live-web-multimodal-information-seeking

# Skill: live-web-multimodal-information-seeking ## 1. Capability Definition & Real Case * **Professional Definition**: The capability to locate, access, and extract a short factual answer from live web content through browser-based GUI interaction, including multimodal sources such as images, videos, or interactive assets. The agent must identify reliable sources, follow a viable browsing trajectory, and avoid unsupported shortcut answers from search snippets or unrelated summaries. * **Dimension Hierarchy**: Goal-Directed GUI Workflow Execution->Live Information Interaction->live-web-multimodal-information-seeking ### Real Case **[Case 1]** * **Initial Environment**: A browser session starts on the open web without the answer preloaded. The task points to a specific website section containing an interactive comparison over a historical document, and the agent must navigate to the intended source, inspect the document closely, and read a fine-grained textual detail from the rendered content. * **Real Question**: The Blog page of AI2's olmOCR website contains a section titled 'Interactive Tool Comparison' that allows the user to compare OCR tools on a 'Historical Document'. What is the exact text written below the horizontal line in the right margin of this document? * **Real Trajectory**: Find the correct blog page, navigate to the specified section, activate the comparison content, open or zoom the historical document far enough to inspect the right margin, and read the requested text directly from the source asset. * **Why this demonstrates the capability**: The answer is not recoverable from generic search snippets because the crucial evidence lives inside the specified rendered document view. The agent must combine source selection, navigation, and fine-grained multimodal inspection to reach the answer. That exact pattern is what live-web multimodal information seeking is meant to measure. ## Pipeline Execution Instructions To synthesize data for this capability, you must strictly follow a 3-phase pipeline. **Do not hallucinate steps.** Read the corresponding reference file for each phase sequentially: 1. **Phase 1: Environment Exploration** Read the exploration guidelines to discover raw knowledge seeds: `references/EXPLORATION.md` 2. **Phase 2: Trajectory Selection** Once Phase 1 is complete, read the selection criteria to evaluate the trajectory: `references/SELECTION.md` 3. **Phase 3: Data Synthesis** Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data: `references/SYNTHESIS.md`