long-form-coherent-summarization

# Skill: long-form coherent summarization ## 1. Capability Definition & Real Case * **Professional Definition**: The ability to summarize very long documents while preserving global coherence across chunk boundaries, maintaining a readable narrative structure rather than a bag of disconnected local summaries. The capability is primarily about long-range continuity: who is being discussed, what happened, why it matters, and how later parts relate to earlier ones after chunk-and-compress processing. * **Dimension Hierarchy**: Document Transformation & Synthesis->Summarization->long-form coherent summarization ### Real Case **[Case 1]** * **Initial Environment**: A recently published book far longer than the model context window is split into chunked inputs, and the system must produce one global summary from the entire book. * **Real Question**: "Summarize the following book-length document." * **Real Trajectory**: The system first generates chunk-level summaries and then either hierarchically merges them or incrementally updates a running summary while compressing when necessary. * **Why this demonstrates the capability**: The benchmark explicitly studies what happens when long summarization requires chunking, merging, updating, and compression. The central quality target is coherence, not just lexical overlap with a reference. This makes it the right case for testing long-range continuity under context pressure. --- **[Case 2]** * **Initial Environment**: A long government report, patent, or transcript is supplied as a single-document summarization task in a long-context evaluation suite. * **Real Question**: "Write a summary of the document that preserves the important information across the whole context." * **Why this demonstrates the capability**: These tasks test whether a model can aggregate information spread across many pages without collapsing into omissions, contradictions, or duplicated points. Unlike short-article summarization, the model must keep section-to-section continuity and relative salience under long input lengths. This demonstrates practical long-document summarization beyond book fiction. ## 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`