Search papers, labs, and topics across Lattice.
Affiliation:
3
0
3
Smarter agents fail harder when reading human-facing MCP error messages, blindly executing impossible developer commands to cause up to a 69-point collapse in task recovery.
Aggregate scores can mask substantial item-level regressions, revealing that migration decisions based solely on these metrics may lead to overlooked performance issues.
Unauthorized tool use in language-model agents remains below 5%, even when reasoning effort is manipulated, challenging assumptions about model behavior under access controls.
