Research map / Large language models / Alignment and preferences
WEPO: Web Element Preference Optimization for LLM-based Web Navigation
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 WEPO, which applies DPO to LLM web navigation by sampling DOM-distance-based non-salient elements as negatives, and reports state-of-the-art step success rate on Mind2Web.
Problem
LLM-based web navigation agents must pick the correct HTML element and operation from many crowded candidates, and standard SFT uses only positive annotations, leaving the redundancy of HTML elements unexploited for contrastive training.
Approach
Unsupervised preference pairs are built from the page's own non-salient elements: negatives are sampled by LCA-based DOM distance from the target element (with operation replacement for TYPE/SELECT), and the DPO objective is applied to the pruned HTML, intent, and history prompt.
Key results
- Llama3-8B WEPO distance-based reaches 63.5% overall SSR on Mind2Web, versus 58.2% for CogAgent and 49.7% for WebAgent (reported as +5.3% and +13.8%).
- Distance-based sampling beats random sampling by 2.3% to 3% SSR across the three held-out test sets.
- Llama3-8B WEPO with k=10 pruning still outperforms Flan-T5-XL MindAct, and adding the ground-truth element raises SSR to over 80%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune Llama3-8B with WEPO distance-based preference learning rather than CogAgent | Mind2Web navigation · Mind2Web overall | SSR: 5.3% | Helped |
| fine-tune Llama3-8B with WEPO rather than WebAgent | Mind2Web navigation · Mind2Web overall | SSR: 13.8% | Helped |
| use distance-based rather than random negative-element sampling (ablation) | Llama3-8B WEPO · Mind2Web overall | SSR: 61.1 -> 63.5 | Helped |
| use distance-based rather than random negative-element sampling (ablation) | Mistral-7B WEPO · Mind2Web overall | SSR: 57.2 -> 59.5 | Helped |
| use distance-based rather than random negative-element sampling (ablation) | Gemma-2B WEPO · Mind2Web overall | SSR: 45.4 -> 48.4 | Helped |
| reduce HTML pruning ratio from k=50 to k=10 (ablation) | Llama3-8B WEPO · Mind2Web cross_domain | SSR: 64.4 -> 49.7 | Hurt |
| reduce HTML pruning ratio from k=50 to k=10 (ablation) | Llama3-8B WEPO · Mind2Web cross_task | SSR: 66.1 -> 55.0 | Hurt |
| reduce HTML pruning ratio from k=50 to k=10 (ablation) | Llama3-8B WEPO · Mind2Web cross_website | SSR: 60.0 -> 46.5 | Hurt |
| include the ground-truth element in the k=10 candidate pool (ablation) | Llama3-8B WEPO · Mind2Web cross_domain | SSR: 49.7 -> 87.2 | Helped |
| include the ground-truth element in the k=10 candidate pool (ablation) | Llama3-8B WEPO · Mind2Web cross_task | SSR: 55.0 -> 88.7 | Helped |
| include the ground-truth element in the k=10 candidate pool (ablation) | Llama3-8B WEPO · Mind2Web cross_website | SSR: 46.5 -> 85.4 | Helped |
| increase negative sample count from n=1 to n=3 with distance-based sampling (ablation) | Llama3-8B WEPO · Mind2Web average across three cross-test sets | SSR: 57.7% -> 63.5% | Helped |
Surprising findings
- Adding the ground-truth element to the candidate pool raises SSR above 80%, suggesting candidate retrieval is the main bottleneck rather than action prediction.
- Gemma-2B WEPO remains more than 10 points behind the larger models, showing a strong scaling effect.
Limitations
- Assumes a single correct element per step (no equivalent multiple paths), which may reduce diversity.
- Relies on a DeBERTa-based candidate generation module for HTML pruning, making candidate retrieval a bottleneck.
- Evaluated only on Mind2Web; generalization to other benchmarks not yet tested.
Experimental setup
Datasets: Mind2Web
Benchmarks: Mind2Web
Metrics: step success rate operation F1 element distance
Baselines: MindAct WebAgent CogAgent Flan-T5-XL MindAct Llama3-8B MindAct SFT
Builds on: Direct Preference Optimization MindAct LoRA
Closest papers by meaning
- Web-Shepherd: Advancing PRMs for Reinforcing Web Agents (NeurIPS 2025 · similarity 0.88)
- WebVLN: Vision-and-Language Navigation on Websites (AAAI 2024 · similarity 0.86)
- WebLINX: Real-World Website Navigation with Multi-Turn Dialogue (ICML 2024 · similarity 0.85)
- Less is More: Improving LLM Alignment via Preference Data Selection (NeurIPS 2025 · similarity 0.85)
- Multimodal Web Navigation with Instruction-Finetuned Foundation Models (ICLR 2024 · similarity 0.85)
- Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation (ICLR 2025 · 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.