Resin improves multi-step reasoning by organizing evidence with entity indexing
RESIN: Multi-Chain Reasoning with Entity-Structured Indexing
Computation and Language
Summary
Multi-step reasoning is hard because the needed facts are often scattered and related in complex ways. The authors designed RESIN, a method that breaks down questions into smaller parts, finds relevant facts for each part, and keeps multiple possible supports ready for later stages. This way, it avoids throwing away useful information too early and helps to put all the pieces together more clearly. RESIN performs better than alternatives on a challenging reasoning benchmark, showing that careful organization of evidence and waiting to choose the final reasoning path works well.
What this means in practice
- •For natural language processing engineers: Improve multi-step question answering systems by retrieving and organizing reasoning evidence in focused, entity-based sets.
- •For knowledge graph developers: Build tools that integrate entity-structured indexes to support complex multi-hop reasoning across dispersed data sources.
Authors
Xuanyu Meng, Jash Rajesh Parekh, Jiawei Han
Abstract
Multi-chain reasoning is a challenging task not only because the multi-chain evidence is dispersed, but also because different parts of reasoning may require distinct, interacting chains of support. Flat retrieval can flood a model with loosely related text, whereas early selection of a single reasoning path can discard evidence that becomes useful only after other facts are considered. We introduce RESIN (Reasoning with Entity Structured Indexing), which uses structure to locate evidence while keeping the original source as the basis for reasoning. RESIN decomposes a question into focused one-hop needs, retrieves a small set of source-grounded units for each hop, and places only this relevant evidence in a compact reasoning buffer. The buffer preserves alternative supports until final composition, making intermediate decisions explicit without forcing a brittle path. On the full HolisQA benchmark, RESIN achieves 82.92 context recall and 76.70 answer relevancy over all questions, with higher reported CR and AR across five domains and higher F1 in four domains than the published baselines. Ablations show that both typed evidence diversity and hop-wise decomposition are essential, supporting a clear conclusion: Precise evidence organization and delayed commitment make multi-chain reasoning effective and auditable. Our code is available at https://github.com/RamonMeng/RESIN.