Chuanneng Sun

Chuanneng Sun, Songjun Huang, Dario Pompili

Large Language Models (LLMs) have demonstrated remarkable abilities in various language tasks, making them promising candidates for decision-making in robotics. Inspired by Hierarchical Reinforcement Learning (HRL), we propose Retrieval-Augmented in-context reinforcement Learning (RAHL), a novel framework that decomposes complex tasks into sub-tasks using an LLM-based high-level policy, in which a complex task is decomposed into sub-tasks by a high-level policy on-the-fly. The sub-tasks, defined by goals, are assigned to the low-level policy to complete. To improve the agent's performance in multi-episode execution, we propose Hindsight Modular Reflection (HMR), where, instead of reflecting on the full trajectory, we let the agent reflect on shorter sub-trajectories to improve reflection efficiency. We evaluated the decision-making ability of the proposed RAHL in three benchmark environments--ALFWorld, Webshop, and HotpotQA. The results show that RAHL can achieve an improvement in performance in 9%, 42%, and 10% in 5 episodes of execution in strong baselines. Furthermore, we also implemented RAHL on the Boston Dynamics SPOT robot. The experiment shows that the robot can scan the environment, find entrances, and navigate to new rooms controlled by the LLM policy.

Chuanneng Sun, Songjun Huang, Dario Pompili

In recent years, Large Language Models (LLMs) have shown great abilities in various tasks, including question answering, arithmetic problem solving, and poem writing, among others. Although research on LLM-as-an-agent has shown that LLM can be applied to Reinforcement Learning (RL) and achieve decent results, the extension of LLM-based RL to Multi-Agent System (MAS) is not trivial, as many aspects, such as coordination and communication between agents, are not considered in the RL frameworks of a single agent. To inspire more research on LLM-based MARL, in this letter, we survey the existing LLM-based single-agent and multi-agent RL frameworks and provide potential research directions for future research. In particular, we focus on the cooperative tasks of multiple agents with a common goal and communication among them. We also consider human-in/on-the-loop scenarios enabled by the language component in the framework.

Ning Liu, Chuanneng Sun, Kristina Klinkner et al.

Direct Preference Optimization (DPO) aligns language models using pairwise preference comparisons, offering a simple and effective alternative to Reinforcement Learning (RL) from human feedback. However, in many practical settings, training data consists of multiple rollouts per prompt, inducing rich preference structure that pairwise DPO fails to exploit. Collapsing such data into independent pairs discards transitivity, introduces redundant or conflicting supervision, and can lead to unstable optimization. We propose Graph Direct Preference Optimization (GraphDPO), a principled generalization of DPO that operates over directed acyclic preference graphs induced by rollout rankings. GraphDPO encodes dominance relations as edges and optimizes a graph-structured Plackett--Luce-inspired objective that aggregates supervision over graph neighborhoods, enforcing transitivity while recovering standard DPO as a special case. To handle discrete or sparse signals, we introduce an equivalence-class construction where responses with identical preferences form graph layers, and intra-layer edges contribute zero loss, preventing spurious gradients. Despite leveraging full graph structure, GraphDPO maintains linear per-prompt complexity via efficient log-sum-exp aggregation. We further incorporate optional ground-truth anchoring by inserting verified solutions as dominant nodes and applying an annealed schedule that stabilizes early training while gradually relaxing oracle supervision. Experiments on reasoning and program synthesis tasks demonstrate superior performance, suggesting that graph-structured preference modeling is a scalable and robust alternative to pairwise and listwise alignment objectives.

Chuanneng Sun, Zeeshan Ahmed, Yingyi Ma et al.

This paper studies contextual biasing with Large Language Models (LLMs), where during second-pass rescoring additional contextual information is provided to a LLM to boost Automatic Speech Recognition (ASR) performance. We propose to leverage prompts for a LLM without fine tuning during rescoring which incorporate a biasing list and few-shot examples to serve as additional information when calculating the score for the hypothesis. In addition to few-shot prompt learning, we propose multi-task training of the LLM to predict both the entity class and the next token. To improve the efficiency for contextual biasing and to avoid exceeding LLMs' maximum sequence lengths, we propose dynamic prompting, where we select the most likely class using the class tag prediction, and only use entities in this class as contexts for next token prediction. Word Error Rate (WER) evaluation is performed on i) an internal calling, messaging, and dictation dataset, and ii) the SLUE-Voxpopuli dataset. Results indicate that biasing lists and few-shot examples can achieve 17.8% and 9.6% relative improvement compared to first pass ASR, and that multi-task training and dynamic prompting can achieve 20.0% and 11.3% relative WER improvement, respectively.