Query-Focused Summarization

Unlike generic summarization, QFS dynamically re-weights the importance of information based on the query. A sentence about battery life becomes highly salient only when the query asks about it. This is achieved through attention mechanisms that calculate semantic similarity between the query embedding and each passage in the source document, creating a query-biased summary.

A critical constraint in QFS is factual consistency, ensuring the generated summary contains only statements directly supported by the source text. This is often enforced via source grounding techniques, where the model must anchor claims to specific spans. Metrics like Attribution Fidelity measure how accurately the summary links back to the original evidence, minimizing hallucination risk.

QFS can be performed using two primary methods:

• Extractive QFS: Selects and concatenates the most relevant existing sentences from the source document based on query overlap.
• Abstractive QFS: Generates entirely new, rephrased sentences that capture the core answer. Modern transformer models often use a hybrid approach, copying key entities while paraphrasing context for fluency.

To prevent repetitive outputs, QFS systems employ Maximum Marginal Relevance (MMR) and Diversity Constraints. MMR selects passages by balancing high relevance to the query against low similarity to already-selected content. This ensures the final summary covers multiple distinct facets of the answer without repeating the same information.

QFS is highly sensitive to the Lost in the Middle phenomenon, where models ignore critical information placed in the center of a long context. To mitigate this, source documents are often re-ordered to place the most query-relevant chunks at the beginning and end of the prompt, exploiting the primacy and recency biases of large language models.

Effective QFS requires strict token budget allocation. The system must strategically distribute the limited output tokens to cover the most salient query-relevant aspects. Techniques like Chain-of-Density (CoD) iteratively refine a summary to pack more entities and details into a fixed length without sacrificing clarity, maximizing information density per token.