Research map / Multi-agent systems and games / LLM agents
HiconAgent: History Context-aware Policy Optimization for GUI Agents
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HCPO, an RL framework for GUI agents that samples variable-length histories (DCS) and trains a dual-branch policy with visual-history-dropped compressed branch aligned by KL (AHC), yielding a 3B model that matches or beats 7B baselines at ~2.5x speedup.
Problem
GUI agents need historical context for sequential navigation, but full history (especially visual observations) adds heavy computation and can distract the policy; prior RL GUI agents mostly drop history observations.
Approach
Dynamic Context Sampling uses an exponential-biased distribution over history lengths during rollouts, and Anchor-guided History Compression trains a compressed branch that drops history screenshots after an early layer while keeping history action tokens, coupled to the full-history branch via a KL alignment loss.
Key results
- HiconAgent-3B beats GUI-R1-7B by +8.46 grounding and +11.32 SR on GUI-Odyssey (52.10 and 50.11).
- Averaged over AndroidControl, AITW and GUI-Odyssey OOD step SR, HiconAgent-3B reaches 51.47 using 3K training samples, the highest among compared models.
- Layer-drop with k=6 reduces FLOPs by about 59.5% versus the uncompressed 3B model (25.21T vs 35.75T), a 2.47x computational speedup per the paper.
Surprising findings
- Inference-only history compression on the 2AO Qwen2.5VL-3B model drops SR by 4.95%, while HCPO training recovers this and exceeds the uncompressed baseline in SR.
- Naive uniform sampling of history lengths causes degeneration of short-history responses during training.
Limitations
- Compression configuration (drop layer k=6) and history window of 2 are fixed; longer histories not explored in main results.
- Evaluation on AndroidControl-High and AITW reports results primarily on the authors' setup; HiconAgent-3B does not clearly beat GUI-R1-7B on AITW.
Experimental setup
Datasets: AMEX AndroidControl AITW GUI-Odyssey
Benchmarks: AndroidControl-High AITW GUI-Odyssey
Metrics: step success rate grounding accuracy action type accuracy FLOPs tokens
Baselines: GUI-R1-3B GUI-R1-7B Qwen2.5VL-3B Qwen2.5VL-7B GPT-4o OS-Atlas-7B UI-shift-7B UI-AGILE infiGUI-3B
Builds on: GRPO GUI-R1 Qwen2.5-VL
Closest papers by meaning
- History-Aware Reasoning for GUI Agents (AAAI 2026 · similarity 0.89)
- GUI-Rise: Structured Reasoning and History Summarization for GUI Navigation (NeurIPS 2025 · similarity 0.89)
- GUI Exploration Lab: Enhancing Screen Navigation in Agents via Multi-Turn Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- SE-GUI: Enhancing Visual Grounding for GUI Agents via Self-Evolutionary Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- GUI-Shift: Enhancing VLM-Based GUI Agents through Self-supervised Reinforcement Learning (ICLR 2026 · similarity 0.85)
- Training High-Level Schedulers with Execution-Feedback Reinforcement Learning for Long-Horizon GUI Automation (CVPR 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.