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

Prune4Web: DOM Tree Pruning Programming for Web Agent

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Software engineering agent benchmarks · 1 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.

Prune4Web prunes massive DOM trees by having an LLM generate keyword-weight parameters for a fixed Python scoring template, then uses a planner, filter, and grounder (SFT + GRPO on Qwen2.5-VL-3B) to reach state-of-the-art web grounding on Multimodal-Mind2Web.

Problem

LLM-based web agents struggle with massive DOM trees (10k-100k tokens), where truncation loses information and full-DOM processing causes attention dilution and latency; existing pruning relies on rigid heuristics or separate ranking LLMs.

Approach

DOM Tree Pruning Programming: the LLM outputs only keywords and base weights for a fixed, hand-designed Python scoring template that ranks DOM elements externally, shortlisting Top-N candidates for a downstream grounder; a two-turn dialogue unified model is trained with SFT and GRPO using filter-retention rewards for the planner.

Key results

• On the low-level sub-task grounding benchmark, Prune4Web raises grounding accuracy from 46.8% (no pruning, FT Qwen2.5-VL-3B) to 88.28%, above GPT-4o with the same filter (80.65%).
• On Multimodal-Mind2Web cross-task, Prune4Web-3B (Two-turn Dialogue Unified) reaches 58.4 element accuracy and 52.4 step SR, above MindAct Flan-T5XL (55.1 / 52.0).
• RFT raises Step SR from 46.5% to 52.4% for the unified two-turn model on the cross-task subset.

Surprising findings

• A fine-tuned Qwen2.5-0.5B filter/grounder performs nearly identically to the 3B model (97.64 vs 97.46 Recall@20; 88.28 grounding accuracy for both).
• Plain LLM Top-N selection yields 0.0% task completion for fine-tuned Qwen2.5-VL-3B, while programmatic filtering gives 5.2%.

Limitations

• Evaluated on a small number of online tasks (30) and relies on GPT-4o as judge for online success.
• Dataset is derived solely from Multimodal-Mind2Web and the grounding benchmark is built from GT sub-tasks, so end-to-end gains are modest in absolute terms.

Experimental setup

Datasets: Multimodal-Mind2Web

Benchmarks: Multimodal-Mind2Web cross-task Multimodal-Mind2Web cross-website Multimodal-Mind2Web cross-domain Low-level sub-task grounding benchmark (custom) 30 online tasks

Metrics: Element Accuracy Operation F1 Step Success Rate Recall@20 Grounding Accuracy LLM-Verified Task Completion Rate

Baselines: GPT-4 GPT-4o SeeAct SeeClick-9.6B MiniCPM-3.1B ScribeAgent-32B GPT-4o UGround EDGE-9.6B MindAct Flan-T5XL

Builds on: GRPO SeeAct Multimodal-Mind2Web LLaMA-Factory

Closest papers by meaning

• A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis (ICLR 2024 · similarity 0.89)
• Web-Shepherd: Advancing PRMs for Reinforcing Web Agents (NeurIPS 2025 · similarity 0.88)
• WALT: Web Agents that Learn Tools (ICLR 2026 · similarity 0.87)
• How to Train Your LLM Web Agent: A Statistical Diagnosis (NeurIPS 2025 · similarity 0.87)
• Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.86)
• AgentOccam: A Simple Yet Strong Baseline for LLM-Based Web Agents (ICLR 2025 · similarity 0.86)