Research Finds Input Formatting Matters More Than Model Size, Boosting LLM Memory Accuracy by Up to 72 Points
Researchers unveil RENDER, showing that how data is formatted before being fed into an LLM affects memory accuracy more than model size
Evaluations of memory and retrieval systems, or RAG (Retrieval-Augmented Generation), for large language models (LLMs) often overlook how input data is presented. Systems may convert conversation histories into memory notes, summaries, structured records, or raw text. Researchers from the University of Waterloo, Stanford University, and independent contributors have now introduced RENDER, a benchmark control that keeps conversations fixed while varying the reader-facing artifact shown to the model.
The study combines a five-level packet ladder to identify where answer-bearing content appears in the input with deterministic templates that simulate real-world formats, including ChatGPT-style conversation lists, LangChain summaries, MemGPT-style typed records, and raw dialogue. Tests across 500 LongMemEval questions and nine models found that matched-budget resolved packets outperformed recency-truncated raw dialogue by 42.4 to 72.6 points. Production-style templates also showed gaps of 24.6 to 48.8 points between the highest- and lowest-scoring formats, underscoring that data presentation can materially change system performance.
The findings show AI developers and users in Thailand that optimizing input formatting can be as important as choosing a larger LLM, helping reduce costs while improving RAG accuracy.