Building agents. @GoogleDeepMind @penn_state | Prev. @AdobeResearch
State College, PA
Joined June 2022
- Tweets34
- Following673
- Followers101
- Likes897
Check out GUIDE lead by @liuziyi93! We explored whether LLMs can be general facilitator for human group chats!
Introducing ๐๐จ๐๐๐, my summer work at @GoogleDeepMind
๐๐ฎ๐ป ๐๐๐ ๐ ๐น๐ฒ๐ฎ๐ฟ๐ป ๐๐ผ ๐ณ๐ฎ๐ฐ๐ถ๐น๐ถ๐๐ฎ๐๐ฒ ๐ต๐๐บ๐ฎ๐ป ๐ด๐ฟ๐ผ๐๐ฝ ๐ฐ๐ผ๐ป๐๐ฒ๐ฟ๐๐ฎ๐๐ถ๐ผ๐ป๐ ๐ฎ๐ฐ๐ฟ๐ผ๐๐ ๐ฑ๐ถ๐๐ฒ๐ฟ๐๐ฒ ๐ฑ๐ผ๐บ๐ฎ๐ถ๐ป๐โ๐๐ถ๐๐ต๐ผ๐๐ ๐บ๐ฎ๐ป๐๐ฎ๐น๐น๐ ๐ฑ๐ฒ๐๐ถ๐ด๐ป๐ฒ๐ฑ ๐๐๐ฟ๐ฎ๐๐ฒ๐ด๐ถ๐ฒ๐? ๐ค
๐ก ๐ง๐ต๐ฒ ๐๐ต๐ฎ๐น๐น๐ฒ๐ป๐ด๐ฒ
Good facilitation goes beyond task success. It should also improve ๐ต๐๐บ๐ฎ๐ป ๐ฒ๐
๐ฝ๐ฒ๐ฟ๐ถ๐ฒ๐ป๐ฐ๐ฒ and promote ๐ณ๐ฎ๐ถ๐ฟ๐ป๐ฒ๐๐, ensuring that everyone feels heard, comfortable, and respected.
We study three settings: open-ended discussions, consensus-building, and negotiation.
๐ง ๐ข๐๐ฟ ๐๐ฝ๐ฝ๐ฟ๐ผ๐ฎ๐ฐ๐ต
Rather than manually engineering task-specific strategies, GUIDE learns a ๐ด๐ฒ๐ป๐ฒ๐ฟ๐ฎ๐น๐ถ๐๐ฎ๐ฏ๐น๐ฒ ๐ณ๐ฎ๐ฐ๐ถ๐น๐ถ๐๐ฎ๐๐ถ๐ผ๐ป ๐ฝ๐น๐ฎ๐๐ฏ๐ผ๐ผ๐ธ from 432 real human conversations, iteratively refining it through reflection and human-aligned LLM evaluation.
๐ ๐๐ฒ๐ ๐๐ถ๐ป๐ฑ๐ถ๐ป๐ด๐
๐ ๐ฒ๐ญ.๐ฐ% offline win rate against task-specific expert facilitators.
๐ค Across ๐ฐ๐ฒ๐ฎ live conversations, GUIDE improves speaking comfort, reduces interpersonal friction, and promotes more equitable experiences, while maintaining comparable task outcomes.
๐ฌ ๐๐ฒ๐๐ ๐ถ๐ ๐บ๐ผ๐ฟ๐ฒ: ๐ฎ๐ฐ.๐ณ% fewer interventions and ๐ฎ๐ต.๐ฌ% fewer words!
Huge thanks to my amazing collaborators and mentors @tzushengkuo @Jiale_Leo @DeqingFu @Nithum and Crystal Qian!
๐ Paper: arxiv.org/abs/2610.16231
#GoogleDeepMind #LLM #HumanAIInteraction #SocialIntelligence #AIResearch
Jiale Liu (on the market!) retweeted
Super excited to share that Who&When Pro has been accepted to NeurIPS 26 E&D Track!!๐ฅณ
If you are interested in #RSI, agent failure attribution, definitely check this work out! We have 10k+ failure trajectory spanning 3 modalities and different application scenarios!
๐ฅ Recursive self improvement is all the heat right now. But before an AI agent can learn from its own failures, there is a more basic question:
Can LLMs really understand their own failures?
Introducing Who&When Pro, a large scale benchmark for failure attribution in AI agents. It contains over 12k failed trajectories across text, image, and video, covering 26 benchmarks and 15 agent frameworks. The benchmark spans both single agent and multi agent systems, from reasoning and coding agents to web agents, GUI agents, and multimodal agents.
And the answer is not as reassuring as we might hope.
Our results show that current LLMs can sometimes locate where a failure happened, while identifying the underlying failure type remains much more difficult. This leaves a major gap for agents that are expected to diagnose, learn from, and improve on their own mistakes.
Super excited to share that Who&When Pro has been accepted to NeurIPS 26 E&D Track!!๐ฅณ
If you are interested in #RSI, agent failure attribution, definitely check this work out! We have 10k+ failure trajectory spanning 3 modalities and different application scenarios!
๐ฅ Recursive self improvement is all the heat right now. But before an AI agent can learn from its own failures, there is a more basic question:
Can LLMs really understand their own failures?
Introducing Who&When Pro, a large scale benchmark for failure attribution in AI agents. It contains over 12k failed trajectories across text, image, and video, covering 26 benchmarks and 15 agent frameworks. The benchmark spans both single agent and multi agent systems, from reasoning and coding agents to web agents, GUI agents, and multimodal agents.
And the answer is not as reassuring as we might hope.
Our results show that current LLMs can sometimes locate where a failure happened, while identifying the underlying failure type remains much more difficult. This leaves a major gap for agents that are expected to diagnose, learn from, and improve on their own mistakes.
Jiale Liu (on the market!) retweeted
๐ฅ Recursive self improvement is all the heat right now. But before an AI agent can learn from its own failures, there is a more basic question:
Can LLMs really understand their own failures?
Introducing Who&When Pro, a large scale benchmark for failure attribution in AI agents. It contains over 12k failed trajectories across text, image, and video, covering 26 benchmarks and 15 agent frameworks. The benchmark spans both single agent and multi agent systems, from reasoning and coding agents to web agents, GUI agents, and multimodal agents.
And the answer is not as reassuring as we might hope.
Our results show that current LLMs can sometimes locate where a failure happened, while identifying the underlying failure type remains much more difficult. This leaves a major gap for agents that are expected to diagnose, learn from, and improve on their own mistakes.
๐ฅ Recursive self improvement is all the heat right now. But before an AI agent can learn from its own failures, there is a more basic question:
Can LLMs really understand their own failures?
Introducing Who&When Pro, a large scale benchmark for failure attribution in AI agents. It contains over 12k failed trajectories across text, image, and video, covering 26 benchmarks and 15 agent frameworks. The benchmark spans both single agent and multi agent systems, from reasoning and coding agents to web agents, GUI agents, and multimodal agents.
And the answer is not as reassuring as we might hope.
Our results show that current LLMs can sometimes locate where a failure happened, while identifying the underlying failure type remains much more difficult. This leaves a major gap for agents that are expected to diagnose, learn from, and improve on their own mistakes.
๐ What did we find?
Failure localization is already challenging, and failure diagnosis is much harder.
On text trajectories, the best model reaches 73.9% accuracy for locating the decisive step, while the best failure type result is only 22.2% macro F1.
Models also frequently assign the wrong cause to a failure. Planning, verification, and coordination errors are often absorbed into broader reasoning error categories, which shows that current models still struggle to recover the actual structure of an agent failure.
Trajectory length makes the problem harder as well. Step localization drops from 94% on short traces below 3K tokens to 50% on traces above 12K tokens.
These results suggest that reliable failure understanding remains an important missing capability for self improving agents.
We hope this benchmark helps push forward research on agent debugging, failure diagnosis, automated feedback, and self evolving agent systems.
๐ Project website whowhenpro.github.io
๐ Paper arxiv.org/abs/2607.09996
๐ท Code github.com/whowhenpro/whowheโฆ
Kudos to my awesome collaborators for making this project happen๐ฅณ:
@ShaokunZhang1, Huajun Xi, @yokeyxz @VyvyenYue @Chi_Wang_ @jiank_uiuc @qingyun_wu @huazheng_wang
Jiale Liu (on the market!) retweeted
๐ #ICML2026 paper from my @AdobeResearch summer '25 internship!
TL;DR: SteER proactively pauses mid-process to invite user steering + adapts a live persona โ +7.83โ22.80% improved alignment.
๐ arxiv.org/abs/2605.24266
#ICML2026 #LLM #DeepResearch #LLMAgents
Excited to share that TeamFusion, my 25 summer intern work at @AdobeResearch, is featured on Adobe Research blog, and accepted to #ACL2026 main conference! ๐๐๐
TL;DR: TeamFusion helps teams make open-ended decisions by simulating teammate discussions and remixing ideas into stronger team-wide outcomes, all with minimal communication overhead.๐คโจ
#Agents #MultiAgentSystems #HumanAI #LLMAgents
We put real professional designers in the loop on 50 actual ad design scenarios. ๐ฉโ๐จ๐จโ๐จ The results:
โ
They agreed with their agent's comments (< 3% strong disagreement)
โ
They preferred TeamFusion's remixed designs over the originals
โ
Team-wide consensus rose by about 17% on average
(It also beat direct-summary baselines across a 500-team public-opinion benchmark.)
My favorite takeaway: AI should not replace teamwork. It should help teams understand each other faster, preserve diverse voices, and make the path to consensus clearer. ๐ฑ
๐ Paper: arxiv.org/abs/2604.19589
๐ฐ Adobe Research blog (check out the demo here): research.adobe.com/news/teamโฆ
Huge thanks to my amazing collaborator Lin Ai, my mentors @victorsb, Haoliang Wang, Sunav Choudhary, Saayan Mitra, and my advisor @qingyun_wu for making this work happen! ๐
RT @ShaokunZhang1: โShow me the codeโ โ โShow me the trace.โ
Awsome agent failures: github.com/ag2ai/Agents_Failโฆ
As agentic AI systems groโฆ
Jiale Liu (on the market!) retweeted
Nemotron-Research-Tool-N1 is finally accepted at ICLR 2026๐ฅ! Super excited about this milestone after it was first announced in May 2025. In 2026, the training recipe for tool-calling models is still extremely valuable for the entire community.
The key contribution of this work is not merely another RL result, but a practical recipe for turning existing SFT tool-calling data into verifiable training signals.
Instead of relying on expensive distilled trajectories, we show that standard SFT corpora can be systematically transformed and then trained with a lightweight verification signal that only checks:
(1) format correctness
(2) executable tool calls
Producing more RL data is one aspect, but how to better leverage existing data is equally important.
Tool-using LLMs can learn to reasonโwithout reasoning traces.
๐ฅ We present Nemotron-Research-Tool-N1, a family of tool-using reasoning LLMs trained entirely via rule-based reinforcement learningโno reasoning supervision, no distillation.
๐ Paper: arxiv.org/pdf/2505.00024
๐ป Code: github.com/NVlabs/Tool-N1
(Please consider giving us a โญ๏ธ to stay updated on the upcoming code release!)
๐ง Why this matters:
Existing tool-call models rely heavily on supervised reasoning traces from stronger modelsโcostly, brittle, and often imitative. We ask:
Can LLMs learn to reason directly from tool success signals?
๐ฆ What we did:
โ Train Qwen2.5-7B/14B with simple binary reward on tool-call correctness + reasoning format in R1-style
โ No reasoning traces needed
โ Evaluate on BFCL, API-Bank, and ACEBench
โ Also study the role of SFT, RL, and widely adopted SFT-then-RL recipes in training Tool-Calling models.
๐ Key findings:
โ Tool-N1-7B/14B obviously outperform GPT-4o and open baselines on all benchmarks
โ Widely adopted SFT+RL paradigm doesnโt necessarily lead to better performance than Pure RL.
โ Binary reward > fine-grained reward, esp. for real-world queries
โ Scaling works: bigger = better gains under our RL setup
๐ Takeaway:
Reasoning doesnโt have to be taught. With just a binary signal, LLMs can learn to reason and act.
Tool-N1 sets a new direction for scalable, supervision-light tool calling model training
Jiale Liu (on the market!) retweeted
๐ค๐The value gap between YOLO agents and a learning AI workforce:
A lot of people donโt realize how much upside theyโre leaving on the table in terms of their agent system. Simply because they are still in the demo stage, or their agents arenโt even running long enough to learn.
As more agents are deployed into real production environments, especially as an AI workforce, reliability, performance, and the ability to continuously learn and improve start to actually matter.
These arenโt one-off executions. Theyโre long-running systems, expected to operate over time, learn from outcomes, and compound gains.
Hereโs a quick sneak peek of what learning from real data can unlock ๐
1. On Prophet Arena (a leaderboard with live market data):
A learning agent built with AG2 outperforms an OpenAI agent using the same model (GPT-5.2) by ~18% in market return
2. A learning AG2 agent using an earlier-generation model ( (GPT-4.1) outperforms a non-learning agent using GPT-5.2 by ~13% in market return
๐ That's the power of a multi-agent styem learn from outcomes, adapts its strategy, and improves over time.
Jiale Liu (on the market!) retweeted
If you're at #ICML2025, come say hi at our Spotlight Poster session! ๐
๐ง "Which Agent Causes Task Failures and When?"
We introduce the first benchmark for automated failure attribution in LLM multi-agent systems โ identifying who failed and when.
๐ East Exhibition Hall A-B #E-2701
๐ Thu July 17 (today), 11amโ1:30pm PDT
๐ lnkd.in/eAMKE72N
#AI #LLM #MultiAgent #AIagents
Jiale Liu (on the market!) retweeted
Tool-using LLMs can learn to reasonโwithout reasoning traces.
๐ฅ We present Nemotron-Research-Tool-N1, a family of tool-using reasoning LLMs trained entirely via rule-based reinforcement learningโno reasoning supervision, no distillation.
๐ Paper: arxiv.org/pdf/2505.00024
๐ป Code: github.com/NVlabs/Tool-N1
(Please consider giving us a โญ๏ธ to stay updated on the upcoming code release!)
๐ง Why this matters:
Existing tool-call models rely heavily on supervised reasoning traces from stronger modelsโcostly, brittle, and often imitative. We ask:
Can LLMs learn to reason directly from tool success signals?
๐ฆ What we did:
โ Train Qwen2.5-7B/14B with simple binary reward on tool-call correctness + reasoning format in R1-style
โ No reasoning traces needed
โ Evaluate on BFCL, API-Bank, and ACEBench
โ Also study the role of SFT, RL, and widely adopted SFT-then-RL recipes in training Tool-Calling models.
๐ Key findings:
โ Tool-N1-7B/14B obviously outperform GPT-4o and open baselines on all benchmarks
โ Widely adopted SFT+RL paradigm doesnโt necessarily lead to better performance than Pure RL.
โ Binary reward > fine-grained reward, esp. for real-world queries
โ Scaling works: bigger = better gains under our RL setup
๐ Takeaway:
Reasoning doesnโt have to be taught. With just a binary signal, LLMs can learn to reason and act.
Tool-N1 sets a new direction for scalable, supervision-light tool calling model training