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

Web-CogReasoner: Towards Knowledge-Induced Cognitive Reasoning for Web Agents

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning

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 Web-CogKnowledge, a Bloom's-taxonomy-inspired framework with a curated Web-CogDataset (factual/conceptual/procedural tasks from 14 websites) and a knowledge-driven CoT, used to fine-tune Qwen2.5-VL-7B into the Web-CogReasoner web agent, evaluated on the new Web-CogBench plus VisualWebBench, WebVoyager, and Online Mind2Web.

Problem

Web agents built on general-purpose multimodal models lack specialized web knowledge and structured cognitive reasoning, limiting performance on knowledge-intensive and generalization web tasks.

Approach

Organize web training data by Bloom's taxonomy into factual, conceptual, and procedural knowledge stages applied as a curriculum via SFT, and pair it with a knowledge-driven chain-of-thought that grounds each reasoning layer in one knowledge type.

Key results

• Web-CogReasoner reaches 84.4% overall on Web-CogBench vs 80.2% for Gemini 2.5 Pro and 69.8% for base Qwen2.5-VL-7B.
• On WebVoyager, the full model with KCoT reaches 42.9% success vs 25.35% without KCoT in the ablation; overall table reports 30.2% against OpenWebVoyager Max at 26.2%.
• On VisualWebBench, overall average is 86.3%, slightly above UI-TARS-7B-SFT at 86.0%.

Surprising findings

• UI-TARS-7B-SFT scores 86.0% on VisualWebBench but only 46.4% on Web-CogBench, suggesting perception alone does not ensure cognitive reasoning.
• Single-stage training on procedural data (S3 only) underperforms the base model on overall Web-CogBench (60.66 vs 69.81).

Limitations

• Relies on imitation learning only; reinforcement learning is left to future work.
• Web-CogBench is small (876 instances) and uses LVM-judged metrics with a limited inter-rater analysis.
• Training data partly generated by LLMs (Qwen2.5-VL-72B, GPT-4o) and online evaluation covers few websites; WebVoyager results are mixed with some categories lower than baselines.

Experimental setup

Datasets: Web-CogDataset MultiUI Mind2Web OpenWebVoyager MultiModal-Mind2Web VisualWebBench WebVoyager Online-Mind2Web

Benchmarks: Web-CogBench VisualWebBench WebVoyager Online-Mind2Web

Metrics: accuracy ROUGE-L LVM judge score success rate average steps

Baselines: Claude Sonnet 4 Gemini 2.5 Pro Qwen2.5-VL-7B UI-TARS-7B-SFT OpenWebVoyager

Builds on: Bloom's Taxonomy Qwen2.5-VL OpenWebVoyager Mind2Web

Closest papers by meaning

• WebFactory: Automated Compression of Foundational Language Intelligence into Grounded Web Agents (ICLR 2026 · similarity 0.87)
• $\tau$-Knowledge: Evaluating Conversational Agents over Unstructured Knowledge (ICML 2026 · similarity 0.86)
• WebArena: A Realistic Web Environment for Building Autonomous Agents (ICLR 2024 · similarity 0.86)
• RealWebAssist: A Benchmark for Long-Horizon Web Assistance with Real-World Users (AAAI 2026 · similarity 0.86)
• Mind2Web 2: Evaluating Agentic Search with Agent-as-a-Judge (NeurIPS 2025 · similarity 0.86)
• REAL: Benchmarking Autonomous Agents on Deterministic Simulations of Real Websites (NeurIPS 2025 · similarity 0.85)