Research2h ago
RENDER Benchmark Tests LLM Memory Formatting
Matched-budget resolved packets beat recency-truncated raw dialogue by up to 72.6 points in the new RENDER benchmark. The study tests how different input formats, like LangChain summaries or raw excerpts, affect model retrieval. Results show that how memory is rendered drastically alters accuracy. Practitioners should optimize input templates to avoid significant performance drops.