
AI research lab @nyuniversity led by @mengyer | Partially run by students | Also find us https://teapaw.org/t.co/IHXOiESm6f on bsky
New York
Joined November 2024
- Tweets91
- Following311
- Followers1.3K
- Likes1.7K
Pinned Tweet
Agentic Learning AI Lab @agentic_ai_lab presents at #ICML2026 in Seoul✈️🇰🇷 Come check out our latest research!
📍Main Conference Papers
Temporal Straightening for Latent Planning
Tue Jul 7 10:30am Hall A #1509
Context Tuning for In-Context Optimization
Wed Jul 8 5pm Hall A #2510
📍Workshop Papers
The Surprising Effectiveness of Deleting Weights in LLM Reasoning and Adaptation
Fri Jul 10 Rm 318 | Foundations of Deep Generative Models Workshop
Seeking the Unfamiliar but Memorable: Conceptual Creativity as Meta-Learning
Sat Jul 11 Rm 401 | Human-AI Co-Creativity Workshop
Aligning LLMs with Human Uncertainty: A Beta-Bernoulli Calibrator for LLM Forecasting
Sat Jul 11 Gr. Ballroom 104–105 | Forecasting as a New Frontier of Intelligence Workshop
Alive and Predicting: A Live Evaluation of Multi-Step Forecasting Agents
Sat Jul 11 Gr. Ballroom 104–105 | Forecasting as a New Frontier of Intelligence Workshop
See threads below for details!🧵👇
agentic learning ai lab retweeted
HyperThink code and weights are out! 👀👀
Train hypernetworks to write per-question weight updates that let LLMs skip their long thinking traces. Thanks @kdgyun425 for leading the release!
github.com/GitGyun/HyperThin…
What if LLMs could think in weights, not tokens?
Introducing HyperThink (#COLM2026, co-led with @kdgyun425): a hypernetwork writes a per-question weight update to the LLM for it to skip the long thinking trace, a step toward accurate System 1 thinking! 🧵👇
📍 Tomorrow at @COLM_conf! Oct 6 (Tue), 4:30 to 6:30 PM, board #97, Franciscan C
What does a JEPA world model's latent space look like? Introducing a new regularizer, Slow SVReg that learns a smooth latent map from watching first person videos. New work by @mengyer @yingwww_ @choang333 @KuangYilun @ylecun
Can a JEPA learn a latent map of the world?
We released a new paper with @yingwww_ @choang333 @KuangYilun @ylecun 🧵
agentic learning ai lab retweeted
Labs disagree on vision encoders for MLLMs! ⚔️
Gemma 4 12B scrapped the encoder.
Kimi K3 dropped encoder pretraining.
Inkling 2 Small revived the encoder Inkling killed.
We trained 200+ models in a controlled study to settle it:
✅ Keep vision parameters, skip the pretraining. And they don't have to be an encoder! 🤔
Paper: arxiv.org/abs/2611.17334 🧵
agentic learning ai lab retweeted
CDS Asst. Prof. Mengye Ren (@mengyer) and his lab released @outerloop_sci, open-source software whose AI agents open pull requests only when their experiments improve a benchmark.
The agents got a GPT model to its target in 77% fewer training steps.
outerloop.science/
agentic learning ai lab retweeted
Context Tuning (ICML 2026) helped language models learn from a few examples by refining their stored context, with model weights held fixed.
Developed by Courant PhD student Jack Lu (@Jacklu_me) and collaborators.
nyudatascience.medium.com/he…
1/2
Was curious if a BF16 trained mask would still hold up on quantized weights, so I ran a quick test using the Qwen2.5-7B GRPO mask
Even when you zero those weights directly inside the FP8 and INT4 checkpoints without touching the scales, almost the entire boost carries over
(on MATH-500, FP8 goes 64.2% -> 74.0% and INT4 goes 60.4% -> 74.2%, vs 74.8% in BF16).
also side finding: in INT4, 25.8% of the 3.16M weights the mask deletes were already rounded to 0 by quantization alone
Even the simplest weight edit, a deletion, can teach an LLM new skills!
Meet Deletion Tuning (#NeurIPS2026): just zeroing <0.1% of weights matches full fine-tuning's accuracy at scale. It's thousands of times smaller and avoids forgetting, a step toward continual learning. 🧵👇
New from our lab, led by @Jacklu_me and @yangzb159 with @yanii’s group at @UCalgaryML: zeroing <0.1% of an LLM’s weights matches full fine-tuning at scale and avoids forgetting!
The Surprising Effectiveness of Deleting Weights in LLM Post-Training ✂️: arxiv.org/abs/2611.10866
Even the simplest weight edit, a deletion, can teach an LLM new skills!
Meet Deletion Tuning (#NeurIPS2026): just zeroing <0.1% of weights matches full fine-tuning's accuracy at scale. It's thousands of times smaller and avoids forgetting, a step toward continual learning. 🧵👇
agentic learning ai lab retweeted
The CDS November Research Feature is here! In her paper "AdaJEPA: An Adaptive Latent World Model," Ying Wang (@yingwww_ ) introduces a world model that keeps learning after training. Read more on AdaJEPA & other CDS research in this month’s feature! t.e2ma.net/webview/130xei/dd…
agentic learning ai lab retweeted
New paper: Outerloop: Open and Trustworthy Multi-Agent Autoresearch. In most open autoresearch systems, the agent grades its own results. Outerloop is a kernel that runs agents as untrusted programs and keeps measurement and review outside their sandboxes.
Will P vs NP be resolved before 2030?
forecast.agenticlearning.ai/…
agentic learning ai lab retweeted
NYU Center for Data Science is hiring Faculty Fellows! Apply by Nov 24 apply.interfolio.com/193498
CDS is hiring Faculty Fellows for fall 2027!
Two-year appointments with independent research, research and travel funds, and one course to teach per year.
Apply by November 24 for full consideration: apply.interfolio.com/193498
agentic learning ai lab retweeted
Outerloop 0.6.0 is out. It now runs locally on a Mac, can use a local git forge instead of GitHub, and keeps a hash-chained, attested record of each agent run. Run "outerloop trajectory <run-id> --verify" to check one.
agentic learning ai lab retweeted
World models can fail when the world changes.
That's why CDS PhD student Ying Wang (@yingwww_), CDS founding director Yann LeCun (@ylecun) & CDS Assistant Professor Mengye Ren (@mengyer) built AdaJEPA to adapt during planning.
nyudatascience.medium.com/a-…
agentic learning ai lab retweeted
New #NVIDIA paper: what if robots never had to wait for their own thoughts?
TacitVLA learns to reason in chain-of-thought or subtasks, but its actions never read them.
→ Keeps the gains of reasoning at up to 3.4× lower latency on 4 benchmarks, and tops RoboMME and RMBench 🧵👇
Come say hi to Jack at #COLM and learn about our joint work with @hongseu33’s group on HyperThink, skipping slow reasoning by turning each question into a fast weight update!
HyperThink: Text-to-Parameter Hypernetworks for Efficient Reasoning
🕸️: hyperthink-llm.github.io/
What if LLMs could think in weights, not tokens?
Introducing HyperThink (#COLM2026, co-led with @kdgyun425): a hypernetwork writes a per-question weight update to the LLM for it to skip the long thinking trace, a step toward accurate System 1 thinking! 🧵👇
📍 Tomorrow at @COLM_conf! Oct 6 (Tue), 4:30 to 6:30 PM, board #97, Franciscan C
11
762
agentic learning ai lab retweeted
What if LLMs could think in weights, not tokens?
Introducing HyperThink (#COLM2026, co-led with @kdgyun425): a hypernetwork writes a per-question weight update to the LLM for it to skip the long thinking trace, a step toward accurate System 1 thinking! 🧵👇
📍 Tomorrow at @COLM_conf! Oct 6 (Tue), 4:30 to 6:30 PM, board #97, Franciscan C
agentic learning ai lab retweeted
Outerloop 0.3.0 is out: run autoresearch agents fully on your own hardware. Here, Claude Code and Codex authors run on a self-hosted GLM-5.3.
pip install outerloop-science==0.3.0
agentic learning ai lab retweeted
Honored to receive the 2nd Annual Google ML and Systems Junior Faculty Award for our work on adaptable ML systems and agents. Grateful to @Google for the support, and to my students. More to come on adaptable agents, in open source and open science.
















