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

WebGen-Agent: Enhancing Interactive Website Generation with Multi-Level Feedback and Step-Level Reinforcement Learning

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution

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 WebGen-Agent, a website-generation code agent that uses VLM screenshot scores and GUI-agent testing scores for multi-level feedback, backtracking and best-step selection, and Step-GRPO that uses those scores as per-step rewards to train small coding LLMs.

Problem

Code agents for website generation rely mainly on execution feedback, which misses visual quality and interactive functionality; small open-source coding LLMs also lag far behind proprietary models on this task.

Approach

Use a VLM's screenshot appearance score and a GUI-agent functional test score as both iterative feedback inside the agent loop (with backtracking and best-step selection) and as dense step-level rewards in GRPO (Step-GRPO), without a cumulative future-reward advantage.

Key results

• WebGen-Agent raises Claude-3.5-Sonnet accuracy on WebGen-Bench from 26.4% (Bolt.diy) to 51.9% and appearance score from 3.0 to 3.9.
• Step-GRPO raises Qwen2.5-Coder-7B-Instruct accuracy from 38.9% (SFT) to 45.4% and appearance score from 3.4 to 3.7.
• Qwen3-Coder-480B-A35B-Instruct within WebGen-Agent reaches the highest accuracy of 58.2% and appearance score of 4.3.

Surprising findings

• Replacing the feedback VLM GPT-4o with the open-source Qwen2.5-VL-32B gives no notable gain, while replacing the coding LLM DeepSeek-V3 with Qwen2.5-VL-32B drops accuracy from 52.6% to 5.6%.
• Adding GUI-agent testing feedback slightly lowers the appearance score (3.6 to 3.4) before backtracking and select-best are added.

Limitations

• Only 7B and 8B models were trained with SFT and Step-GRPO due to limited compute.
• Does not consider website response speed or network conditions.
• Scores from the VLM feedback are only about 89-96% accurate by human annotation.

Experimental setup

Datasets: WebGen-Bench WebGen-Instruct

Benchmarks: WebGen-Bench

Metrics: accuracy appearance score Yes/Partial/No/Start Failed rates

Baselines: OpenHands Aider Bolt.diy WebGen-LM-7B WebGen-LM-14B WebGen-LM-32B Naive outcome GRPO Step-GRPO with cumulative advantage

Builds on: GRPO Bolt.diy WebGen-Bench DeepSeekMath GRPO

Closest papers by meaning

• WebGen-Bench: Evaluating LLMs on Generating Interactive and Functional Websites from Scratch (NeurIPS 2025 · similarity 0.90)
• Vision2Web: A Hierarchical Benchmark for Visual Website Development with Agent Verification (ICML 2026 · similarity 0.90)
• FullStack-Agent: Enhancing Agentic Full-Stack Web Coding via Development-Oriented Testing and Repository Back-Translation (ICML 2026 · similarity 0.89)
• WebGym: Scaling Training Environments for Long-Horizon Visual Web Agents with Realistic Tasks (CVPR 2026 · similarity 0.89)
• A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis (ICLR 2024 · similarity 0.87)
• WebFactory: Automated Compression of Foundational Language Intelligence into Grounded Web Agents (ICLR 2026 · similarity 0.87)