Teng Xiao
I work at the Allen Institute for AI (AI2) and the University of Washington with Prof. Noah A. Smith and Prof. Hanna Hajishirzi. I completed my PhD at Pennsylvania State University, advised by Prof. Vasant Honavar.
I am interested in machine learning and reinforcement learning. I am currently working on: (i) Alignment and Reasoning for LLM Agents; (ii) Long-horizon LLM Agent for Open-ended Tasks. My work is regularly published at ICML, ICLR, and NeurIPS.
News
- 2026Context Language Models is out: language models that natively manage their own context (code, blog).
- 2026Two new preprints on agent harnesses: Rethinking the Evaluation of Harness Evolution for Agents and Co-Harness.
- 2026Simple Denoising Diffusion Language Models accepted at ICML 2026.
- 2026Meta-Reinforcement Learning with Self-Reflection for Agentic Search accepted at COLM 2026.
- 2026Tmax: A Simple Recipe for Terminal Agents is out (blog).
- 2025Olmo 3 is out. I am a core contributor on RL post-training.
More news
- 2025Honorable Mention in AI Dissertations at the Annual Penn State AI Awards — one of three recipients across the university.
- 2025SimPER is used by the EXAONE series models from LG AI Research; the SimPER-trained EXAONE Deep 32B surpasses DeepSeek-R1 (671B) on reasoning benchmarks.
- 2025Joined the Allen Institute for AI as a researcher.
- 2025Defended my PhD at Penn State.
- 2025DSPO selected as an Oral at ICLR 2025 (top 1.8%).
- 2025SimPER and DIL accepted at ICLR 2025.
- 2025Serving as Area Chair for NeurIPS and ACL.
- 2024Cal-DPO accepted at NeurIPS 2024.
- 2022Decoupled Self-supervised Learning for Graphs selected as a Spotlight at NeurIPS 2022 (top 5%).
Selected Publications
Organised by theme; the full list is on Google Scholar. * equal contribution, † advising role.
Self-Evolution for LLMs
-
Context Language Models
Models manage their own context: 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus.
-
Rethinking the Evaluation of Harness Evolution for Agents
Automatic harness evolution fails to beat simple test-time scaling on Terminal-Bench 2.1; task-specific evolution beats search by 80% on ARC-AGI-3.
- Co-Harness: Co-Evolving Harnesses and Model Weights for LLM Agents
- Meta-Reinforcement Learning with Self-Reflection for Agentic Search
Reinforcement Learning for LLMs
-
Tmax: A Simple Recipe for Terminal Agents
33% on Terminal-Bench 2.0 with a 9B model.
- Simple Denoising Diffusion Language Models
-
Olmo 3
Core contributor to RL post-training of AI2's fully open models.
- Internalizing World Models via Self-Play Finetuning for Agentic RL
- Inference-time Alignment in Continuous Space
- DSPO: Direct Score Preference Optimization for Diffusion Model Alignment
-
SimPER: A Minimalist Approach to Preference Alignment without Hyperparameters
Used by LG AI Research's EXAONE models; the SimPER-trained EXAONE Deep 32B surpasses DeepSeek-R1 (671B) on reasoning.
- On a Connection Between Imitation Learning and RLHF
- Cal-DPO: Calibrated Direct Preference Optimization for Language Model Alignment
Structured Knowledge Representation Learning
Academic Service
- Area Chair
- NeurIPS, ACL
- Program Committee & Reviewer
- ICLR (2022–2026), NeurIPS (2022–2025), ICML (2023, 2024), WSDM (2023–2025), AAAI (2022, 2023), SIGIR (2021–2023), ACL ARR (2024), CIKM (2023), TheWebConf (2022, 2023), COLM (2024)
- Journal Reviewer
- ACM Transactions on Intelligent Systems and Technology, ACM Transactions on Information Systems