Jianshu Zhang โœˆ๏ธICLR2025๐Ÿ‡ธ๐Ÿ‡ฌ (@sterzhang) 's Twitter Profile
Jianshu Zhang โœˆ๏ธICLR2025๐Ÿ‡ธ๐Ÿ‡ฌ

@sterzhang

Senior Year Undergraduate @WHU_1893 ๐Ÿ–๏ธ | Incoming PhD Student in Computer Science @NorthwesternU ๐Ÿ’œ

ID: 1770670160078397440

calendar_today21-03-2024 04:34:10

35 Tweet

86 Takipรงi

178 Takip Edilen

Zihan Wang - on RAGEN (@wzihanw) 's Twitter Profile Photo

In the last two months, RAGEN has powered Agent RL training frameworks for over 300,000 people. Now, weโ€™re introducing VAGENโ€”the first open-source framework that trains *Visual* Agents using multi-turn Reinforcement Learning! ๐Ÿš€(1/n)

In the last two months, RAGEN has powered Agent RL training frameworks for over 300,000 people.
Now, weโ€™re introducing VAGENโ€”the first open-source framework that trains *Visual* Agents using multi-turn Reinforcement Learning! ๐Ÿš€(1/n)
Kangrui Wang (@james_kkw) 's Twitter Profile Photo

Super excited to introduce VAGEN!! We trained a 3B VLM agent in Sokoban and it can sometimes solve 6-step game! Honored be part of the team!

Shujin Wu (@shujin_wu) 's Twitter Profile Photo

๐Ÿ‡Introducing Alice, our most recent work on advancing weak-to-strong generalization! Instead of students passively absorbing what teachers feed them, Alice puts stronger student models in the driver's seat - it incentivizes student models to self-generate supervision based on

๐Ÿ‡Introducing Alice, our most recent work on advancing weak-to-strong generalization! Instead of students passively absorbing what teachers feed them, Alice puts stronger student models in the driver's seat - it incentivizes student models to self-generate supervision based on
Shizhe Diao (@shizhediao) 's Twitter Profile Photo

Thrilled to share my first project at NVIDIA! โœจ Todayโ€™s language models are pre-trained on vast and chaotic Internet texts, but these texts are unstructured and poorly understood. We propose CLIMB โ€” Clustering-based Iterative Data Mixture Bootstrapping โ€” a fully automated

Thrilled to share my first project at NVIDIA! โœจ

Todayโ€™s language models are pre-trained on vast and chaotic Internet texts, but these texts are unstructured and poorly understood. We propose CLIMB โ€” Clustering-based Iterative Data Mixture Bootstrapping โ€” a fully automated
Wei Liu โœˆ๏ธ ICLR2025 (@weiliu99) 's Twitter Profile Photo

โ€œWhat is the answer of 1 + 1?โ€ Large Reasoning Models (LRMs) may generate 1500+ tokens just to answer this trivial question. Too much thinking ๐Ÿคฏ Can LRMs be both Faster AND Stronger? Yes. Introducing LASER๐Ÿ’ฅ: Learn to Reason Efficiently with Adaptive Length-based Reward Shaping

โ€œWhat is the answer of 1 + 1?โ€
Large Reasoning Models (LRMs) may generate 1500+ tokens just to answer this trivial question.
Too much thinking ๐Ÿคฏ
Can LRMs be both Faster AND Stronger?
 Yes.
Introducing LASER๐Ÿ’ฅ: Learn to Reason Efficiently with Adaptive Length-based Reward Shaping
Young-Jun Lee (@passing2961) 's Twitter Profile Photo

๐Ÿšจ #Alert Our recent work is accepted in #ICCV2025. ๐ŸŽ‰ Huge thanks to our second author ๐Ÿ’› Byung-Kwan Lee (Byung-Kwan Lee), one of the best #KAIST colleagues I've ever seen, and our third author ๐Ÿ’› Jianshu Zhang (Jianshu Zhang), as well as amazing collaborations with #KAIST, #NAVER,

๐Ÿšจ #Alert

Our recent work is accepted in #ICCV2025.

๐ŸŽ‰ Huge thanks to our second author ๐Ÿ’› Byung-Kwan Lee (<a href="/BKLEE_NANO/">Byung-Kwan Lee</a>), one of the best #KAIST colleagues I've ever seen, and our third author ๐Ÿ’› Jianshu Zhang (<a href="/SterZhang/">Jianshu Zhang</a>), as well as amazing collaborations with #KAIST, #NAVER,
Zhaochen Su (@suzhaochen0110) 's Twitter Profile Photo

Excited to share our new survey on the reasoning paradigm shift from "Think with Text" to "Think with Image"! ๐Ÿง ๐Ÿ–ผ๏ธ Our work offers a roadmap for more powerful & aligned AI. ๐Ÿš€ ๐Ÿ“œ Paper: arxiv.org/pdf/2506.23918 โญ GitHub (400+๐ŸŒŸ): github.com/zhaochen0110/Aโ€ฆ

Excited to share our new survey on the reasoning paradigm shift from "Think with Text" to "Think with Image"! ๐Ÿง ๐Ÿ–ผ๏ธ
Our work offers a roadmap for more powerful &amp; aligned AI. ๐Ÿš€
๐Ÿ“œ Paper: arxiv.org/pdf/2506.23918
โญ GitHub (400+๐ŸŒŸ): github.com/zhaochen0110/Aโ€ฆ
May Fung (@may_f1_) 's Twitter Profile Photo

๐Ÿง  How can AI evolve from statically ๐˜ต๐˜ฉ๐˜ช๐˜ฏ๐˜ฌ๐˜ช๐˜ฏ๐˜จ ๐˜ข๐˜ฃ๐˜ฐ๐˜ถ๐˜ต ๐˜ช๐˜ฎ๐˜ข๐˜จ๐˜ฆ๐˜ด โ†’ dynamically ๐˜ต๐˜ฉ๐˜ช๐˜ฏ๐˜ฌ๐˜ช๐˜ฏ๐˜จ ๐˜ธ๐˜ช๐˜ต๐˜ฉ ๐˜ช๐˜ฎ๐˜ข๐˜จ๐˜ฆ๐˜ด as cognitive workspaces, similar to the human mental sketchpad? ๐Ÿ” Whatโ€™s the ๐—ฟ๐—ฒ๐˜€๐—ฒ๐—ฎ๐—ฟ๐—ฐ๐—ต ๐—ฟ๐—ผ๐—ฎ๐—ฑ๐—บ๐—ฎ๐—ฝ from tool-use โ†’ programmatic

๐Ÿง  How can AI evolve from statically ๐˜ต๐˜ฉ๐˜ช๐˜ฏ๐˜ฌ๐˜ช๐˜ฏ๐˜จ ๐˜ข๐˜ฃ๐˜ฐ๐˜ถ๐˜ต ๐˜ช๐˜ฎ๐˜ข๐˜จ๐˜ฆ๐˜ด โ†’ dynamically ๐˜ต๐˜ฉ๐˜ช๐˜ฏ๐˜ฌ๐˜ช๐˜ฏ๐˜จ ๐˜ธ๐˜ช๐˜ต๐˜ฉ ๐˜ช๐˜ฎ๐˜ข๐˜จ๐˜ฆ๐˜ด as cognitive workspaces, similar to the human mental sketchpad?
๐Ÿ” Whatโ€™s the ๐—ฟ๐—ฒ๐˜€๐—ฒ๐—ฎ๐—ฟ๐—ฐ๐—ต ๐—ฟ๐—ผ๐—ฎ๐—ฑ๐—บ๐—ฎ๐—ฝ from tool-use โ†’ programmatic
Jianshu Zhang โœˆ๏ธICLR2025๐Ÿ‡ธ๐Ÿ‡ฌ (@sterzhang) 's Twitter Profile Photo

Excited to witness a new breakthrough in linking cues across multi-image, which shows performance boost in our VLM2-Bench! ๐Ÿ‘๐Ÿป Welcome check this paper out as well as explore new approaches that can achieve higher performance in our VLM2-Bench!

May Fung (@may_f1_) 's Twitter Profile Photo

Heading out to #ACL2025 in Vienna with six main/finding papers to present! ๐Ÿ‡ฆ๐Ÿ‡นโœˆ๏ธ๐Ÿคฉ Would love to chat about research on multimodal model reasoning and agent, as well as opportunities in my group HKUST NLP. Please DM if you'd like to meet!

Heading out to #ACL2025 in Vienna with six main/finding papers to present! ๐Ÿ‡ฆ๐Ÿ‡นโœˆ๏ธ๐Ÿคฉ

Would love to chat about research on multimodal model reasoning and agent, as well as opportunities in my group <a href="/hkustNLP/">HKUST NLP</a>.

Please DM if you'd like to meet!
May Fung (@may_f1_) 's Twitter Profile Photo

HKUST NLP UIUC NLP ACL 2025 [1/n] "๐˜”๐˜ข๐˜ต๐˜ค๐˜ฉ๐˜ช๐˜ฏ๐˜จ ๐˜ค๐˜ถ๐˜ฆ๐˜ด ๐˜ง๐˜ฐ๐˜ณ ๐˜ช๐˜ฅ๐˜ฆ๐˜ฏ๐˜ต๐˜ช๐˜ค๐˜ข๐˜ญ ๐˜ฐ๐˜ฃ๐˜ซ๐˜ฆ๐˜ค๐˜ต๐˜ด, ๐˜ฅ๐˜ช๐˜ด๐˜ต๐˜ช๐˜ฏ๐˜ค๐˜ต ๐˜ข๐˜ต๐˜ต๐˜ณ๐˜ช๐˜ฃ๐˜ถ๐˜ต๐˜ฆ๐˜ด ๐˜ง๐˜ฐ๐˜ณ ๐˜ถ๐˜ฏ๐˜ช๐˜ฒ๐˜ถ๐˜ฆ ๐˜ฐ๐˜ฏ๐˜ฆ๐˜ด." Such ๐™˜๐™ง๐™ค๐™จ๐™จ-๐™˜๐™ค๐™ฃ๐™ฉ๐™š๐™ญ๐™ฉ ๐™ซ๐™ž๐™จ๐™ช๐™–๐™ก ๐™ง๐™š๐™–๐™จ๐™ค๐™ฃ๐™ž๐™ฃ๐™œ is extremely simple and straightforward for the human cognitive process,

<a href="/hkustNLP/">HKUST NLP</a> <a href="/uiuc_nlp/">UIUC NLP</a> <a href="/aclmeeting/">ACL 2025</a> [1/n] "๐˜”๐˜ข๐˜ต๐˜ค๐˜ฉ๐˜ช๐˜ฏ๐˜จ ๐˜ค๐˜ถ๐˜ฆ๐˜ด ๐˜ง๐˜ฐ๐˜ณ ๐˜ช๐˜ฅ๐˜ฆ๐˜ฏ๐˜ต๐˜ช๐˜ค๐˜ข๐˜ญ ๐˜ฐ๐˜ฃ๐˜ซ๐˜ฆ๐˜ค๐˜ต๐˜ด, ๐˜ฅ๐˜ช๐˜ด๐˜ต๐˜ช๐˜ฏ๐˜ค๐˜ต ๐˜ข๐˜ต๐˜ต๐˜ณ๐˜ช๐˜ฃ๐˜ถ๐˜ต๐˜ฆ๐˜ด ๐˜ง๐˜ฐ๐˜ณ ๐˜ถ๐˜ฏ๐˜ช๐˜ฒ๐˜ถ๐˜ฆ ๐˜ฐ๐˜ฏ๐˜ฆ๐˜ด." Such ๐™˜๐™ง๐™ค๐™จ๐™จ-๐™˜๐™ค๐™ฃ๐™ฉ๐™š๐™ญ๐™ฉ ๐™ซ๐™ž๐™จ๐™ช๐™–๐™ก ๐™ง๐™š๐™–๐™จ๐™ค๐™ฃ๐™ž๐™ฃ๐™œ is extremely simple and straightforward for the human cognitive process,
Canyu Chen (@canyuchen3) 's Twitter Profile Photo

Excited to speak at today's Agentic AI Summit! Happy to catch up if you also attend! ๐Ÿ“ Frontier Stage ๐Ÿ“…4:50pm PT "Lightning Talks" Session ๐Ÿ”—Project website: agent-trust.camel-ai.org ๐Ÿ”—Slides: drive.google.com/file/d/1zC2hm0โ€ฆ

Excited to speak at today's Agentic AI Summit! Happy to catch up if you also attend!

๐Ÿ“ Frontier Stage 
๐Ÿ“…4:50pm PT "Lightning Talks" Session

๐Ÿ”—Project website: agent-trust.camel-ai.org
๐Ÿ”—Slides: drive.google.com/file/d/1zC2hm0โ€ฆ
Jianshu Zhang โœˆ๏ธICLR2025๐Ÿ‡ธ๐Ÿ‡ฌ (@sterzhang) 's Twitter Profile Photo

Life update: Today marks the beginning of my PhD journey at Northwestern University! Excited for the road ahead. ๐ŸŽ“๐Ÿ’œ #PhDlife #Northwestern

Manling Li (@manlingli_) 's Twitter Profile Photo

World Model Reasoning for VLM Agents (NeurIPS 2025, Score 5544) We release VAGEN to teach VLMs to build internal world models via visual state reasoning: - StateEstimation: what is the current state? - TransitionModeling: what is next? MDP โ†’ POMDP shift to handle the partial

Zihan Wang - on RAGEN (@wzihanw) 's Twitter Profile Photo

๐Ÿš€Excited to share our NeurIPS 2025 paper VAGEN, a scalable RL framework that trains VLM agents to reason as world models. VLM agents often act without tracking the world: they lose state, fail to anticipate effects, and RL wobbles under sparse, late rewards. Our solution is

Andrej Karpathy (@karpathy) 's Twitter Profile Photo

I quite like the new DeepSeek-OCR paper. It's a good OCR model (maybe a bit worse than dots), and yes data collection etc., but anyway it doesn't matter. The more interesting part for me (esp as a computer vision at heart who is temporarily masquerading as a natural language