🚨 Excited to announce TRAAC, an online difficulty-adaptive, attention-based method that handles the tradeoff of under & overthinking in reasoning models to improve both accuracy and efficiency. Underthinking ❌: Models terminate reasoning too early on harder problems, leading

🚨 Excited to share Gistify! Often the easiest way to understand large/complicated repos is by playing around with test cases and tracing back through the code that is executed. Gistify tasks models with turning a codebase and an entry-point (e.g. command, unit test) into a

I'll be presenting ✨MAgICoRe✨ virtually tonight at 7 PM ET / 8 AM CST (Gather Session 3)! I'll discuss 3 key challenges in LLM refinement for reasoning, and how MAgICoRe tackles them jointly: 1⃣ Over-correction on easy problems 2⃣ Failure to localize & fix its own errors 3⃣

🚨 Check out our awesome students/postdocs' papers at #EMNLP2025 and say hi to them 👋! Also, I will give a keynote (virtually) on "Attributable, Conflict-Robust, and Multimodal Summarization with Multi-Source Retrieval" at the NewSumm workshop. -- Jaehong (in-person) finished

🚨 Excited to announce Gistify!, where a coding agent must extract the gist of a repository: generate a single, executable, and self-contained file that faithfully reproduces the behavior of a given command (e.g., a test or entrypoint). ✅ It is a lightweight, broadly applicable

Thrilled to have our paper “Data-scarce Behavior Editing of Language Models” accepted at #EMNLPFindings2025! 🎉 We propose TaRot, a gradient-free method to edit LLM behavior efficiently — no retraining or large datasets needed. Super fun collab with Subhabrata Dutta &

🚀 LCS2 Sneak Peek Series for #EMNLPFindings2025 🚀 📝 Data-scarce Behavior Editing of Language Models 👥 @joykiratsingh, Subhabrata Dutta, @Tanmoy_Chak 📌 Paper: https://t.co/QmyNyZ0h4Z 🎥 Video: https://t.co/yjXaJM9T43

It was an honor and pleasure to give a keynote at the 28th European Conference on Artificial Intelligence (#ECAI2025) in beautiful Bologna, and engage in enthusiastic discussions about trustworthy + calibrated agents, collaborative reasoning + privacy, and controllable multimodal

🚨 Excited to share new work on inferring symbolic world models from observations! OneLife can infer world models in stochastic, complex environments by proposing rules via LLM and reweighting code-based environment laws from observations collected in a single interaction

🚨 Excited to share our new work ✨ OneLife ✨, which investigates how an agent can infer executable symbolic world models 🌐 from a single unguided trajectory in a stochastic environment. I’m especially excited about our planning + evaluation contributions: 1️⃣ We support

How can an agent reverse engineer the underlying laws of an unknown, hostile & stochastic environment in “one life”, without millions of steps + human-provided goals / rewards? In our work, we: 1️⃣ infer an executable symbolic world model (a probabilistic program capturing

🚨 New Paper Alert! Introducing SciVideoBench — a comprehensive benchmark for scientific video reasoning! 🔬SciVideoBench: 1. Spans Physics, Chemistry, Biology & Medicine with authentic experimental videos. 2. Features 1,000 challenging MCQs across three reasoning types:

🚨 Thrilled to introduce Self-Improving Demonstrations (SID) for Goal-Oriented Vision-and-Language Navigation — a scalable paradigm where navigation agents learn to explore by teaching themselves. ➡️ Agents iteratively generate and learn from their own successful trajectories ➡️

Landed in Montreal 🇨🇦 for #COLM2025 to present my first-author work on task-conditioned mixed-precision quantization: “Task-Circuit Quantization” (Thursday 11am, Poster Session 5). I'm applying to PhD programs this cycle and am excited to chat about this or other interests (LLM

I am attending #COLM2025 🇨🇦 this week to present our work on: Unit Test Generation: 📅 Oct 8th (Wed), 4:30 PM, #79 RAG with conflicting evidence: 📅 Oct 9th (Thu), 11 AM, #71 PS: I'm on the industry job market for RS roles, so you can reach me via DM or in-person to chat! 😄

✈️ Arrived at #COLM2025 where I'll be helping to present the following 4 papers. I'm also recruiting multiple PhD students for my new lab at UT Austin -- happy to chat about research, PhD applications, or postdoc openings in my former postdoc lab at UNC! -- Learning to Generate

🚨 Check out our awesome students/postdocs' papers at #COLM2025 and say hi to them (several are on the job market or hiring) --> -- Archiki, David are on the post-PhD job market! -- Elias finished his postdoc & is now faculty at UT-Austin CS and looking to admit PhD students!

🚨 "Think the right amount" for improving both reasoning accuracy and efficiency! --> Large reasoning models under-adapt = underthink on hard problems and overthink on easy ones --> ✨TRAAC✨ is an online RL, difficulty-adaptive, attention-based compression method that prunes

🚨 TRAAC uses an online difficulty-adaptive, attention-based compression method to address a core problem in long thinking: an inability to adapt to problem difficulty! Leads to underthinking on hard problems, overthinking on easy ones, reducing accuracy and efficiency. TRAAC

Models often think too much on easy problems and not enough on harder reasoning problems. Our new method ✨TRAAC✨ fixes this by teaching models to adaptively compress their "thinking budget" to the difficulty of the task during GRPO rollouts. Result? The model uses