AI Group @ucsantabarbara. Profs. @xwang_lk, @YuhengBu, Xifeng Yan, @WilliamWangNLP, @CodeTerminator. Account run by Student Social Committee.
Santa Barbara, CA
Joined July 2021
- Tweets382
- Following639
- Followers2.5K
- Likes649
UCSB AI retweeted
๐ ๐๐๐ซ๐ฌ๐จ๐ง๐๐ฅ ๐๐ ๐๐ง๐ญ๐ฌ ๐๐ซ๐ ๐๐๐ค๐ข๐ง๐ ๐๐๐: ๐๐จ๐จ๐ ๐ฅ๐ ร ๐๐๐๐ ๐๐ง๐ญ๐ซ๐จ๐๐ฎ๐๐ ๐๐ญ๐ซ๐๐๐ฆ๐๐ข๐๐ ๐ญ๐จ ๐๐ง๐๐๐ซ๐ฌ๐ญ๐๐ง๐ ๐๐ฏ๐จ๐ฅ๐ฏ๐ข๐ง๐ ๐๐ฌ๐๐ซ ๐๐ง๐ญ๐๐ง๐ญ ๐๐ก๐ซ๐จ๐ฎ๐ ๐ก ๐๐ง๐ญ๐๐ซ๐๐๐ญ๐ข๐จ๐ง
๐คThe promise of personal agents is compelling: an assistant that understands what we need as we go about our day. But understanding a user is a moving target.
๐ณย When we say "help me cook this dish," the next useful action depends on what we do next. A step gets completed. A mistake happens. We change our mind. The agent has to keep up with both the world and our intent, deciding when to help and when to stay quiet.
๐ฅย Introducing StreamVibe: Proactive Video Agents for Evolving User Intent.
๐งฉ Three key pieces power StreamVibe: โ StreamVibe-VL learns when and how to help; โก StreamVibe-Harness tracks evolving user intents and task progress; โข StreamVibe-Bench tests proactive assistance in real-world, long-horizon interactions.
๐ฌย Watch the demo. More on the ideas and results below. ๐งต
๐ย stream-vibe.github.io/
UCSB AI retweeted
๐ Huge congratulations to @WengZhaoti39773 and our UCSB students on winning the ๐ข๐๐๐๐๐ฎ๐ป๐ฑ๐ถ๐ป๐ด ๐ฃ๐ฎ๐ฝ๐ฒ๐ฟ ๐๐๐ฎ๐ฟ๐ฑ at the COLM 2026 Lifelong Agent Workshop for our ๐๐ฟ๐ผ๐๐ฝ-๐๐๐ผ๐น๐๐ถ๐ป๐ด ๐๐ด๐ฒ๐ป๐๐ paper!
This is Zhaotianโs ๐๐ฒ๐ฟ๐ ๐ณ๐ถ๐ฟ๐๐ ๐ฝ๐ฟ๐ผ๐ท๐ฒ๐ฐ๐ in our group; what an incredible achievement!
๐๐ผ๐น๐น๐ฒ๐ฐ๐๐ถ๐๐ฒ ๐ฒ๐๐ผ๐น๐๐๐ถ๐ผ๐ป ๐ถ๐ ๐ท๐๐๐ ๐๐ต๐ฒ ๐ฏ๐ฒ๐ด๐ถ๐ป๐ป๐ถ๐ป๐ด!
๐ Honored that Group-Evolving Agents won the Outstanding Paper Award at the Lifelong Agents Workshop @ #COLM2026!
Agents that evolve as a group, sharing experience, keep improving open-endedly ๐งฌ๐ค
Grateful to my wonderful coauthors and @xwang_lk !
Had a great time at #COLM2026
๐ arxiv.org/abs/2602.04837
๐ป github.com/UCSB-AI/GEA
UCSB AI retweeted
We got a Spotlight Position Paper at NeurIPS 2026! What a week full of events. And I got to author this paper with some of the best minds in AI.
The paper will be available at the end of this month! Stay tuned
๐๐๐ฎ๐ซ๐๐๐ 2026 ๐๐ฉ๐จ๐ญ๐ฅ๐ข๐ ๐ก๐ญ! Huge congrats to @AlfonAmayuelas on our position paper "Sampling Is Not Curiosity: Why LLM Agents Should Investigate" being selected as ๐จ๐ง๐ ๐จ๐ ๐ญ๐ก๐ ๐จ๐ง๐ฅ๐ฒ 11 ๐ฉ๐๐ฉ๐๐ซ๐ฌ ๐๐ฌ ๐๐ฉ๐จ๐ญ๐ฅ๐ข๐ ๐ก๐ญ!
Paper coming soon!
UCSB AI retweeted
๐ค I will be giving an Oral Presentation today! @ Lifelong Agent Workshop ยท Fri Oct 9, 11:45 AMโ12:00 PM ยท Continental Ballroom 6
โจ Group-Evolving Agents: Open-Ended Self-Improvement via Experience Sharing
GEA treats a group of agents, not an individual, as the unit of evolution. Agents share experience as they self-improve, outperforming prior state-of-the-art open-ended self-improving methods and matching or exceeding top human-designed agent frameworks, all with zero human intervention.
AI agents aren't biological individuals. So why make them evolve like one?
They can directly share experience and learned artifacts. Not bounded by reproduction, lineage, or genes.
Meet ๐๐๐, accepted to ๐๐ข๐๐ ๐ฎ๐ฌ๐ฎ๐ฒ ๐
๐ณ๐ญ.๐ฌ% SWE-bench Verified / ๐ด๐ด.๐ฏ% Polyglot ยท zero human intervention
๐งต๐
Most self-evolving agent systems follow a similar pattern: select a parent, refine it, produce an offspring, repeat. Evolution unfolds as a tree.
It's great at generating diversity, but that diversity gets trapped. Agents explore independently, and instead of serving as stepping stones, their discoveries stay stuck in local branches. Most variants are short-lived.
๐๐
๐ฝ๐น๐ผ๐ฟ๐ฎ๐๐ถ๐ผ๐ป ๐ต๐ฎ๐ฝ๐ฝ๐ฒ๐ป๐; ๐ฟ๐ฒ๐๐๐ฒ ๐ฎ๐ป๐ฑ ๐ฎ๐ฐ๐ฐ๐๐บ๐๐น๐ฎ๐๐ถ๐ผ๐ป ๐ฟ๐ฎ๐ฟ๐ฒ๐น๐ ๐ฑ๐ผ.
It's time to rethink the evolution of AI agents: why keep evolving them like biological individuals?
AI agents aren't bound by reproduction, lineage, or genes. They can directly share trajectories, tools, workflows, and learned artifacts, aggregating complementary skills instantly. ๐ช๐ต๐ ๐ป๐ผ๐ ๐ฟ๐ฒ๐ฑ๐ฒ๐๐ถ๐ด๐ป ๐ฒ๐๐ผ๐น๐๐๐ถ๐ผ๐ป ๐ฎ๐ฟ๐ผ๐๐ป๐ฑ ๐๐ต๐ฎ๐ ๐๐ต๐ฒ๐ ๐ฐ๐ฎ๐ป ๐ฎ๐ฐ๐๐๐ฎ๐น๐น๐ ๐ฑ๐ผ?
That's GEA.
GEA makes a ๐ด๐ฟ๐ผ๐๐ฝ ๐ผ๐ณ ๐ฎ๐ด๐ฒ๐ป๐๐ ๐๐ต๐ฒ ๐ณ๐๐ป๐ฑ๐ฎ๐บ๐ฒ๐ป๐๐ฎ๐น ๐๐ป๐ถ๐ ๐ผ๐ณ ๐ฒ๐๐ผ๐น๐๐๐ถ๐ผ๐ป. Each round, a parent group is selected under a Performance-Novelty criterion that balances competence with exploratory diversity. All members within the group then pool their experience, model patches, failure modes, eval logs, and solutions, into a shared pool, and the whole group jointly produces the next generation.
Exploration is no longer wasted. It gets consolidated.
The results show a significant improvement over prior state-of-the-art self-evolving methods, and GEA matches or surpasses top human-designed frameworks with ๐ป๐ผ ๐ต๐๐บ๐ฎ๐ป ๐ถ๐ป ๐๐ต๐ฒ ๐น๐ผ๐ผ๐ฝ.
The analysis is the more interesting part. GEA's gains come from explicitly reusing diversity, not from lucky outliers: stronger performance under the same number of evolved agents, more robust to framework-level bugs (repaired in ๐ญ.๐ฐ iterations vs ๐ฑ), and improvements that target workflows and tools rather than overfitting to one model, so they transfer consistently across GPT- and Claude-series backbones.
The key to open-ended evolution is not only generating enough diversity. What matters more is whether discoveries accumulate and get reused. ๐๐๐ ๐น๐ฒ๐๐ ๐ฒ๐
๐ฝ๐น๐ผ๐ฟ๐ฎ๐๐ถ๐ผ๐ป ๐๐๐ฎ๐ฟ๐ ๐ฐ๐ผ๐บ๐ฝ๐ผ๐๐ป๐ฑ๐ถ๐ป๐ด.
๐ arXiv: arxiv.org/abs/2602.04837
Grateful to my advisor @xwang_lk and my wonderful coauthors @anton_iades @deepaknathani11 @zhenzhangzz @XiaoSophiaPu .
Interning in Palo Alto this summer. See you at COLM 2026 ๐
UCSB AI retweeted
Presenting GEA at #COLM2026 this week! ๐งฌ
โจ Group-Evolving Agents: Open-Ended Self-Improvement via Experience Sharing
GEA treats a group of agents, not an individual, as the unit of evolution. Agents share experience as they self-improve, outperforming prior state-of-the-art open-ended self-improving methods and matching or exceeding top human-designed agent frameworks, all with zero human intervention.
๐ Poster ยท Wed Oct 7, 4:30 PM ยท Imperial Ballroom ยท #21
๐ค Oral @ Lifelong Agent Workshop ยท Fri Oct 9, 11:45 AMโ12:00 PM ยท Continental Ballroom 6
My research focuses on recursive self-improvement and autoresearch. Stop by the poster to chat, or DM me to grab a coffee! โ
๐ arxiv.org/abs/2602.04837
AI agents aren't biological individuals. So why make them evolve like one?
They can directly share experience and learned artifacts. Not bounded by reproduction, lineage, or genes.
Meet ๐๐๐, accepted to ๐๐ข๐๐ ๐ฎ๐ฌ๐ฎ๐ฒ ๐
๐ณ๐ญ.๐ฌ% SWE-bench Verified / ๐ด๐ด.๐ฏ% Polyglot ยท zero human intervention
๐งต๐
Most self-evolving agent systems follow a similar pattern: select a parent, refine it, produce an offspring, repeat. Evolution unfolds as a tree.
It's great at generating diversity, but that diversity gets trapped. Agents explore independently, and instead of serving as stepping stones, their discoveries stay stuck in local branches. Most variants are short-lived.
๐๐
๐ฝ๐น๐ผ๐ฟ๐ฎ๐๐ถ๐ผ๐ป ๐ต๐ฎ๐ฝ๐ฝ๐ฒ๐ป๐; ๐ฟ๐ฒ๐๐๐ฒ ๐ฎ๐ป๐ฑ ๐ฎ๐ฐ๐ฐ๐๐บ๐๐น๐ฎ๐๐ถ๐ผ๐ป ๐ฟ๐ฎ๐ฟ๐ฒ๐น๐ ๐ฑ๐ผ.
It's time to rethink the evolution of AI agents: why keep evolving them like biological individuals?
AI agents aren't bound by reproduction, lineage, or genes. They can directly share trajectories, tools, workflows, and learned artifacts, aggregating complementary skills instantly. ๐ช๐ต๐ ๐ป๐ผ๐ ๐ฟ๐ฒ๐ฑ๐ฒ๐๐ถ๐ด๐ป ๐ฒ๐๐ผ๐น๐๐๐ถ๐ผ๐ป ๐ฎ๐ฟ๐ผ๐๐ป๐ฑ ๐๐ต๐ฎ๐ ๐๐ต๐ฒ๐ ๐ฐ๐ฎ๐ป ๐ฎ๐ฐ๐๐๐ฎ๐น๐น๐ ๐ฑ๐ผ?
That's GEA.
GEA makes a ๐ด๐ฟ๐ผ๐๐ฝ ๐ผ๐ณ ๐ฎ๐ด๐ฒ๐ป๐๐ ๐๐ต๐ฒ ๐ณ๐๐ป๐ฑ๐ฎ๐บ๐ฒ๐ป๐๐ฎ๐น ๐๐ป๐ถ๐ ๐ผ๐ณ ๐ฒ๐๐ผ๐น๐๐๐ถ๐ผ๐ป. Each round, a parent group is selected under a Performance-Novelty criterion that balances competence with exploratory diversity. All members within the group then pool their experience, model patches, failure modes, eval logs, and solutions, into a shared pool, and the whole group jointly produces the next generation.
Exploration is no longer wasted. It gets consolidated.
The results show a significant improvement over prior state-of-the-art self-evolving methods, and GEA matches or surpasses top human-designed frameworks with ๐ป๐ผ ๐ต๐๐บ๐ฎ๐ป ๐ถ๐ป ๐๐ต๐ฒ ๐น๐ผ๐ผ๐ฝ.
The analysis is the more interesting part. GEA's gains come from explicitly reusing diversity, not from lucky outliers: stronger performance under the same number of evolved agents, more robust to framework-level bugs (repaired in ๐ญ.๐ฐ iterations vs ๐ฑ), and improvements that target workflows and tools rather than overfitting to one model, so they transfer consistently across GPT- and Claude-series backbones.
The key to open-ended evolution is not only generating enough diversity. What matters more is whether discoveries accumulate and get reused. ๐๐๐ ๐น๐ฒ๐๐ ๐ฒ๐
๐ฝ๐น๐ผ๐ฟ๐ฎ๐๐ถ๐ผ๐ป ๐๐๐ฎ๐ฟ๐ ๐ฐ๐ผ๐บ๐ฝ๐ผ๐๐ป๐ฑ๐ถ๐ป๐ด.
๐ arXiv: arxiv.org/abs/2602.04837
Grateful to my advisor @xwang_lk and my wonderful coauthors @anton_iades @deepaknathani11 @zhenzhangzz @XiaoSophiaPu .
Interning in Palo Alto this summer. See you at COLM 2026 ๐
UCSB AI retweeted
1/6 ๐ง ๐ฌ๐ผ๐๐ฟ ๐๐ ๐๐๐๐ถ๐๐๐ฎ๐ป๐ ๐๐ฎ๐ป ๐ฅ๐ฒ๐บ๐ฒ๐บ๐ฏ๐ฒ๐ฟ ๐ช๐ต๐ฎ๐ ๐ฌ๐ผ๐ ๐ฆ๐ฎ๐ถ๐ฑโ๐ฎ๐ป๐ฑ ๐ฆ๐๐ถ๐น๐น ๐๐ด๐ป๐ผ๐ฟ๐ฒ ๐ช๐ต๐ฎ๐ ๐ฌ๐ผ๐ ๐ช๐ฎ๐ป๐.
Long-running assistants accumulate facts, preferences, and constraints across conversations. But remembering them is not enough: a response can sound fluent and relevant while still recommending exactly what the user has ruled out.
We present ๐ ๐ฒ๐บ๐๐ผ๐น๐ฑ: ๐๐ฒ๐ฎ๐ฟ๐ป๐ถ๐ป๐ด ๐๐ผ๐บ๐ฝ๐ฎ๐ฐ๐ ๐ฆ๐ผ๐ณ๐ ๐ ๐ฒ๐บ๐ผ๐ฟ๐ ๐ณ๐ผ๐ฟ ๐๐ผ๐ป๐ด-๐๐ผ๐ป๐๐ฒ๐
๐ ๐ฃ๐ฒ๐ฟ๐๐ผ๐ป๐ฎ๐น๐ถ๐๐ฎ๐๐ถ๐ผ๐ป ๐๐ถ๐ฎ ๐ข๐ป-๐ฃ๐ผ๐น๐ถ๐ฐ๐ ๐ข๐ฝ๐๐ถ๐บ๐ถ๐๐ฎ๐๐ถ๐ผ๐ป. MemFold trains a model to use compact soft memory by learning from the answers it produces, combining task-level rewards with more detailed token-level guidance.
๐ข๐๐ฟ ๐ธ๐ฒ๐ ๐ณ๐ถ๐ป๐ฑ๐ถ๐ป๐ด๐ ๐ฅ:
๐ฏ ๐๐๐ฑ๐ด๐ฒ ๐บ๐ฒ๐บ๐ผ๐ฟ๐ ๐ฏ๐ ๐๐ต๐ฒ ๐ฎ๐ป๐๐๐ฒ๐ฟ๐ ๐ถ๐ ๐๐๐ฝ๐ฝ๐ผ๐ฟ๐๐. Reconstruction alone does not tell us whether a model will respect a userโs preferences.
๐ ๐ข๐ป๐ฒ ๐๐ฒ๐ ๐ผ๐ณ ๐ฟ๐ฒ๐๐ฝ๐ผ๐ป๐๐ฒ๐, ๐๐๐ผ ๐น๐ฒ๐ฎ๐ฟ๐ป๐ถ๐ป๐ด ๐๐ถ๐ด๐ป๐ฎ๐น๐. GRPO rewards task outcomes, while token-level guidance provides more detailed feedback on the same generated responses.
๐ ๐๐ฎ๐ฟ๐ด๐ฒ๐ฟ ๐ด๐ฎ๐ถ๐ป๐ ๐๐ถ๐๐ต ๐น๐ผ๐ป๐ด๐ฒ๐ฟ ๐ต๐ถ๐๐๐ผ๐ฟ๐ถ๐ฒ๐. Across three Qwen backbones, the margin over the strongest baseline widens from PersonaMem-32K to PersonaMem-128K.
๐ ๐ง๐ฟ๐ฎ๐ป๐๐ณ๐ฒ๐ฟ ๐ฏ๐ฒ๐๐ผ๐ป๐ฑ ๐๐ต๐ฒ ๐๐ฟ๐ฎ๐ถ๐ป๐ถ๐ป๐ด ๐ฑ๐ฎ๐๐ฎ. MemFold transfers to PrefEval and LongMemEval without target-domain training.
๐ ๐๐ต๐ฒ๐ฐ๐ธ ๐ผ๐๐ ๐๐ต๐ฒ ๐ฝ๐ฎ๐ฝ๐ฒ๐ฟ, ๐ฝ๐ฟ๐ผ๐ท๐ฒ๐ฐ๐, ๐ฎ๐ป๐ฑ ๐ฐ๐ผ๐ฑ๐ฒ:
๐ Paper: huggingface.co/papers/2609.3โฆ
๐ Project: memfold.github.io
๐ป Code: github.com/Johnny221B/memfolโฆ
UCSB AI retweeted
AI agents aren't biological individuals. So why make them evolve like one?
They can directly share experience and learned artifacts. Not bounded by reproduction, lineage, or genes.
Meet ๐๐๐, accepted to ๐๐ข๐๐ ๐ฎ๐ฌ๐ฎ๐ฒ ๐
๐ณ๐ญ.๐ฌ% SWE-bench Verified / ๐ด๐ด.๐ฏ% Polyglot ยท zero human intervention
๐งต๐
Most self-evolving agent systems follow a similar pattern: select a parent, refine it, produce an offspring, repeat. Evolution unfolds as a tree.
It's great at generating diversity, but that diversity gets trapped. Agents explore independently, and instead of serving as stepping stones, their discoveries stay stuck in local branches. Most variants are short-lived.
๐๐
๐ฝ๐น๐ผ๐ฟ๐ฎ๐๐ถ๐ผ๐ป ๐ต๐ฎ๐ฝ๐ฝ๐ฒ๐ป๐; ๐ฟ๐ฒ๐๐๐ฒ ๐ฎ๐ป๐ฑ ๐ฎ๐ฐ๐ฐ๐๐บ๐๐น๐ฎ๐๐ถ๐ผ๐ป ๐ฟ๐ฎ๐ฟ๐ฒ๐น๐ ๐ฑ๐ผ.
It's time to rethink the evolution of AI agents: why keep evolving them like biological individuals?
AI agents aren't bound by reproduction, lineage, or genes. They can directly share trajectories, tools, workflows, and learned artifacts, aggregating complementary skills instantly. ๐ช๐ต๐ ๐ป๐ผ๐ ๐ฟ๐ฒ๐ฑ๐ฒ๐๐ถ๐ด๐ป ๐ฒ๐๐ผ๐น๐๐๐ถ๐ผ๐ป ๐ฎ๐ฟ๐ผ๐๐ป๐ฑ ๐๐ต๐ฎ๐ ๐๐ต๐ฒ๐ ๐ฐ๐ฎ๐ป ๐ฎ๐ฐ๐๐๐ฎ๐น๐น๐ ๐ฑ๐ผ?
That's GEA.
GEA makes a ๐ด๐ฟ๐ผ๐๐ฝ ๐ผ๐ณ ๐ฎ๐ด๐ฒ๐ป๐๐ ๐๐ต๐ฒ ๐ณ๐๐ป๐ฑ๐ฎ๐บ๐ฒ๐ป๐๐ฎ๐น ๐๐ป๐ถ๐ ๐ผ๐ณ ๐ฒ๐๐ผ๐น๐๐๐ถ๐ผ๐ป. Each round, a parent group is selected under a Performance-Novelty criterion that balances competence with exploratory diversity. All members within the group then pool their experience, model patches, failure modes, eval logs, and solutions, into a shared pool, and the whole group jointly produces the next generation.
Exploration is no longer wasted. It gets consolidated.
The results show a significant improvement over prior state-of-the-art self-evolving methods, and GEA matches or surpasses top human-designed frameworks with ๐ป๐ผ ๐ต๐๐บ๐ฎ๐ป ๐ถ๐ป ๐๐ต๐ฒ ๐น๐ผ๐ผ๐ฝ.
The analysis is the more interesting part. GEA's gains come from explicitly reusing diversity, not from lucky outliers: stronger performance under the same number of evolved agents, more robust to framework-level bugs (repaired in ๐ญ.๐ฐ iterations vs ๐ฑ), and improvements that target workflows and tools rather than overfitting to one model, so they transfer consistently across GPT- and Claude-series backbones.
The key to open-ended evolution is not only generating enough diversity. What matters more is whether discoveries accumulate and get reused. ๐๐๐ ๐น๐ฒ๐๐ ๐ฒ๐
๐ฝ๐น๐ผ๐ฟ๐ฎ๐๐ถ๐ผ๐ป ๐๐๐ฎ๐ฟ๐ ๐ฐ๐ผ๐บ๐ฝ๐ผ๐๐ป๐ฑ๐ถ๐ป๐ด.
๐ arXiv: arxiv.org/abs/2602.04837
Grateful to my advisor @xwang_lk and my wonderful coauthors @anton_iades @deepaknathani11 @zhenzhangzz @XiaoSophiaPu .
Interning in Palo Alto this summer. See you at COLM 2026 ๐
UCSB AI retweeted
Key to realizing Auto Research Agents that can make novel discoveries in AI is understanding and improving their exploration capabilities.
To this end, we built Heuresis, a composable framework that combines coding agents with arbitrary search algorithms within a flexible loop.
UCSB AI retweeted
I will be attending @icmlconf in Seoul ๐ฐ๐ท next week presenting SAW-Bench!!
Would love to chat about embodied perception, spatial reasoning, world modeling, and general video understanding ๐
DMs are open ๐
โ๏ธTue, Jun 7, 10:30 am -- 12:15 pm
๐Hall A #4309
Human perception is inherently situated โ we understand the world relative to our own body, viewpoint, and motion.
To deploy multimodal foundation models in embodied settings, we ask:
โCan these models reason in the same observer-centric way?โ
We study this through SAW-Bench: a novel benchmark for observer-centric situated awareness:
- 786 real world egocentric videos
- 2,071 human-annotated QA pairs
Across all tasks, we evaluate 24 state-of-the-art MFMs:
๐ Best model: 53.9%
๐ง Humans: 91.6%
Models systematically:
โ Confuse head rotation with physical movement
โ Collapse under multi-turn trajectories
โ Fail to maintain persistent world-state memory
๐ We see that maintaining a stable observer-centric representation remains challenging.
As MFMs are increasingly integrated into embodied agents, situated awareness becomes essential for reliable real-world interaction.
We release SAW-Bench and encourage further research toward improving observer-centric reasoning in multimodal foundation models.
UCSB AI retweeted
Auto-research is a search problem.
๐๐๐๐ซ๐๐ก ๐ข๐ฌ ๐ ๐๐จ๐ซ๐ฆ ๐จ๐ ๐ก๐จ๐ฅ๐ข๐ฌ๐ญ๐ข๐ ๐ซ๐๐๐ฌ๐จ๐ง๐ข๐ง๐ ๐จ๐ฏ๐๐ซ ๐๐ง ๐๐ฏ๐จ๐ฅ๐ฏ๐ข๐ง๐ ๐ฅ๐๐ง๐๐ฌ๐๐๐ฉ๐ ๐จ๐ ๐ฉ๐จ๐ฌ๐ฌ๐ข๐๐ข๐ฅ๐ข๐ญ๐ข๐๐ฌ.
A scientist does not simply solve a task. They generate hypotheses, design experiments, interpret failures, update beliefs, and decide where to explore next. The central challenge is navigating the search space efficiently.
This perspective is largely missing from todayโs autonomous research systems, where reasoning is often confined to a single trajectory.
In ARTS, we introduce a reasoning-guided tree search framework with test-time learning, enabling agents to reason about the search process itself. We show that a fine-tuned 4B model can achieve performance comparable to frontier closed-model research agents on MLGym and MLEBench, while operating at substantially lower inference cost.
More broadly, we believe that progress in autonomous research will increasingly come from better reasoning-driven search, not just larger models.
UCSB AI retweeted
While AI scientists are getting better, two core challenges remain: (i) deciding what to explore next and (ii) handling the growing context.
ARTS addresses both with reasoning-guided exploration and test-time RL to learn from long search histories.
UCSB AI retweeted
A good researcher has intuition about when to drop a line of work and when to push. Current auto-discovery systems delegate this decision to scalar scores, which cannot represent this judgment. ARTS (Agentic Reasoning for Tree Search) reasons over the search state to decide.
UCSB AI retweeted
Excited to be at @icmlconf in Seoul๐ฐ๐ท๐!
Presenting my work on Reasoning, RL, Co-Evolution & automated discovery.
Would love to chat about RL, Agents, Reasoning, Automated Discovery, Co-Evolution & anything else.
Come say hi by my posters July 8 and 9 in Hall A! ๐
UCSB AI retweeted
Interesting observation: even after TTT, diversity doesn't collapse. We test-time-train the ARTS "scientist" on a simple, performance-only reward. Intuition says this should collapse solution diversity. But it doesn't, this is attributed to verbalized sampling which still surfaces the low-probability solution.
UCSB AI retweeted
Automated discovery is a search problem over combinatorially large spaces. It requires reasoning to explore promising ideas while handling long contexts.
Introducing ARTS, a search algorithm that (i) reasons to guide exploration and (ii) test-time trains to handle long context
UCSB AI retweeted
Had so much fun working on this! ABC provides robot data and base models for all!
Introducing ABC: open data, training, and infrastructure for robotics.
We release the largest teleop dataset to date, and extensively investigate design decisions, pretraining, and post-training techniques.
@arthurallshire @Cinnabar233 @adamrasb @redstone_hong @davidrmcall
UCSB AI retweeted
Excited to share that our paper SAW-Bench: Learning Situated Awareness in the Real World received the Best Paper Award Runner-Up at the #CVPR2026 WMAS workshop!
Congratulations to all co-authors, and thanks to the organizers and reviewers for the recognition.
Special thanks to @jieneng_chen for hosting me at WMAS!
Human perception is inherently situated โ we understand the world relative to our own body, viewpoint, and motion.
To deploy multimodal foundation models in embodied settings, we ask:
โCan these models reason in the same observer-centric way?โ
We study this through SAW-Bench: a novel benchmark for observer-centric situated awareness:
- 786 real world egocentric videos
- 2,071 human-annotated QA pairs
Across all tasks, we evaluate 24 state-of-the-art MFMs:
๐ Best model: 53.9%
๐ง Humans: 91.6%
Models systematically:
โ Confuse head rotation with physical movement
โ Collapse under multi-turn trajectories
โ Fail to maintain persistent world-state memory
๐ We see that maintaining a stable observer-centric representation remains challenging.
As MFMs are increasingly integrated into embodied agents, situated awareness becomes essential for reliable real-world interaction.
We release SAW-Bench and encourage further research toward improving observer-centric reasoning in multimodal foundation models.
UCSB AI retweeted
๐คIt is time to rethink how we evaluate agent memory
๐ As agents become longer horizon and more autonomous, memory is no longer just a module for storing past chats.
๐ ๏ธ It determines how agents track changing worlds, learn from past actions, revise outdated information, and reuse experience for future decisions.
๐ This raises three key questions:
Are human designed write store retrieve memory pipelines still the best choice?
If harnesses such as Codex, Claude Code, and OpenClaw already let agents observe, act, call tools, write files, and reorganize context, can memory also be managed by the harness itself?
Do current evaluations really cover how agent memory is used in realistic settings? Many benchmarks are still text centric or single modal, with limited pressure from screenshots, GUIs, tool feedback, and environment changes.
โ Is final QA accuracy enough?
๐ฅ We present WorldMemArena, a multimodal benchmark for evaluating agent memory through action world interaction.
๐ Key insights:
๐งฉ Memory is a lifecycle, not a static cache.
๐ Better memory storage does not necessarily lead to better final performance.
๐ผ๏ธ Multimodal memory remains a major bottleneck for current systems.
๐ Real agentic trajectories expose the fragility of memory systems.
โ๏ธ Harness-based memory is more flexible, but still costly and unstable.
UCSB AI retweeted
๐๐ก๐ ๐๐ข๐ญ๐ญ๐๐ซ ๐๐๐ฌ๐ฌ๐จ๐ง ๐จ๐ ๐๐ ๐๐ง๐ญ๐ข๐ ๐๐๐ฆ๐จ๐ซ๐ฒ: memory should be a derived capability that exists because it makes an agent better at acting over time.
๐๐จ๐ซ๐ฅ๐๐๐๐ฆ๐๐ซ๐๐ง๐ is designed around this principle. Rather than evaluating memory as a storage problem, WorldMemArena evaluates memory through ๐๐๐ญ๐ข๐จ๐งโ๐ฐ๐จ๐ซ๐ฅ๐ ๐ข๐ง๐ญ๐๐ซ๐๐๐ญ๐ข๐จ๐ง, instrumenting the full write โ maintain โ retrieve โ use lifecycle across 400 multimodal, multi-session tasks.
And it exposes the findings that should mark the end of the storage-centric era:
โ Storage โ use. Better memory storage and retrieval do not necessarily produce better task performance. Optimizing the component we designed does not optimize the capability we actually care about.
โ Harness-based memory performs best where memory is hardest. Agents that can write files, reorganize context, create artifacts, and interact with persistent environments adapt most effectively in long-horizon settings. They are costly and unstable today, which is exactly what many Bitter Lesson transitions look like before scaling and learning take over.
The deeper move is in what gets measured. Memory shouldn't get a score; it should be inferred from capability: how much does remembering improve performance over time.
WorldMemArena drags evaluation off the static object and into the actionโworld loop, the only place you can tell whether an agent has developed memory or is just simulating it convincingly.
๐คIt is time to rethink how we evaluate agent memory
๐ As agents become longer horizon and more autonomous, memory is no longer just a module for storing past chats.
๐ ๏ธ It determines how agents track changing worlds, learn from past actions, revise outdated information, and reuse experience for future decisions.
๐ This raises three key questions:
Are human designed write store retrieve memory pipelines still the best choice?
If harnesses such as Codex, Claude Code, and OpenClaw already let agents observe, act, call tools, write files, and reorganize context, can memory also be managed by the harness itself?
Do current evaluations really cover how agent memory is used in realistic settings? Many benchmarks are still text centric or single modal, with limited pressure from screenshots, GUIs, tool feedback, and environment changes.
โ Is final QA accuracy enough?
๐ฅ We present WorldMemArena, a multimodal benchmark for evaluating agent memory through action world interaction.
๐ Key insights:
๐งฉ Memory is a lifecycle, not a static cache.
๐ Better memory storage does not necessarily lead to better final performance.
๐ผ๏ธ Multimodal memory remains a major bottleneck for current systems.
๐ Real agentic trajectories expose the fragility of memory systems.
โ๏ธ Harness-based memory is more flexible, but still costly and unstable.
UCSB AI retweeted
Finally, this got done. It always felt off when people were treating VLAs as a separate class from multimodal LLMs. Better late than never.
Excited to share Qwen-VLA paper, our exploration of generalist Vision-Language-Action models.
It extends Qwenโs multimodal backbone from visual understanding and reasoning to continuous action generation and trajectory prediction.
Paper:
arxiv.org/pdf/2605.30280
