Search papers, labs, and topics across Lattice.
2
0
4
Aggregate-score leaderboards can mislead, as they fail to predict agent performance in real-world scenarios, revealing a critical flaw in current evaluation practices.
Hallucinations in multi-agent LLM workflows are far more complex and frequent than previously measured, with nearly half involving multiple types simultaneously, demanding a shift from final-answer checks to trajectory-aware auditing.
