Research map / Multi-agent systems and games / LLM agents

History-Aware Reasoning for GUI Agents

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: GUI agent memory and trajectories

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 History-Aware Reasoning (HAR), a reflective RL framework that uses teacher-synthesized correction guidelines and a memory-augmented hybrid reward to give a 3B GUI agent (HAR-GUI-3B) short-term memory of episode history.

Problem

Native GUI agents' explicit System-2 reasoning ignores historical interactions in long-horizon episodes, treating each screen in isolation, which hurts GUI automation performance.

Approach

Build a reflection scenario where the agent sees its own past errors plus teacher-written correction guidelines, train it with GRPO using a hybrid reward (format, distance-scaled action reward, and a model-judged memory reward checking whether CoT references history), then align to inference format with a task-mixing RL stage that includes grounding.

Key results

• HAR-GUI-3B reaches 70.2 overall step success rate on AITW, above MP-GUI (69.2) and InfiGUI-R1-3B (67.7).
• On ScreenSpot-V2, HAR-GUI-3B scores 86.2 average, above UI-R1-3B (85.4) and OS-Atlas-7B (84.1).
• On GUI-Odyssey, HAR-GUI-3B reaches 62.31 overall SSR versus 58.39 for Qwen2.5-VL-7B.

Surprising findings

• Training with episodic reasoning RL alone degraded grounding ability on ScreenSpot, which the task-mixing strategy mitigated.
• Prompt-constrained focus on history during RL (GRPO*) degraded performance compared with unconstrained GRPO.

Limitations

• The memory reward relies on a large external judge model (Qwen3-235B-A22B), adding cost.
• The OOD benchmark is small (415 tasks) and restricted to CLICK-only episodes in one app ecosystem.
• Teacher-generated guidelines depend on Qwen2.5-VL-72B quality and were not validated beyond the paper's experiments.

Experimental setup

Datasets: MP-GUI data OS-Atlas AITW Mind2Web GUI-Odyssey

Benchmarks: AITW Mind2Web GUI-Odyssey ScreenSpot ScreenSpot-V2 ScreenQA Screen2Words WebSRC Widget Captioning Taperception Chinese OOD Alipay mini-program benchmark

Metrics: step success rate element accuracy operation F1 accuracy CIDEr SQuAD F1 ROUGE-L

Baselines: UI-R1-3B GUI-R1-3B InfiGUI-R1-3B MP-GUI ShowUI SeeClick OmniParser UI-TARS-2B OS-Atlas-7B Qwen2.5-VL-3B Qwen2.5-VL-7B UI-TARS-1.5-7B

Builds on: GRPO Qwen2.5-VL System-2 CoT distillation curriculum learning

Closest papers by meaning

• GUI-Rise: Structured Reasoning and History Summarization for GUI Navigation (NeurIPS 2025 · similarity 0.90)
• HiconAgent: History Context-aware Policy Optimization for GUI Agents (CVPR 2026 · similarity 0.89)
• GUI Exploration Lab: Enhancing Screen Navigation in Agents via Multi-Turn Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
• SE-GUI: Enhancing Visual Grounding for GUI Agents via Self-Evolutionary Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
• GUI-Eyes: Tool-Augmented Perception for Visual Grounding in GUI Agents (AAAI 2026 · similarity 0.87)
• GUI-Shift: Enhancing VLM-Based GUI Agents through Self-supervised Reinforcement Learning (ICLR 2026 · similarity 0.87)