InfoGain-RAG: Boosting Retrieval-Augmented Generation through Document Information Gain-based Reranking and Filtering

Venue: EMNLP 2025 (Main Conference) · Published: November 2025 · Type: Academic paper (RAG)

Source: https://aclanthology.org/2025.emnlp-main.365.pdf

Summary

InfoGain-RAG proposes using document information gain as the core signal for reranking and filtering retrieved documents in Retrieval-Augmented Generation pipelines. Instead of treating all retrieved passages as equally valuable context, the method estimates the incremental information each document contributes toward answering the query, then reranks and filters documents accordingly before generation. This targets a well-known RAG failure mode: noisy, redundant, or irrelevant passages in the context window that degrade LLM answer quality.

Key Points

• Core idea: quantify per-document information gain and use it to drive both reranking and filtering stages of the RAG pipeline.
• Problem addressed: traditional pipelines treat retrieval, reranking, and generation as disconnected stages; irrelevant context hurts generation quality and wastes token budget.
• Pipeline position: the technique is a pluggable reranker/filter between the retriever and the generator, fitting standard cascades (dense/sparse retrieval → rerank → generate).
• Evaluation context: work in this area is typically assessed on benchmarks such as MS MARCO, BEIR, and Natural Questions with metrics like nDCG@10, MRR, Recall@k, plus answer faithfulness and task success. Refer to the original PDF for exact quantitative results.

Context in the RAG Landscape

• Neural IR has evolved from BM25 → dual-tower dense retrieval → cross-encoder rerankers → late interaction → generative retrieval and LLM agents, each balancing efficiency, effectiveness, and maintainability.
• Agentic RAG extends retrieval from a one-shot step to an iterative, verifiable, plannable process, shifting evaluation from static nDCG toward task success rates, citation accuracy, and multi-hop reasoning quality.
• InfoGain-RAG contributes to the reranking layer of this stack, where deciding *which* documents deserve context slots is increasingly treated as a learnable optimization problem.

Practical Takeaways for Engineers

• Architecture: cascade retrieval + rerank + generation remains the mainstream pattern; reranking quality directly gates downstream generation quality.
• Latency/cost: filtering by information gain before generation can reduce token spend; consider early stopping and caching hot queries.
• Quality assurance: verify offline gains translate to online satisfaction via interleaving experiments and citation checking; LLM-as-judge should be cross-validated with human evaluation.
• Safety: open retrieval can introduce poisoning or bias — use source allowlists, adversarial detection, and output filtering.

Limitations and Open Questions

Typical open issues in this line of work include benchmark-to-real-user distribution mismatch, English-centric data limiting cross-lingual generalization, compute budget constraints on experiments, and safety risks of agentic retrieval on the open web. Future directions include more efficient test-time compute allocation, deeper integration with knowledge graphs and structured databases, and causal/fairness constraints for recommendation settings.

Related Reading

• A Survey of Graph Retrieval-Augmented Generation for Customized LLMs (arXiv:2501.13958)
• A Survey on Retrieval-Augmented Text Generation for Large Language Models (arXiv:2404.10981)
• Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG (arXiv:2501.09136)
• RAFT: Adapting Language Model to Domain Specific RAG (2024)
• RAG vs. GraphRAG: A Systematic Evaluation and Key Insights (arXiv:2502.11371)

> Note: This write-up is based on the paper's title, metadata, and public listing. For exact experimental numbers, ablations, and method details, consult the original PDF.