Jonathan Light

Jonathan Light, Yuanzhe Liu, Ziniu Hu

Offline reinforcement learning often requires a quality dataset that we can train a policy on. However, in many situations, it is not possible to get such a dataset, nor is it easy to train a policy to perform well in the actual environment given the offline data. We propose using data distillation to train and distill a better dataset which can then be used for training a better policy model. We show that our method is able to synthesize a dataset where a model trained on it achieves similar performance to a model trained on the full dataset or a model trained using percentile behavioral cloning. Our project site is available at https://datasetdistillation4rl.github.io . We also provide our implementation at https://github.com/ggflow123/DDRL .

Jonathan Light, Min Cai, Weiqin Chen et al.

Traditional reinforcement learning and planning typically requires vast amounts of data and training to develop effective policies. In contrast, large language models (LLMs) exhibit strong generalization and zero-shot capabilities, but struggle with tasks that require detailed planning and decision-making in complex action spaces. We introduce STRATEGIST, a novel approach that integrates the strengths of both methods. Our approach leverages LLMs to search and update high-level strategies (as text), which are then refined and executed by low-level Monte Carlo Tree Search (MCTS). STRATEGIST is a generalizable framework to optimize the strategy through population-based self-play simulations without the need for any training data. We demonstrate the effectiveness of STRATEGIST in learning optimal strategies for competitive, multi-turn games with partial information, including Game of Pure Strategy (GOPS) and multi-agent, hidden-identity discussion games like The Resistance: Avalon. Our results show that agents equipped with STRATEGIST outperform those trained with traditional RL methods, other LLM-based skill acquisition techniques, pre-existing LLM agents across both game environments and achieves comparable performance against human players.

Minbiao Han, Jonathan Light, Steven Xia et al.

Organizations often lack sufficient data to effectively train machine learning (ML) models, while others possess valuable data that remains underutilized. Data markets promise to unlock substantial value by matching data suppliers with demand from ML consumers. However, market design involves addressing intricate challenges, including data pricing, fairness, robustness, and strategic behavior. In this paper, we propose a pragmatic data-augmented AutoML market that seamlessly integrates with existing cloud-based AutoML platforms such as Google's Vertex AI and Amazon's SageMaker. Unlike standard AutoML solutions, our design automatically augments buyer-submitted training data with valuable external datasets, pricing the resulting models based on their measurable performance improvements rather than computational costs as the status quo. Our key innovation is a pricing mechanism grounded in the instrumental value - the marginal model quality improvement - of externally sourced data. This approach bypasses direct dataset pricing complexities, mitigates strategic buyer behavior, and accommodates diverse buyer valuations through menu-based options. By integrating automated data and model discovery, our solution not only enhances ML outcomes but also establishes an economically sustainable framework for monetizing external data.

Jonathan Light, Min Cai, Sheng Shen et al.

In this paper, we explore the potential of Large Language Models (LLMs) Agents in playing the strategic social deduction game, Resistance Avalon. Players in Avalon are challenged not only to make informed decisions based on dynamically evolving game phases, but also to engage in discussions where they must deceive, deduce, and negotiate with other players. These characteristics make Avalon a compelling test-bed to study the decision-making and language-processing capabilities of LLM Agents. To facilitate research in this line, we introduce AvalonBench - a comprehensive game environment tailored for evaluating multi-agent LLM Agents. This benchmark incorporates: (1) a game environment for Avalon, (2) rule-based bots as baseline opponents, and (3) ReAct-style LLM agents with tailored prompts for each role. Notably, our evaluations based on AvalonBench highlight a clear capability gap. For instance, models like ChatGPT playing good-role got a win rate of 22.2% against rule-based bots playing evil, while good-role bot achieves 38.2% win rate in the same setting. We envision AvalonBench could be a good test-bed for developing more advanced LLMs (with self-playing) and agent frameworks that can effectively model the layered complexities of such game environments.

Zhengyao Gu, Jonathan Light, Raul Astudillo et al.

Post-training large foundation models with reinforcement learning typically relies on massive and heterogeneous datasets, making effective curriculum learning both critical and challenging. In this work, we propose ACTOR-CURATOR, a scalable and fully automated curriculum learning framework for reinforcement learning post-training of large language models (LLMs). ACTOR-CURATOR learns a neural curator that dynamically selects training problems from large problem banks by directly optimizing for expected policy performance improvement. We formulate problem selection as a non-stationary stochastic bandit problem, derive a principled loss function based on online stochastic mirror descent, and establish regret guarantees under partial feedback. Empirically, ACTOR-CURATOR consistently outperforms uniform sampling and strong curriculum baselines across a wide range of challenging reasoning benchmarks, demonstrating improved training stability and efficiency. Notably, it achieves relative gains of 28.6% on AIME2024 and 30.5% on ARC-1D over the strongest baseline and up to 80% speedup. These results suggest that ACTOR-CURATOR is a powerful and practical approach for scalable LLM post-training.

Dan Zhang, Min Cai, Jonathan Light et al.

Reward models are central to both reinforcement learning (RL) with language models and inference-time verification. However, existing reward models often lack temporal consistency, leading to ineffective policy updates and unstable RL training. We introduce TDRM, a method for learning smoother and more reliable reward models by minimizing temporal differences (TD) for training-time reinforcement learning and inference-time verification. Experiments show that TD-trained process reward models (PRMs) improve performance across Best-of-N (up to 6.6%) and tree-search (up to 23.7%) settings. When combined with Reinforcement Learning with Verifiable Rewards (RLVR), TD-trained PRMs lead to more data-efficient RL -- achieving comparable performance with just 2.5k data to what baseline methods require 50.1k data to attain -- and yield higher-quality language model policies in 8 model variants (5 series), e.g., Qwen2.5-(0.5B, 1,5B), GLM4-9B-0414, GLM-Z1-9B-0414, Qwen2.5-Math-(1.5B, 7B), and DeepSeek-R1-Distill-Qwen-(1.5B, 7B). We release all code at https://github.com/THUDM/TDRM.

Jonathan Light, Wei Cheng, Benjamin Riviere et al.

Inference scaling methods for LLMs often rely on decomposing problems into steps (or groups of tokens), followed by sampling and selecting the best next steps. However, these steps and their sizes are often predetermined or manually designed based on domain knowledge. We propose dynamic decomposition, a method that adaptively and automatically partitions solution and reasoning traces into manageable steps during inference. By more effectively allocating compute -- particularly through subdividing challenging steps and prioritizing their sampling -- dynamic decomposition significantly improves inference efficiency. Experiments on benchmarks such as APPS, MATH, and LiveCodeBench demonstrate that dynamic decomposition outperforms static approaches, including token-level, sentence-level, and single-step decompositions, reducing the pass@10 error rate by 5.0%, 6.7%, and 10.5% respectively. These findings highlight the potential of dynamic decomposition to improve a wide range of inference scaling techniques.

Jonathan Light, Yue Wu, Yiyou Sun et al.

We frame code generation as a black-box optimization problem within the code space and demonstrate how optimization-inspired techniques can enhance inference scaling. Based on this perspective, we propose SCATTERED FOREST SEARCH (SFS), a novel approach that improves solution diversity and better exploits feedback during evolutionary search. Our theoretical analysis illustrates how these methods help avoid local optima during optimization, leading to more efficient exploration. Extensive experiments on HumanEval, MBPP, APPS, CodeContests, and Leetcode reveal significant performance gains. For instance, our method achieves a pass@1 rate of 67.1% on HumanEval+ and 87.2% on HumanEval with GPT-3.5, marking improvements of 8.6% and 4.3% over the state-of-the-art, while also halving the iterations needed to find the correct solution. Furthermore, our approach scales more efficiently than existing search techniques, including tree search, line search, and repeated sampling.

Tianchun Wang, Zichuan Liu, Yuanzhou Chen et al.

Large language model (LLM) scaling inference is key to unlocking greater performance, and leveraging diversity has proven an effective way to enhance it. Motivated by the observed relationship between solution accuracy and meaningful response diversity, we systematically study the effect of prompt diversity in scaling inference. We theoretically explain why diversified sampling improves Best-of-$N$ scaling, showing that responses generated from meaningful diverse prompts after Best-of-$N$ selection exhibit significantly lower error rates than those produced from stationary prompts. To promote solution diversity, we analyze perturbation fidelity and show that moderately relevant perturbations improve performance, providing guidance for effective perturbation design. Further, we present a set of effective perturbations, including task-level and query-level ones, and analyze the conditions under which they succeed. We systematically evaluate diversified sampling across tasks, finding relative gains of 10.8% in EM@100 for reasoning, 9.6% for mathematics, and 9.5% in Pass@100 for code generation.

Jonathan Light, Sixue Xing, Yuanzhe Liu et al.

Effective extraction of the world knowledge in LLMs for complex decision-making tasks remains a challenge. We propose a framework PIANIST for decomposing the world model into seven intuitive components conducive to zero-shot LLM generation. Given only the natural language description of the game and how input observations are formatted, our method can generate a working world model for fast and efficient MCTS simulation. We show that our method works well on two different games that challenge the planning and decision making skills of the agent for both language and non-language based action taking, without any training on domain-specific training data or explicitly defined world model.

Hongliang Chi, Qiong Wu, Zhengyi Zhou et al.

Data selection has emerged as a crucial downstream application of data valuation. While existing data valuation methods have shown promise in selection tasks, the theoretical foundations and full potential of using data values for selection remain largely unexplored. In this work, we first demonstrate that data values applied for selection can be naturally reformulated as a sequential-decision-making problem, where the optimal data value can be derived through dynamic programming. We show this framework unifies and reinterprets existing methods like Data Shapley through the lens of approximate dynamic programming, specifically as myopic reward function approximations to this sequential problem. Furthermore, we analyze how sequential data selection optimality is affected when the ground-truth utility function exhibits monotonic submodularity with curvature. To address the computational challenges in obtaining optimal data values, we propose an efficient approximation scheme using learned bipartite graphs as surrogate utility models, ensuring greedy selection is still optimal when the surrogate utility is correctly specified and learned. Extensive experiments demonstrate the effectiveness of our approach across diverse datasets.