Proactive User-in-the-Loop Clarification for High-Ambiguity Daily Tasks
Add a clarification action to the model's policy and train it via reinforcement learning to trigger dialogue when encountering high-ambiguity daily GUI tasks.
Time
Compute level
Resources
Prerequisites
Scope
The Gap
The AndroidDaily benchmark results (Section 4, Table 6) demonstrate a significant performance drop for Step-GUI-8B when moving from low-ambiguity tasks to high-ambiguity tasks (success rates falling by over 12 points).
Limitation
The current action space (Section 2.2.1) is purely execution-oriented (Click, Swipe, Type) and lacks a mechanism for the agent to identify underspecified user intents and seek clarification.
Proposed Solution
Extend the model's action space to include a 'CLARIFY' action. Train the model using RLVR with a reward function that penalizes incorrect terminal actions in high-uncertainty states but provides a positive reward for triggering a clarification dialogue when the task instruction entropy exceeds a threshold. This requires integrating a multi-turn conversation head into the GUI-MCP protocol.
Why This Matters
Practical GUI agents cannot assume perfect user instructions; the ability to ask for missing parameters (e.g., 'Which email address should I send this to?') is critical for real-world reliability.
How to Get Started
- Extend the model's discrete action space to include a 'CLARIFY' token within the GUI-MCP protocol.
- Synthesize a dataset of high-ambiguity instructions paired with appropriate clarification dialogue trees.
- Implement a reinforcement learning reward function that penalizes incorrect terminal actions and rewards proper clarification triggering.
- Fine-tune Step-GUI-8B using the reward function and evaluate success rates on the AndroidDaily benchmark.
Code linked to this paper on Hugging Face: github.com/stepfun-ai/gelab-zero
Suggested by AI. Adapt it to your setup.
Has This Been Done?
KnowU-Bench: Towards Interactive, Proactive, and Personalized Mobile Agent Evaluation and Implicit Turn-Wise Policy Optimization for Proactive User-LLM Interaction explore proactive behaviors and interactive agent evaluations. However, they do not specifically integrate a reinforcement learning mechanism with explicit task instruction entropy thresholds and a 'CLARIFY' action head inside a GUI-MCP framework as proposed.
- Implicit Turn-Wise Policy Optimization for Proactive User-LLM Interaction · 2026 · 5 citations
- KnowU-Bench: Towards Interactive, Proactive, and Personalized Mobile Agent Evaluation · 2026 · 27 citations
We searched OpenAlex for “Add a clarification action to the model's policy and train it via reinforcement learning to trigger dialogue when encountering high-ambiguity daily GUI tasks.” and “Proactive User-in-the-Loop Clarification for High-Ambiguity Daily Tasks” and an AI compared the idea with the top 10 results. A small search can miss relevant work.
Source Paper
Step-GUI Technical ReportHaolong Yan, Jia Wang, Xin Huang +94 more · 2025
The paper introduces Step-GUI, a family of multimodal models (4B/8B) specifically optimized for GUI automation through a self-evolving training pipeline. This pipeline uses a Calibrated Step Reward System (CSRS) to convert model-generated trajectories into high-quality training data at significantly lower costs while introducing GUI-MCP for secure, cross-platform deployment. Evaluation on the new AndroidDaily benchmark shows Step-GUI achieves state-of-the-art performance in real-world mobile and desktop scenarios.
Related Ideas
- Dense Process Reward Optimization for Long-Horizon GUI AgentsMobile-Agent-v3: Foundamental Agents for GUI Automation🟠Medium
- Verifiable Tree-Search with Partial State FeedbackDeepPlanning: Benchmarking Long-Horizon Agentic Planning with Verifiable Constraints🟠Medium
- Task-Specific Reward Experts for MR-GRPODeepGen 1.0: A Lightweight Unified Multimodal Model for Advancing Image Generation and Editing🟡Fairly easy
- Asynchronous RL for Multi-Turn Agentic TasksAReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning🟡Fairly easy
- Safety-Tuned Process Reward Models for Infeasible Task RefusalEnterpriseOps-Gym: Environments and Evaluations for Stateful Agentic Planning and Tool Use in Enterprise Settings🟠Medium