Rongqian Chen
Publications
IntentScore: Intent-Conditioned Action Evaluation for Computer-Use Agents
Computer-Use Agents (CUAs) leverage large language models to execute GUI operations on desktop environments, yet they generate actions without evaluating action quality, leading to irreversible errors that cascade through subsequent steps. We propose IntentScore, a plan-aware reward model that learns to score candidate actions from 398K offline GUI interaction steps spanning three operating systems. IntentScore trains with two complementary objectives: contrastive alignment for state-action relevance and margin ranking for action correctness. Architecturally, it embeds each candidate's planning intent in the action encoder, enabling discrimination between candidates with similar actions but different rationales. IntentScore achieves 97.5% pairwise discrimination accuracy on held-out evaluation. Deployed as a re-ranker for Agent S3 on OSWorld, an environment entirely unseen during training, IntentScore improves task success rate by 6.9 points, demonstrating that reward estimation learned from heterogeneous offline trajectories generalizes to unseen agents and task distributions.
Knowing When to Ask: Resolving Uncertainty in Human-Robot Joint Planning via Explicit Dialogue and Implicit Intent Cues
Effective human-robot collaboration in open-world environments requires joint planning under uncertainty about the task, the environment, and the human teammate. Communication is the most direct means of resolving such uncertainty, yet most existing systems support only one-way communication: robots listen and act, treating humans as passive supervisors rather than conversational teammates capable of two-way dialogue. We propose a unified human-robot joint planning system in which the robot actively resolves uncertainty through two complementary communication channels. When uncertainty is decision-critical, an uncertainty-mitigation joint planning module engages the human in clarification dialogue: it grounds ambiguous instructions via an LLM-assisted active elicitation mechanism, enumerates traversability hypotheses through a hypothesis-augmented A* search, and computes a cost-optimal querying policy via dynamic programming, so that the robot asks only the questions whose answers actually matter for the plan. When explicit dialogue is unnecessary or impractical, a real-time intent-aware collaboration module instead reads implicit, nonverbal cues, maintaining a probabilistic belief over the human's latent task intent from spatial and directional signals to enable coordination-aware task selection without any communication overhead. We validate the proposed system in both Gazebo simulations and real-world UAV deployments, integrated with a voice dialogue interface and a Vision-Language Model (VLM)-based 3D semantic perception pipeline. Experimental results show that cost-optimal clarification dialogue cuts the interaction cost by 51.9% while maintaining a 100% task success rate, and implicit intent reading reduces the cooperative task execution time by 25.4% compared to the baselines.
Agent Alpha: Tree Search Unifying Generation, Exploration and Evaluation for Computer-Use Agents
While scaling test-time compute through trajectory-level sampling has significantly improved Graphical User Interface (GUI) agents, the lack of regressive ability prevents the reuse of partial successes and the recovery from early missteps. In this paper, we introduce Agent Alpha, a unified framework that synergizes generation, exploration, and evaluation through step-level Monte Carlo Tree Search (MCTS). It enables active modeling or exploiting structures of the planning space. By integrating alpha-UCT guided search into the interaction loop, Agent Alpha enables deliberate planning, facilitating early pruning of suboptimal branches and efficient prefix reuse. We also employ comparison-driven evaluation to mitigate absolute scoring biases and diversity-constrained expansion to maintain a compact, informative search space. Regret bound of alpha-UCT is analyzed. On the OSWorld benchmark, Agent Alpha achieves a state-of-the-art success rate of $\sim 77\%$, significantly outperforming trajectory-level baselines under equivalent compute.