TL;DR

LLM×MapReduce is a training-free divide-and-conquer framework that processes arbitrarily long documents by splitting them into chunks, extracting structured intermediate answers with confidence scores, and resolving inter-chunk conflicts at aggregation time.

Summary

Large language models struggle with documents that exceed their effective context window, forcing a choice between truncation and expensive retraining. LLM×MapReduce splits long texts into chunks processed in a Map stage (extracting structured information with a confidence score per chunk), optionally compresses results through a Collapse stage, then resolves contradictions in a Reduce stage via in-context confidence calibration. On InfiniteBench (>100K token inputs), Llama3-70B with LLM×MapReduce scores 68.66% average — outperforming GPT-4 (57.34%) and all open-source long-context baselines — while handling sequences up to 1.28M tokens with lower GPU memory overhead than standard decoding.

Key contributions

  1. Proposes LLM×MapReduce, a training-free three-stage (Map → Collapse → Reduce) pipeline enabling any LLM to process documents that exceed its native context limit without fine-tuning.
  2. Introduces a structured information protocol — a four-field output (Extracted Information, Rationale, Answer, Confidence Score) — that preserves inter-chunk dependencies across the pipeline stages.
  3. Demonstrates in-context confidence calibration, aligning chunk-level confidence scores to a consistent standard using prompt-embedded examples, enabling principled conflict resolution without parameter updates.
  4. Shows scalability to 1.28M-token sequences on a model with only 8K native context, achieving near-perfect retrieval (99.56%) and strong mathematical reasoning (91.43%) on InfiniteBench.

When to cite

  1. When arguing that training-free methods can match or exceed fine-tuned long-context LLMs on document understanding benchmarks.
  2. When explaining how to resolve contradictory evidence across chunks in a divide-and-conquer long-document pipeline.
  3. When discussing inference efficiency strategies for very long documents that avoid repeated full-context decoding and reduce GPU memory requirements.
  4. When citing work on structured prompting formats that preserve reasoning chains across independently processed text segments.

LLM Prompt Engineering A Survey of Large Language Models REACT - SYNERGIZING REASONING AND ACTING IN LANGUAGE MODELS