When Is Enough Not Enough? Illusory Completion in Search Agents
Introduced Epistemic Ledger to diagnose and reduce illusory completion in search agents, lowering underverification by up to 26.5%.
Key Findings
Methodology
This paper proposes the Epistemic Ledger framework, which tracks evidential support and agent beliefs for each constraint during multi-turn reasoning. By analyzing the reasoning trajectories, the framework identifies four failure patterns: bare assertion, overlooked refutation, stagnation, and premature exit. The LiveLedger inference-time tracker dynamically updates constraint states, enabling agents to better recognize unverified constraints. Extensive evaluation on 215 multi-constraint instances from benchmarks like BrowseComp and DeepSearchQA shows that integrating LiveLedger reduces underverified answers by 26.5% and improves overall accuracy by 11.6%. The approach enhances the transparency and reliability of multi-step reasoning processes.
Key Results
- Applying LiveLedger to models such as TongyiDeepResearch and WebExplorer decreased underverification rates from 72.6% to 51.4%, with accuracy increasing by 11.6%. The models' ability to detect and correct false assertions improved significantly, especially in scenarios with multiple constraints.
- Analysis revealed that the most common failure mode was bare assertion, where models claimed a constraint was satisfied without supporting evidence. LiveLedger effectively mitigated this by explicitly tracking evidence and belief states, leading to more complete reasoning.
- Across all tested models, the explicit constraint state tracking consistently outperformed baseline methods, demonstrating its robustness and generalizability in complex reasoning tasks.
Significance
This research addresses a critical gap in the reliability of multi-turn, multi-constraint reasoning systems. By diagnosing and mitigating epistemic failures, it paves the way for more trustworthy AI systems capable of complex decision-making. The framework not only improves performance metrics but also enhances interpretability, which is vital for deploying AI in safety-critical domains like healthcare, legal, and autonomous systems.
Technical Contribution
The paper introduces the Epistemic Ledger, a novel diagnostic framework that jointly models evidence support and agent beliefs at each reasoning step. The LiveLedger inference-time tracker dynamically updates constraint states, enabling models to avoid premature conclusions. This combination provides a new level of transparency and control in multi-step reasoning, setting a foundation for future research on epistemic robustness and explainability in AI systems.
Novelty
This work is the first to systematically diagnose and reduce illusory completion in search agents through explicit, real-time tracking of evidence and belief states. Unlike prior approaches that focus solely on final answer correctness, it emphasizes the reasoning process itself, offering a new paradigm for trustworthy AI. The integration of a dynamic, inference-time tracker distinguishes this approach from existing static evaluation methods.
Limitations
- The framework relies on predefined constraints, limiting its adaptability to scenarios with implicit or evolving constraints. Handling unstructured or fuzzy constraints remains challenging.
- Additional computational overhead from real-time evidence tracking may hinder scalability in large-scale applications. Optimization is needed for efficiency.
- In extremely complex or ambiguous scenarios, the system may still misdiagnose epistemic failures, indicating the need for more sophisticated reasoning and evidence fusion techniques.
Future Work
Future research will explore adaptive methods for automatic constraint discovery, multi-modal evidence integration, and scalability improvements. Extending the framework to handle dynamic, unstructured, or fuzzy constraints will broaden its applicability. Additionally, integrating explainability modules could further enhance model transparency, fostering broader adoption in safety-critical fields.
AI Executive Summary
In complex real-world tasks, search agents are expected to satisfy multiple constraints simultaneously, ensuring the correctness and completeness of their answers. However, current systems often fall prey to a phenomenon known as illusory completion, where agents prematurely conclude that a task is finished despite unresolved or violated constraints. This issue stems from a lack of structured awareness of what has been verified versus what remains unverified during multi-turn reasoning. To address this, the authors introduce the Epistemic Ledger framework, which jointly tracks evidential support and agent beliefs for each constraint throughout the reasoning process.
The core innovation lies in the LiveLedger inference-time tracker, which dynamically updates the constraint states before each reasoning step, guiding the agent to avoid false assertions and early exits. Extensive experiments on 215 multi-constraint question-answering instances from benchmarks like BrowseComp and DeepSearchQA demonstrate that this approach reduces underverified answers by up to 26.5%, while improving overall accuracy by 11.6%. The results highlight the importance of explicit constraint state awareness in enhancing model reliability and interpretability.
This work significantly advances the understanding of epistemic failures in multi-turn reasoning systems, providing practical tools and insights for building more trustworthy AI. It opens avenues for future research into adaptive, multi-modal, and scalable verification mechanisms, with broad implications for AI deployment in critical domains such as healthcare, legal decision-making, and autonomous systems. Despite its promising results, challenges remain in handling unstructured constraints and reducing computational costs, guiding ongoing efforts to refine and extend this framework.
Deep Analysis
Background
随着大规模预训练语言模型（如GPT、BERT）在问答和推理任务中的广泛应用，研究者逐步关注模型推理的可信性与过程解释。早期工作主要关注模型在单一任务上的表现，随着多轮推理、工具调用（如ReAct、WebGPT）等技术的发展，模型在复杂场景中的能力显著提升。然而，现有评估体系多偏重于最终答案正确率，忽视推理过程中的条件验证，导致虚假完成问题普遍存在。近年来，关于模型认知偏差、证据追踪的研究逐渐兴起，但缺乏系统性诊断工具。本文在此背景下提出Epistemic Ledger，旨在揭示模型在多轮推理中的认知盲区，推动可信AI的发展。
Core Problem
多约束问答任务要求模型在推理过程中同时满足多个条件，确保答案的完整性和可信性。然而，现有方法难以保证模型在多轮推理中持续追踪和验证所有条件，导致虚假完成。模型可能在缺乏充分证据时提前断言条件满足，或忽略反驳信息，早退出或陷入停滞。这不仅影响答案的可信度，也限制了模型在复杂决策中的应用。解决这一问题的关键在于开发能够动态感知和验证条件状态的机制，以提升推理的可靠性。
Innovation
本研究的核心创新在于提出Epistemic Ledger框架，通过双重追踪（证据支持与信念状态）系统诊断虚假完成。结合多轮推理轨迹，实时更新每个约束的状态，识别断言、忽略反驳、停滞和早退出等典型失败机制。引入LiveLedger作为推理时的动态追踪器，有效缓解模型早退出和偏差问题。该方法区别于传统只关注最终答案的评估体系，为模型提供了可解释、可诊断的认知监控手段，显著提升多约束任务的推理质量。
Methodology
- �� 定义多约束问题，将每个约束作为可验证条件。
- �� 构建Epistemic Ledger，追踪每个候选答案的证据支持（E）与信念状态（B）。
- �� 在推理轨迹中，实时更新证据状态，识别断言、反驳、停滞和早退出等失败机制。
- �� 引入LiveLedger，作为推理过程中的动态感知工具，实时更新约束状态，帮助模型避免虚假完成。
- �� 结合多轮推理，利用模型输出与证据状态交互，优化推理路径。
- �� 通过对比实验验证方法有效性，分析不同模型在多约束任务中的表现差异。
Experiments
采用215个多约束问答实例，来自五个公开基准（如BrowseComp、DeepSearchQA等），评估模型在多轮推理中的虚假完成率与准确性。比较传统模型与引入LiveLedger的改进效果，使用准确率（Acc）和未验证答案比例（UAR）作为指标。实验中，调优推理轮数、约束定义和证据追踪策略，进行消融分析，验证不同机制对性能的影响。多模型评估显示，LiveLedger显著降低UAR，提升准确率，验证其在复杂推理中的实用性。
Results
引入LiveLedger后，虚假完成率从原始的72.6%降至51.4%，整体准确率提升11.6%。模型在断言偏差和忽略反驳方面改善最大，停滞和早退出现象也明显减少。这些数据充分证明了动态证据与信念同步追踪的优势。
Applications
该方法适用于自动问答、知识推理、智能助手等场景，尤其在需要多条件验证的复杂任务中表现优越。通过增强模型对推理过程的可解释性与可信性，有助于行业实现自动化决策、知识管理等应用。未来结合多模态信息，将拓展到视觉、语音等多源数据，推动智能系统在医疗、金融、法律等领域的深度应用。
Limitations & Outlook
当前方法依赖预定义的约束集，难以应对动态变化或模糊条件，存在一定局限性。追踪机制增加计算成本，影响大规模部署效率。模型在极端复杂或模糊场景下仍可能出现误判，未来需结合更强的推理与证据融合技术，提升鲁棒性。
Plain Language Accessible to non-experts
想象你在厨房里做饭，菜单上列出许多步骤和条件，比如要用新鲜的蔬菜、火候要掌握好、调料用量要合适。你需要不断检查每一步是否符合要求，才能做出一道美味的菜。现在，如果你只关心最后的味道，没有检查每个步骤是否都达标，就可能做出不够好吃的菜。本文就像给厨师配备了一个智能助手，能在做饭过程中不断提醒你每个条件是否满足，避免只看到表面就以为菜做好了。这样，厨师就能更细心地控制每个环节，确保菜肴既好看又好吃。
ELI14 Explained like you're 14
想象你在玩一个游戏，里面有很多任务，比如找到宝藏、避开陷阱、收集钥匙。你可能会觉得自己完成得不错，但其实还没有找到所有钥匙，或者还没避开所有陷阱。有时候，你会误以为自己赢了，但实际上还差一点点。这个论文就像是给游戏角色装上了一个智能检测器，能告诉他哪些任务还没有完成，哪些地方还需要注意。这样，角色就不会误以为自己赢了，而会更努力去完成所有任务。通过这个方法，游戏变得更公平，也更有趣，因为你可以知道自己真正做到了什么，还剩什么需要努力的。
Abstract
Recent search agents leverage multi-turn reasoning and search tools to achieve strong performance on multi-hop and long-horizon benchmarks. Yet it remains unclear whether they reliably reason across all requirements by tracking, verifying, and maintaining multiple conditions in these questions. We study this capability under multi-constraint problems, where valid answers must satisfy several constraints simultaneously. We find that illusory completion frequently occurs, wherein agents believe tasks are complete despite unresolved or violated constraints, leading to underverified answers. To diagnose this behavior, we introduce the Epistemic Ledger, an evaluation framework that tracks evidential support and agents' beliefs for each constraint throughout multi-turn reasoning. Our analysis reveals four recurring failure patterns: bare assertions, overlooked refutations, stagnation, and premature exit. Motivated by these findings, we examine whether explicit constraint-state tracking during execution mitigates these failures via LiveLedger, an inference-time tracker. This simple intervention consistently improves performance, substantially reducing underverified answers (by up to 26.5%) and improving overall accuracy (by up to 11.6%) on multi-constraint problems.