active information gathering

This work addresses the challenge in goal-oriented dialogue systems of simultaneously optimizing information gathering and goal confirmation over multi-turn interactions. Existing approaches often suffer from either insufficient flexibility due to rigid predefined structures or limited long-horizon decision-making capabilities when relying solely on large language models. To overcome this trade-off, the paper proposes the CUP framework, which explicitly models uncertainty as a guiding signal for sequential decision-making. CUP synergistically combines the action generation capacity of large language models with a structured planner that evaluates the long-term informational value of actions, thereby jointly optimizing exploration and commitment. Experimental results demonstrate that the method significantly improves task success rates and reduces dialogue turns across multiple benchmarks, effectively breaking the longstanding tension between flexibility and long-horizon planning.

This study addresses how agents balance immediate action against seeking clarifying information under uncertainty. The authors propose a computational model based on expected regret that formally characterizes, for the first time, how contextual uncertainty and action costs jointly influence human decisions to ask clarification questions. Using an experimental paradigm that integrates both linguistic clarification and non-linguistic action choices, the research demonstrates that individuals’ propensity to seek clarification increases with the potential loss associated with erroneous actions, thereby validating the proposed rational trade-off mechanism. These findings reveal a cognitive strategy wherein humans proactively reduce uncertainty in high-stakes situations to avoid significant losses, offering both theoretical and empirical support for understanding metacognitive decision-making in communicative contexts.

To address the insufficient decision robustness of LLM-based planning agents in partially observable environments—caused by discrepancies between internal dynamics and real-world conditions—this paper proposes InfoSeeker, the first framework enabling large language model agents to proactively seek information under model-reality mismatch. InfoSeeker supports hypothesis testing, environmental change detection, and cognitive calibration. It jointly models task planning and information-seeking actions, establishing a decision architecture with self-verification and dynamic adaptation capabilities. Evaluated on both custom and established benchmarks—including robotic manipulation and web navigation—InfoSeeker achieves a 74% absolute performance gain over baseline methods, maintains high sample efficiency, and is compatible with multiple mainstream LLMs.

In hierarchical reasoning, agents often veer into incorrect branches due to undetected information gaps at intermediate decision points. This work proposes the ACTION-RATING framework, which integrates clarification requests into the action space by sharing an ordinal rating scale with navigation actions, enabling help-seeking behaviors to directly compete with and be observable alongside execution actions at every decision point. The approach reveals two emergent information-seeking strategies—compulsory and opportunistic—and, through a controlled answer channel design, empirically disentangles the location of help requests from answer quality. Evaluated on the HTS classification task, the method improves information-seeking effectiveness from 50% to 74% and boosts top-10 coding accuracy by up to 16.2%, with strategy structures remaining stable even as answer quality degrades.

This work addresses the challenge faced by large language model agents in adaptively balancing the trade-offs between taking action and seeking clarification when handling ambiguous user requests—a problem exacerbated by existing approaches that either rely on task-specific hyperparameter tuning or ignore variations in risk. To overcome this limitation, the study introduces the theory of Value of Information (VoI) into human-AI interaction and proposes a parameter-free decision framework. During inference, the framework dynamically evaluates the expected utility gain from asking a question against the associated user cognitive cost, enabling context-aware, adaptive clarification strategies. By explicitly balancing task risk, query ambiguity, and user burden, the method matches or surpasses the best manually tuned baselines across four diverse tasks—20 Questions, medical diagnosis, flight booking, and e-commerce—with utility improvements of up to 1.36 points in high-risk scenarios.

This study addresses the challenge that large language models face in eliciting critical user information concealed due to privacy concerns, ambiguity, or social inhibitions during open-ended conversations. To tackle this, the work formulates the task as a sequential decision-making problem and introduces RPS, a lightweight reinforcement learning framework that dynamically selects optimal prompts from a predefined pool to adaptively guide users toward disclosure. The key contributions include a learnable prompt selection policy, the construction of ILegal—the first benchmark dataset for information elicitation grounded in real legal documents—and empirical validation demonstrating that RPS significantly outperforms static prompting baselines in both synthetic environments and the ILegal dataset. These results establish the efficacy of policy-driven adaptive prompting for effective information elicitation.

This work addresses the poorly understood “aha” mechanism in large language model (LLM) reasoning, particularly the role of uncertainty expression in facilitating effective inference. We propose an information-theoretic framework that decouples reasoning into procedural information processing and cognitive verbalization, formally characterizing the latter as a key mechanism for achieving informational sufficiency—explicitly externalizing uncertainty to drive continuous information acquisition and downstream control. Integrating information-theoretic analysis, programmatic modeling, and quantification of verbalization, our empirical results demonstrate that strong reasoning capabilities stem not from specific surface tokens but from the external articulation of uncertainty. Cognitive verbalization effectively overcomes informational stagnation, offering a unified account of insight-like “aha” phenomena and post-training behaviors, thereby establishing a novel paradigm for designing reasoning-capable models.

This work addresses the challenge of overconfident and erroneous responses from large language models in high-stakes domains such as healthcare and law, often triggered by ambiguous user queries. To mitigate this, the authors propose a novel active information acquisition framework that integrates domain-specific document retrieval with strategic follow-up questioning. For the first time, Dempster–Shafer evidence theory is incorporated into this process, enabling interpretable fusion of incomplete or conflicting multi-source information through a structured evidence network. This approach explicitly models uncertainty and prevents premature decision-making. Experimental results demonstrate that the method outperforms strong baselines on legal and medical tasks, significantly enhancing decision reliability and system efficiency while reducing the number of required interaction rounds.

Existing planning tools struggle to align users’ surface-level requests with their deeper goals and lack mechanisms for sustained tracking, reflective guidance, and accountability. This work proposes a novel personalized planning system that integrates conversational AI for reflective prompting, a persistent goal database, and a multi-source information synchronization processor to dynamically refine recommendations through continuous behavioral pattern analysis. The system innovatively couples goal articulation, long-term tracking, and need-alignment mechanisms to effectively bridge the gap between intention and action. Evaluation employs a phased strategy combining automated testing, simulated user interactions, and longitudinal human studies, demonstrating significant improvements in goal attainment rates, plan realism, and alignment between goals and personal values.

This study addresses the challenge of generating natural language explanations that effectively guide human navigation and decision-making in uncertain environments. We propose the first utility-driven procedural explanation framework, which translates natural language explanations into policy priors and value maps for planning agents operating in partially observable settings. Explanation quality is quantified through path efficiency and replanning frequency. Integrating large language models, POMDP-based planning algorithms, and a preregistered behavioral experiment, we collected 1,200 explanation utterances across 24 maps. Results demonstrate that high-quality explanations significantly improve human navigation performance compared to both low-quality explanations and a no-explanation baseline, establishing—for the first time—a direct link between executable planning performance and explanation quality.