Memory Heuristics Fail in LLM Group Decision Making
October 26, 2026
The CrossProfile benchmark reveals that common memory heuristics for LLM groups—relevance and rarity—fail to preserve critical information for decision-makers. In testing across Qwen2.5, Llama-3.1, and DeepSeek, relevance-based filters often incorrectly include decoys, leading to significant performance loss.
HOW THIS AFFECTS YOU
●
builderBe cautious when using simple relevance filters to compress multi-agent conversation history for downstream tasks.
●
researcherThis highlights fundamental flaws in how LLMs summarize group discussions for context windows.