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

Learning to Adapt: Self-Improving Web Agent via Cognitive-Aware Exploration

CVPR 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 0 citations (OpenAlex)

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 SCALE, a self-driven web-agent exploration framework where one MLLM plays Selector, Predictor, and Judger roles to find its cognitive boundaries, plus a graph-based SCALE-Hop exploration strategy and the SCALE-20k dataset.

Problem

Existing MLLM web agents rely on handcrafted execution pipelines or costly expert trajectories and struggle to adapt to dynamic, diverse real-world websites; they also do not systematically probe their own cognitive boundaries.

Approach

An MLLM assumes three adversarial roles (Selector proposes challenging actions, Predictor forecasts outcomes, Judger checks predictions against observed results); mismatched predictions are turned into SFT data to update Selector and Predictor iteratively, with SCALE-Hop adding a graph of visited states and verification-guided backtracking for global exploration.

Key results

• SCALE improves InternVL2.5-8B by 231.8% and Qwen2.5-VL-7B by 176.3% on average task success rate over zero-shot.
• On WebVoyager, SCALE gets 7.9 SR for Qwen2.5-VL-7B versus 0.6 zero-shot and 6.7 for OS-Genesis.
• SCALE-Hop raises visited node coverage (overall 876 vs 277 without SCALE-Hop), reported as a 216% increase.

Surprising findings

• Random-walk exploration achieves higher shopping SR (14.8) than SCALE for some settings, and w/o SCALE-Hop matches or beats SCALE on Reddit (7.1 vs 4.8).

Limitations

• Relies on GPT-4o for dataset construction and evaluation, and on the same MLLM for all three roles, so the judger is not independently trained.
• Results on WebVoyager are low in absolute terms (success rates under 10%).
• Reported improvement percentages are relative to very low zero-shot baselines.

Experimental setup

Datasets: SCALE exploration traces on 19 real-world websites VisualWebArena WebVoyager

Benchmarks: VisualWebArena WebVoyager

Metrics: success rate average steps node coverage Sentence-BERT cosine diversity

Baselines: GPT-4o zero-shot Augvis ViGoRL GPT trajectory imitation OS-Genesis Tree Search LLaVA-NeXT-8B zero-shot Random Walk

Builds on: Set-of-Mark OS-Genesis ReAct Tree-of-Thought

Closest papers by meaning

• WebPilot: A Versatile and Autonomous Multi-Agent System for Web Task Execution with Strategic Exploration (AAAI 2025 · similarity 0.88)
• Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation (ICLR 2025 · similarity 0.88)
• WebGym: Scaling Training Environments for Long-Horizon Visual Web Agents with Realistic Tasks (CVPR 2026 · similarity 0.88)
• ScaleEnv: Scaling Environment Synthesis from Scratch for Generalist Interactive Tool-Use Agent Training (ICML 2026 · similarity 0.87)
• Go-Browse: Training Web Agents with Structured Exploration (ICLR 2026 · similarity 0.87)
• Safe and Scalable Web Agent Learning via Recreated Websites (ICML 2026 · similarity 0.87)