#TianMin
LightNews LightNews
nataliepeluso.com
Bottom-up and generative computations both uniquely explain social brain responses

New preprint from Manasi Malik, Minjae Kim, Tianmin Shu, Shari Liu, Leyla Isik

#neuroskyence #socialcognition

www.biorxiv.org/content/10.6...
www.biorxiv.org
February 22, 2026 at 10:22 PM
goldkorn.bsky.social
Here is a classic, scary, and funny Chinese New Year's animation from the last of the golden days of the Chinese creative internet in 2011. More links and background in following tweet 1/2
哐哐哐之新年贺卡 小兔乖乖
YouTube video by Tianmin Qu
www.youtube.com
January 28, 2025 at 2:55 PM
spmontecarlo.bsky.social
www.arxiv.org/abs/2602.07632
'Scalable Mean-Field Variational Inference via Preconditioned Primal-Dual Optimization'
- Jinhua Lyu, Tianmin Yu, Ying Ma, Naichen Shi
February 11, 2026 at 1:49 PM
paper.bsky.social
[6/30] 577 Upvotes, 5 Comments, 3 Posts, arXiv:2609.06986

🆕Continual Learning Mechanisms Compose for Long-Horizon Memorization

Zheyuan Zhang, Alvin Zhang, Daniel Khashabi, Tianmin Shu
Language models may need to internalize information that arrives over time and retain it through many subsequent updates. To study this challenge, we introduce long-horizon memorization, a setting in which a model learns 100 query-answer tasks through continual supervised fine-tuning without retaining earlier training examples or receiving task identifiers at inference. Sequential updates cause catastrophic forgetting, and no single continual learning mechanism we evaluate maintains strong retention at this horizon. We hypothesize that mechanisms addressing complementary sources of forgetting will be more effective when composed. We organize these compositions along two design dimensions. Data, function, and weight anchors specify what prior information each update should preserve, while low-rank allocation rules determine where successive updates are retained. To test this hypothesis systematically, we construct three distinct 100-task memorization datasets. We introduce task-level successive halving to search the combinatorial design space and use a factorial experiment to measure individual and interaction effects. Our best method combines all three anchors with merged LoRA, ranks among the top 3 methods in all datasets, and raises average final retention from 1.2% under naive sequential fine-tuning to 34.9%, a 28-fold improvement. The data anchor and merged LoRA provide the largest average gains and interact super-additively on all three datasets. Together, these results show that composing complementary mechanisms substantially improves long-horizon memorization beyond what any individual mechanism achieves.
September 27, 2026 at 2:45 PM
arjun.bsky.social
This guy.
The man in front of the tanks from Tianmin Square.
June 4, 2024 at 1:58 PM
jhucompsci.bsky.social
Congratulations to @anjalief.bsky.social, @gligoric.bsky.social, @ghadfield.bsky.social, Chien-Ming Huang, @tiziano.bsky.social, and Tianmin Shu on winning 2026 @jhu.edu Discovery Awards! Learn more about their proposed projects below:
CS faculty receive 2026 Johns Hopkins Discovery Awards
Winning projects—chosen from 324 proposals—will accelerate the research of 132 individuals from across the university.
www.cs.jhu.edu
July 28, 2026 at 6:51 PM
athman.bsky.social
No tengo ni pajolera idea de lo que dicen porque no entiendo el chino ni sé de qué hablan en concreto, pero se puede aplicar a cualquier estado opresor. Israel es lo primero que me viene a la mente, pero vale para todo régimen fascista y totalitario.
Eat the rich!
youtu.be/rnw5BvxSDmM?...
哐哐哐之新年贺卡 小兔乖乖
YouTube video by Tianmin Qu
youtu.be
January 28, 2025 at 10:37 AM
drkoraly.bsky.social
August's Editor's Choice:

An asynchronous, hands-off workflow for looking time experiments with infants

Gal Raz, Sabrina Piccolo, Janine Medrano, Shari Liu, Kirsten Lydic, Catherine Mei, Victoria Nguyen, Tianmin Shu, & Rebecca Saxe

@rebeccasaxe.bsky.social

psycnet.apa.org/record/2024-...
APA PsycNet
psycnet.apa.org
August 14, 2024 at 2:41 PM
ghadfield.bsky.social
2/ With Tianmin Shu, Seth Lazar and Dylan Hadfield-Menell: we're building institutions that let real communities, not model developers, define how AI behaves.
April 15, 2026 at 6:45 PM
entropydrivenmindset.win
Zhuo Tianmin clashes with Xi Yazhou over forced resettlement, accusing him of mimicking wartime tactics—but is it a smokescreen?

Ch.571 Initial Results
https://youtu.be/JRDsiQu1_DE?t=8207s

#MingDynasty #1600s #HistoricalFiction
Illumine Lingao — Chapter 571 Initial Results
Zhuo Tianmin clashes with Xi Yazhou over forced resettlement, accusing him of mimicking wartime tactics—but is it a smokescreen?
youtu.be
August 11, 2026 at 7:09 PM
arxiv-cs-cl.bsky.social
Xiaoyu Deng, Ye Zhang, Tianmin Guo, Yongzhe Zhang, Zhengjian Kang, Hang Yang
ChallengeMe: An Adversarial Learning-enabled Text Summarization Framework
https://arxiv.org/abs/2502.05084
February 10, 2025 at 6:36 AM
mathpr-bot.bsky.social
Reza Gheissari, Curtis Grant, Tianmin Yu: Mixing times of Langevin dynamics for spiked matrix models https://arxiv.org/abs/2604.20008 https://arxiv.org/pdf/2604.20008 https://arxiv.org/html/2604.20008
April 23, 2026 at 6:49 AM
tmlr-pub.bsky.social
Augmented Vision-Language Models: A Systematic Review

Anthony C Davis, Burhan A. Sadiq, Tianmin Shu, Chien-Ming Huang

Action editor: Li Fuxin

https://openreview.net/forum?id=DFnPi77v6J

#visual #vision #trained
March 19, 2026 at 8:20 AM
clio2.bsky.social
Unless we are prepared to be the one standing in front of a tank in Tianmin Square, with no guarantee--or it becomes a literal civil war with organized forces on both sides.
October 8, 2025 at 4:43 PM
badsciencemonk.bsky.social
I got that when I asked about Tianmin Square, the Uighur and, bizarrely, to name a politician who looks like Winnie The Poo.

In each case it started to give an answer before redacting it.
January 28, 2025 at 8:57 AM
paper.bsky.social
[23/30] 76 Likes, 2 Comments, 2 Posts
2412.09624, cs․CV | cs․RO, 12 Dec 2024

🆕GenEx: Generating an Explorable World

Taiming Lu, Tianmin Shu, Junfei Xiao, Luoxin Ye, Jiahao Wang, Cheng Peng, Chen Wei, Daniel Khashabi, Rama Chellappa, Alan Yuille, Jieneng Chen
December 18, 2024 at 12:05 AM
jhucompsci.bsky.social
In “SimWorld: An Open-ended Simulator for Agents in Physical and Social Worlds,” Tianmin Shu, @marstin.bsky.social, @zhitinghu.bsky.social, @lianhui.bsky.social, and more introduce a new simulator for developing and evaluating agents in rich, real-world-like settings 🌎: (4/17)
NeurIPS Poster SimWorld: An Open-ended Simulator for Agents in Physical and Social WorldsNeurIPS 2025
neurips.cc
December 1, 2025 at 11:22 PM
jhucompsci.bsky.social
🧡 Welcome, new faculty! 🧡
Erik Rye, Gagan Garg, Vladimir Braverman, and Kai Presler-Marshall post together on the Gertrude's patio behind a round cocktail table. Haimin Hu, Tianmin Shu, and Ziang Xiao hold beverages on the Gertrude's patio. Other faculty can be seen mingling in the background. Laura Henneman and Megan Dakwa pose together on the Gertrude's patio. Erik Rye and Nicolas Christianson talk as Renjie Zhao and Haimin Hu eat in the background.
September 18, 2026 at 1:17 PM
ghadfield.bsky.social
1/ Who decides how AI systems behave? With Tianmin Shu, @sethlazar.org, and @dhadfieldmenell.bsky.social, we were named runner-up in the LaudeInstitute's inaugural Moonshot program and awarded a seed grant to find out. www.laude.org/moonshots
Laude | Moonshots
We asked the most consequential AI researchers in the world how they would use AI to solve humanity's hardest problems. 125 proposals and 600 researchers later, meet the Moonshots // ONE awardees.
www.laude.org
April 15, 2026 at 6:45 PM
entropydrivenmindset.win
As the Senate vote threatens to gridlock the nation, I'm listening to a story about a critical design flaw that could bring an entire project to a halt.

Ch.72 Water and Power Part 2
https://youtu.be/-5JtARw84oU?t=20229s

#ChineseHistory #audiobook #MingDynasty
Illumine Lingao — Chapter 72 Water and Power Part 2
Zhuo Tianmin demands a rebuild, risking delays—but Wang Luobin’s inspection reveals a critical design flaw threatening the entire hydroelectric project. Will they...
youtu.be
March 17, 2026 at 1:00 PM
hopkinsengineer.bsky.social
Live from #CVPR2025! @jhu-bdps.bsky.social's Alan Yuille sharing his computer vision research at Hopkins Engineering's booth 1317. Stop by to learn how we're harnessing the power and promise of AI, data science and machine learning. Check out our faculty in attendance:
A large group in a convention booth watching as a professor describes his work, gesturing towards slides on a screen. Promotional flyer for “Data Science and Artificial Intelligence at Hopkins – Meet the Experts” event, listing speakers, their photos, affiliations, session times, and research areas. Featured experts include Tinoosh Mohsenin (edge AI, robotics), Rama Chellappa (AI, computer vision), Alan Yuille (AI, cognitive science), Vishal Patel (deep learning, biometrics), Reno Kriz (natural language generation), and Tianmin Shu (embodied AI). Hosted by Johns Hopkins University.
June 13, 2025 at 5:52 PM
markparthun.bsky.social
Amazing work from Dr. Tianmin Fu from the Department of Biological Chemistry and Pharmacology at Ohio State University.

www.cell.com/cell/abstrac...
DdmDE eliminates plasmid invasion by DNA-guided DNA targeting
As a component of a bacterial anti-plasmid defense system, DdmE is an Argonaute-like protein that functions as a DNA-guided DNA-targeting module. It recruits DdmD, an autoinhibited dimer with helicase...
www.cell.com
November 27, 2024 at 4:07 AM
nanoatp.bsky.social
[7/30] 492 Upvotes, 5 Comments, 3 Posts, arXiv:2609.06986

🆕Continual Learning Mechanisms Compose for Long-Horizon Memorization

Zheyuan Zhang, Alvin Zhang, Daniel Khashabi, Tianmin Shu
Language models may need to internalize information that arrives over time and retain it through many subsequent updates. To study this challenge, we introduce long-horizon memorization, a setting in which a model learns 100 query-answer tasks through continual supervised fine-tuning without retaining earlier training examples or receiving task identifiers at inference. Sequential updates cause catastrophic forgetting, and no single continual learning mechanism we evaluate maintains strong retention at this horizon. We hypothesize that mechanisms addressing complementary sources of forgetting will be more effective when composed. We organize these compositions along two design dimensions. Data, function, and weight anchors specify what prior information each update should preserve, while low-rank allocation rules determine where successive updates are retained. To test this hypothesis systematically, we construct three distinct 100-task memorization datasets. We introduce task-level successive halving to search the combinatorial design space and use a factorial experiment to measure individual and interaction effects. Our best method combines all three anchors with merged LoRA, ranks among the top 3 methods in all datasets, and raises average final retention from 1.2% under naive sequential fine-tuning to 34.9%, a 28-fold improvement. The data anchor and merged LoRA provide the largest average gains and interact super-additively on all three datasets. Together, these results show that composing complementary mechanisms substantially improves long-horizon memorization beyond what any individual mechanism achieves.
September 18, 2026 at 12:04 AM
nanoatp.bsky.social
[6/30] 483 Upvotes, 3 Comments, 3 Posts, arXiv:2609.06986

🆕Continual Learning Mechanisms Compose for Long-Horizon Memorization

Zheyuan Zhang, Alvin Zhang, Daniel Khashabi, Tianmin Shu
Language models may need to internalize information that arrives over time and retain it through many subsequent updates. To study this challenge, we introduce long-horizon memorization, a setting in which a model learns 100 query-answer tasks through continual supervised fine-tuning without retaining earlier training examples or receiving task identifiers at inference. Sequential updates cause catastrophic forgetting, and no single continual learning mechanism we evaluate maintains strong retention at this horizon. We hypothesize that mechanisms addressing complementary sources of forgetting will be more effective when composed. We organize these compositions along two design dimensions. Data, function, and weight anchors specify what prior information each update should preserve, while low-rank allocation rules determine where successive updates are retained. To test this hypothesis systematically, we construct three distinct 100-task memorization datasets. We introduce task-level successive halving to search the combinatorial design space and use a factorial experiment to measure individual and interaction effects. Our best method combines all three anchors with merged LoRA, ranks among the top 3 methods in all datasets, and raises average final retention from 1.2% under naive sequential fine-tuning to 34.9%, a 28-fold improvement. The data anchor and merged LoRA provide the largest average gains and interact super-additively on all three datasets. Together, these results show that composing complementary mechanisms substantially improves long-horizon memorization beyond what any individual mechanism achieves.
September 17, 2026 at 12:06 AM
paper.bsky.social
[14/30] 483 Likes, 3 Comments, 3 Posts
2609.06986, cs․LG, 07 Sep 2026

🆕Continual Learning Mechanisms Compose for Long-Horizon Memorization

Zheyuan Zhang, Alvin Zhang, Daniel Khashabi, Tianmin Shu
Language models may need to internalize information that arrives over time and retain it through many subsequent updates. To study this challenge, we introduce long-horizon memorization, a setting in which a model learns 100 query-answer tasks through continual supervised fine-tuning without retaining earlier training examples or receiving task identifiers at inference. Sequential updates cause catastrophic forgetting, and no single continual learning mechanism we evaluate maintains strong retention at this horizon. We hypothesize that mechanisms addressing complementary sources of forgetting will be more effective when composed. We organize these compositions along two design dimensions. Data, function, and weight anchors specify what prior information each update should preserve, while low-rank allocation rules determine where successive updates are retained. To test this hypothesis systematically, we construct three distinct 100-task memorization datasets. We introduce task-level successive halving to search the combinatorial design space and use a factorial experiment to measure individual and interaction effects. Our best method combines all three anchors with merged LoRA, ranks among the top 3 methods in all datasets, and raises average final retention from 1.2% under naive sequential fine-tuning to 34.9%, a 28-fold improvement. The data anchor and merged LoRA provide the largest average gains and interact super-additively on all three datasets. Together, these results show that composing complementary mechanisms substantially improves long-horizon memorization beyond what any individual mechanism achieves.
September 17, 2026 at 12:04 AM