My work spans language-model training and evaluation, multi-agent learning, and human and machine cognition.
* indicates equal contribution. See also Google Scholar.
Under review at ICLR 2027
U-WIN enables agents to explore new environments and train on self-written notes and self-generated trajectories without access to downstream tasks, rewards, or correctness feedback. Across legal, codebase, and math environments, downstream performance improved near log-linearly as training scaled from 10M to 100M tokens.
COLM 2026
RL-based mid-training with dense outcome- and process-level rewards from a reference-guided LLM judge. On held-out AIME 2026 with Qwen3-4B, ExpRL improved pass@1 over the strongest baseline at each stage by 3.6 points after priming and 4.7 points after identical sparse-GRPO post-training.
ICML 2026
Identified persistent length, uncertainty, position, sycophancy, and model-style biases in language reward models. Null-space probe projections reduced three of these biases without retraining or loss on RewardBench-2.
EACL 2026 · Long Paper
A creative-writing evaluation benchmark with 43k training pairs and a 2.5k-pair debiased, human-labeled test set. Trained reward models improved human-preference agreement by 5 points over the strongest zero-shot judge.
NeurIPS 2025 Efficient Reasoning Workshop · Spotlight
A GRPO-compatible adaptive length penalty reduced DeepScaleR-1.5B generation length by more than 50% at matched accuracy across MATH-500, AIME, and OlympiadBench, while allocating 5.35× more tokens to hard prompts than easy ones.
NeurIPS 2025 Datasets & Benchmarks Track · Spotlight
An execution harness for 212 research-code challenges from 20 recent ML papers. Across 32 evaluated models, the strongest reached 33.0% on the HARD split, and 43 tasks remained unsolved by every model.
Preprint, 2025
A training roadmap combining linearized search traces, process supervision, synthetic data, and reinforcement learning, supported by evidence of in-context search behavior in frontier reasoning models.
ICLR 2025 · Poster
An LLM agent with a theory-of-mind scaffold that outperformed multi-agent RL baselines on 3 of 4 environments and 6 of 8 eight-agent scenarios across 30 Melting Pot tasks.
NeurIPS 2023 ALOE Workshop · Poster
Comparing centralized, decentralized, asynchronous, and self-supervised multi-agent RL pipelines reveals how centralized-information assumptions can overstate performance in realistic mixed-motive settings.
Computational Cognitive Neuroscience 2025 Proceedings
An 868-hour longitudinal egocentric-video dataset spanning ages 5 months to 3 years, with annotations and evaluation pipelines for speech transcription, diarization, pose estimation, and out-of-distribution transfer.
Behavior Research Methods, Volume 56, 2024
NeurIPS 2022 Datasets & Benchmarks Track
Naturalistic real-time and lifelong visual-learning benchmarks identify memory mechanisms as critical for learning from sparse, low-diversity input streams.
Cognition, 2020
Proceedings on Privacy Enhancing Technologies, 2021
Proceedings of the Annual Meeting of the Cognitive Science Society, 2024
Manuscript, 2024