GPTNT

Properties

axis: multi-agent
verification: det-test
whatItTests: Real-time, asynchronous collaboration benchmark built on the cooperative video game Keep Talking and Nobody Explodes: one agent sees the bomb but not the manual, the other holds the manual but cannot see the bomb, and both must communicate live under a countdown to defuse procedurally generated puzzles.
discriminates: [spawn-and-isolate]
saturation: open
patternLabFit: The strict information asymmetry (bomb-holder vs instruction-holder) plus real-time, no-turn-taking constraint is a sharp discriminator for Pattern Lab's spawn-and-isolate pattern and for measuring roundsToFirstPass under genuine time pressure rather than simulated turn order; the benchmark's design lets a runner selectively withhold the manual, the partner agent, or both to isolate learned vs in-the-moment reasoning.
source: GPTNT: Benchmarking Real-Time Collaboration Between Multimodal Agents on Keep Talking And Nobody Explodes (arXiv:2606.28514)
notes: Distinct from tracked entries: GPTNT is real-time (no turn-taking) and built on an actual commercial cooperative video game with genuine information asymmetry between exactly two agents, a mechanism none of MAST/MultiAgentBench/BattleAgentBench test. Very recent (June 2026); moderate confidence pending wider citation uptake.

axis: multi-agent
verification: det-test
whatItTests: Real-time, asynchronous collaboration benchmark built on the cooperative video game Keep Talking and Nobody Explodes: one agent sees the bomb but not the manual, the other holds the manual but cannot see the bomb, and both must communicate live under a countdown to defuse procedurally generated puzzles.
discriminates: [spawn-and-isolate]
saturation: open
patternLabFit: The strict information asymmetry (bomb-holder vs instruction-holder) plus real-time, no-turn-taking constraint is a sharp discriminator for Pattern Lab's spawn-and-isolate pattern and for measuring roundsToFirstPass under genuine time pressure rather than simulated turn order; the benchmark's design lets a runner selectively withhold the manual, the partner agent, or both to isolate learned vs in-the-moment reasoning.
source: GPTNT: Benchmarking Real-Time Collaboration Between Multimodal Agents on Keep Talking And Nobody Explodes (arXiv:2606.28514)
notes: Distinct from tracked entries: GPTNT is real-time (no turn-taking) and built on an actual commercial cooperative video game with genuine information asymmetry between exactly two agents, a mechanism none of MAST/MultiAgentBench/BattleAgentBench test. Very recent (June 2026); moderate confidence pending wider citation uptake.