how to reason through hard problems without losing their minds. ๐คฏ
Better than SFT && RLVR.
Read more: huggingface.co/papers/2510.2โฆ
#llms #RL #reasoning
Jun Yan retweeted
New paper from @Google is a major memory breakthrough for AI agents.
ReasoningBank helps an AI agent improve during use by learning from its wins and mistakes.
To succeed in real-world settings, LLM agents must stop making the same mistakes. ReasoningBank memory framework helps agents learn from both successes and failures and turn them into reasoning strategies that generalize.
Traditionally, most agents handle tasks in a stream but do not keep lessons, so they repeat errors.
The system turns each action log into a short memory item with a title, a description, and a concrete next step.
Before a new task, the agent retrieves the most relevant items using embedding search and appends them to the system prompt.
After each run, a separate LLM labels success or failure, and the system writes out strategies and pitfalls.
New items go into the bank with consolidation that removes duplicates and sharpens the content.
MaTTS is a memory aware test time scaling setup that spends extra compute on parallel rollouts or step by step self refinement.
Those extra passes create contrasting traces that make the memory stronger and more precise.
Across WebArena, Mind2Web, and a software engineering benchmark with Gemini 2.5 and Claude 3.7, success rises and average steps drop.
On successful cases, the average step count drops by 2, which means less wandering when the agent is on track.
----
Paper โ arxiv. org/abs/2509.25140
Paper Title: "ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory"
Jun Yan retweeted
๐Drop from Gemini VibeCoding Team:
Why does code functionality correctness NOT necessarily translate to ๐ฏ๐ฒ๐๐๐ฒ๐ฟ ๐๐๐ฒ๐ฟ ๐ฝ๐ฟ๐ฒ๐ณ๐ฒ๐ฟ๐ฒ๐ป๐ฐ๐ฒ ๐ณ๐ผ๐ฟ ๐ฐ๐ผ๐ฑ๐ถ๐ป๐ด tasks?
In ๐ฉ๐ถ๐ฏ๐ฒ๐๐ต๐ฒ๐ฐ๐ธ๐ฒ๐ฟ: arxiv.org/abs/2510.07315, we find that the human preference correlates the best with a composite score of functional correctness and instruction following capability.
To evaluate the instruction following capability of LLM for coding, we propose a taxonomy of ๐ฏ๐ฌ ๐๐ฒ๐ฟ๐ถ๐ณ๐ถ๐ฎ๐ฏ๐น๐ฒ ๐ฐ๐ผ๐ฑ๐ถ๐ป๐ด ๐ถ๐ป๐๐๐ฟ๐๐ฐ๐๐ถ๐ผ๐ป๐ together with their corresponding deterministic checkers.
We evaluated 31 leading LLMs and
show that even the strongest models struggle to comply with multiple instructions and exhibit clear functional regression!
Please take a look at our paper and let us know how you like it!
We also look forward to seeing the community extend the work on real user coding prompts to understand how we can improve ๐ฟ๐ฒ๐ฎ๐น-๐๐ผ๐ฟ๐น๐ฑ ๐๐๐ฒ๐ฟ ๐ฐ๐ผ๐ฑ๐ถ๐ป๐ด ๐ฒ๐
๐ฝ๐ฒ๐ฟ๐ถ๐ฒ๐ป๐ฐ๐ฒ, let us know if you want to collaborate on these topics! (0/1)
Jun Yan retweeted
Extremely honored to be part of this incredible team โค๏ธ! Here is one of my favorite samples ; )
it's an honor to work with such an incredible group of researchers.
massive shoutout to @LiLiunian, @DmytroOk , avi, @ezhang887 , @ryw999 @troyluhman and eric, eric mintun, @bram_wallace @JureZbontar @mmmbchang @gabriel1 @dangengdg and @cmikeh2 for their incredible work on sora 2.
the sora research team is very special.
Jun Yan retweeted
Thanks for sharing our work @arankomatsuzaki ! Really excited about how reasoning-based memory drives and scales for self-evolving agents ๐ซ :
๐ฌReasoningBank stores insights from both successful and failure trajectories;
๐ ๏ธ MaTTS builds on this powerful experience learner, and positions memory-driven experience as a new dimension for agent scaling by leveraging the synergy between memory and TTS.
How do we build personalized LLM agents using memory? Come find out at our #ACL2025 poster today!
My colleagues @anmourchen and Long will present. Come say hi to chat about the paper and all things LLM agents.
โฐTue, July 29th, 16:00-17:30
๐ Hall 4/5, Session 10
#ACL2025NLP
Memory is crucial for building personalized agents. I'll be at ACL 2025 next week to present our paper in reflective memory management, see you in Vienna!
Paper link: arxiv.org/pdf/2503.08026
#LLM #memory #agent #ACL2025
Jun Yan retweeted
@ReviewAcl @emnlpmeeting Urgent help needed.
acFZ: initial score 3
๐ง Complete silence during discussion.
โฐ 4am PST, 9 min before deadline: quietly drops to 2.
with โThanks for the rebuttal. I have updated the score.โ
โ ๏ธ No explanation. No notice. No chance to respond.
(0/n)
Jun Yan retweeted
Is a single accuracy number all we can get from model evals?๐ค
๐จDoes NOT tell where the model fails
๐จDoes NOT tell how to improve it
Introducing EvalTree๐ณ
๐identifying LM weaknesses in natural language
๐weaknesses serve as actionable guidance
(paper&demo ๐in๐งต)
[1/n]
Jun Yan retweeted
I'll be at #NeurIPS2024! My group has papers analyzing how LLMs use Fourier Features for arithmetic (@tianyi_zhou12 ) and how TFs learn higher-order optimization for ICL (@DeqingFu), plus workshop papers on backdoor detection (@jun_yannn) and LLMs + PDDL (@BillJohn1235813)