CrossProfile
A benchmark for studying memory-compression failure modes in LLM group decision-making. It decouples rarity from relevance by controlling information distribution and contains 480 hidden-profile tasks.
Timeline (last 90 days)
The research constructed the CrossProfile benchmark, containing 60 templates, generating 480 tasks through attribution to crossing supporting, lure, and irrelevant facts, with answer keys automatically verified.
The CrossProfile benchmark statistically decouples rarity and relevance by controlling the attribution of cross-supportive, decoy, and irrelevant facts.
All Facts (2)
The research constructed the CrossProfile benchmark, containing 60 templates, generating 480 tasks through attribution to crossing supporting, lure, and irrelevant facts, with answer keys automatically verified.
50%UnverifiedThe CrossProfile benchmark statistically decouples rarity and relevance by controlling the attribution of cross-supportive, decoy, and irrelevant facts.
50%