Papers by Mutian Bao
CARL: Constraint-Aware Reinforcement Learning for Planning with LLMs (2026.findings-acl)
Copied to clipboard
Qiuyi Qi, Jinjian Zhang, Mutian Bao, Tian Liang, Guocong Li, Dongnan Liu, Wei Zhou, Jie Liu, Ming Kong, Linjian Mo, Feng Zhang, Qiang Zhu
| Challenge: | Existing approaches to constraint-aware planning fail to enhance the model’s intrinsic focus on constraints. |
| Approach: | They propose a constraint-aware reinforcement learning framework that encourages constraint focus and penalizes neglect of LLMs. |
| Outcome: | The proposed framework outperforms existing frameworks and state-of-the-art reasoning models in a number of real-world applications. |
UrbanGeoEval: A City-Scale Benchmark for Evaluating Large Language Models in Geospatial Reasoning (2026.acl-long)
Copied to clipboard
| Challenge: | Extensive experiments on 18 widely used LLMs uncover critical insights: (1) models exhibit severe geographic biases and resolution gaps; (2) failures in complex multi-hop tasks stem from brittle foundational spatial skills rather than high-level logic deficits. |
| Approach: | They propose a dual-module framework that disentangles factual recall and spatial logic from the model's real capabilities in urban environments. |
| Outcome: | Extensive tests on 18 widely used LLMs reveal that models exhibit severe geographic biases and resolution gaps, and failures in complex multi-hop tasks often stem from brittle foundational spatial skills rather than high-level logic deficits. |
STAPO: Selective Trajectory-Aware Policy Optimization for LLM Agent Training (2026.acl-long)
Copied to clipboard
Qiuyi Qi, Tian Liang, Mutian Bao, Jinjian Zhang, Dongnan Liu, Wei Zhou, Linjian Mo, Ming Kong, Jie Liu, Feng Zhang, Qiang Zhu
| Challenge: | Prior work has explored step-level supervision using Shannon-entropy-based uncertainty signals, which conflate inherent state complexity with agent confidence. |
| Approach: | They propose a hierarchical group-based RL framework that leverages normalized entropy to locate outlier steps associated with trajectory neglect and optimizes them via a mechanism of trajectory-aware reward and trajectory-independent penalty. |
| Outcome: | Experiments on ALFWorld, WebShop, and Search-Augmented QA show that STAPO achieves state-of-the-art performance while substantially alleviating trajectory neglect. |