Publications

A scalable, fully automated curriculum-learning framework for RL post-training of LLMs. Learns a neural curator that selects training problems to directly maximize expected policy improvement, formulated as a non-stationary bandit with regret guarantees. Delivers +28.6% on AIME2024, +30.5% on ARC-1D, and up to 80% speedup over the strongest baseline. Outstanding Paper Award and Oral (top 3.9% of accepted papers) at the ICLR 2026 Logical Reasoning Workshop.

A data-augmented AutoML market that automatically augments a buyer's training data with external datasets and prices the resulting model by its measured performance gain rather than compute. Pricing is grounded in the instrumental value of the sourced data, which sidesteps direct dataset pricing, mitigates strategic buyer behavior, and accommodates diverse valuations through menu-based options.

Frames data selection as a sequential decision-making problem, with data values acting as encodings of the optimal selection order. Shows that methods like Data Shapley are myopic linear approximations of this problem, explains when they fail as utility curvature increases, and introduces a bipartite graph-based surrogate for scalable greedy selection with provable guarantees, evaluated on classical ML benchmarks and large-scale LLM fine-tuning data selection.

Adaptive, compute-aware reasoning - dynamically expands or contracts thought steps for efficient problem-solving. Delivers +33% accuracy vs DeepSeek-R1 at equal tokens and 5-10% lower error on APPS, MATH, and LiveCodeBench.

LLM + MCTS hybrid that learns to optimize its own decision process through self-play and strategy refinement. Outperforms RL and existing LLM agents, achieves human-level performance in complex games. Highlighted in the State of AI Report 2024 as a breakthrough in autonomous agent design.

Introduces Retrospection-Only Fine-Tuning (ROFT): an agent attempts a task, writes a retrospective explanation, and is fine-tuned on that explanation alone - no verifier or reward-based update. With Qwen3.5-4B, reaches 49.2% on SWE-bench Verified and 26.8% on SWE-bench Pro after 20 updates, beating GRPO after 40 updates.

Introduces JAZ, a minimal LLM agent framework built on a single language primitive, invoke: a function whose body is written by the LLM at call time, with recursive calls and the full interaction history available as variables in a Python REPL. With prompting alone and no dedicated memory or meta-harness, JAZ beats Letta (MemGPT) by 8% at half the cost on long-horizon StuLife recall and beats ACE by 4% at lower cost on AppWorld self-improvement. Oral (top 2% of submissions) at the NeurIPS 2026 Meta-Agents Workshop.

Reframes code generation as black-box optimization: uses Scattering, Foresting, and Scouting to boost exploration + feedback exploitation. Achieves SOTA on HumanEval, MBPP, APPS, CodeContests, LeetCode (e.g. 87.2% pass@1 on HumanEval) while halving iterations vs prior methods.

Decomposes language-described environments into intuitive components, enabling zero-shot LLM world models for efficient MCTS in multi-agent settings.

Introduces AvalonBench to benchmark LLM agents in strategic social deduction games and highlights the gap between current agents and engineered bots.

Synthesizes compact, high-quality offline RL datasets that retain policy performance comparable to training on full datasets.