cs.AIFeb 26, 2026

Requesting Expert Reasoning: Augmenting LLM Agents with Learned Collaborative Intervention

Authors: Zhiming Wang, Jinwei He, Feng Lu

Organizations: State Key Laboratory of VR Technology and Systems, School of CSE, Beihang University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Large Language Model (LLM) based agents excel at general reasoning but often fail in specialized domains where success hinges on long-tail knowledge absent from their training data. While human experts can provide this missing knowledge, their guidance is often unstructured and unreliable, making its direct integration into an agent's plan problematic. To address this, we introduce AHCE (Active Human-Augmented Challenge Engagement), a framework for on-demand Human-AI collaboration. At its core, the Human Feedback Module (HFM) employs a learned policy to treat the human expert as an interactive reasoning tool. Extensive experiments in Minecraft demonstrate the framework's effectiveness, increasing task success rates by 32% on normal difficulty tasks and nearly 70% on highly difficult tasks, all with minimal human intervention. Our work demonstrates that successfully augmenting agents requires learning how to request expert reasoning, moving beyond simple requests for help.

Figures & tables

Figure 1: Illustration of the multi-layered knowledge gap in domain-specific tasks. An agent’s failure can stem from a missing factual rule (top path, trying to mine stone without a pickaxe) or a missing strategic heuristic (bottom path, not knowing to dig down for stone). Our approach (bottom path) leverages on-demand human expertise to address both types of failures, enabling successful task completion.
Figure 2: The architecture of our MP5-core baseline, adapted from MP5 for a rigorous zero-shot evaluation. Key modifications include excising the Knowledge Memory to ensure true zero-shot planning, and replacing the original Performer with the vision-only MineDreamer [ 28 ] to operate without privileged game data.
Figure 3: Two primary failure modes for the autonomous agent. a planning failure from a domain-knowledge gap (left, mining stone without a pickaxe), and an execution failure from an environmental trap (right, searching for wood in a desert).
Figure 4: An overview of our AHCE framework, which prioritizes autonomous self-correction (the red loop). When a critical impasse is detected, the system seamlessly transitions to solicit targeted human feedback (the green loop), enabling the agent to overcome knowledge gaps and achieve the task.
Figure 5: The operational logic of the Problem Identification Module (Top) and Query Execution Modules (Bottom) .
Figure 6: The overview of HFM. (a) The GRPO pipeline. (b) The detail of the rollout generation process.
Task Level# Reasoning StepsExample Task
Easy1–3craft crafting table
Normal4–5craft wooden sword
Hard6–9craft stone pickaxe
Table 1: Categorization of tasks in Minecraft by level and complexity, defined by the approximate number of reasoning steps (i.e., minimum sub-objectives) required for completion.
MethodEasyMediumHard
Success RateHuman Time (s)Total Time (s)Human RatioSuccess RateHuman Time (s)Total Time (s)Human RatioSuccess RateHuman Time (s)Total Time (s)Human Ratio
MP5-core ∗ [ 17 ]100%0251.40%64%0373.00%10%0-0%
AHCE-log100%0251.40%86%81.0499.516.2%68%310.11513.720.5%
AHCE-Qwen-7B-Instruct100%0251.40%94%57.2439.813.0%78%122.31325.89.2%
AHCE-Qwen-32B-Instruct100%0251.40%96%32.7433.47.5%82%79.41265.66.3%
Table 2: Main experimental results comparing our AHCE methods against baselines across tasks of varying difficulty. ( ∗ ) indicates that this method is implemented by ourselves, for details see section 3 .
Figure 7: Ablation study on the failure threshold nmaxn_{\text{max}} , which governs the agent’s autonomy before requesting human help. We evaluate the trade-off between task success rate and human interaction time on two tasks: medium difficulty (craft wooden sword) and hard difficulty (craft stone pickaxe). As nmaxn_{\text{max}} increases, the agent acts more autonomously, reducing human effort but risking failure—especially on complex tasks.

Explore similar work

CardsList

  1. Agora: Enhancing LLM Agent Reasoning Via Auction-Based Task Allocation

    Jul 10, 2026Kaiji Zhou, Aleš Leonardis, Yue FengLLM Agent OrchestrationLLM Agent Routing

    University of Birmingham

  2. AgentCollab: A Self-Evaluation-Driven Collaboration Paradigm for Efficient LLM Agents

    Mar 27, 2026Wenbo Gao, Renxi Liu, Xian Wang +8LLM AgentsLLM Agent Workflow Optimization

    Huawei · Hong Kong Polytechnic University

  3. Hierarchical Experimentalist Agents

    Jun 28, 2026Abhranil Chandra, Utsav Dhanuka, Varun Gandhi +2LLM Agent Skill LearningAI Agent Benchmarks

    University of Massachusetts Amherst