BattleAgentBench

Properties

axis: multi-agent
verification: mixed
whatItTests: Fine-grained evaluation across seven sub-stages of three difficulty levels, testing single-agent navigation, paired-agent task execution, and multi-agent collaboration AND competition capability, evaluated on four closed-source and seven open-source models.
saturation: open
patternLabFit: The staged single-to-paired-to-multi-agent difficulty progression is a candidate structure for graduated Pattern Lab collaboration tests that isolate where cooperation breaks down as agent count and adversarial pressure increase.
source: BattleAgentBench: A Benchmark for Evaluating Cooperation and Competition Capabilities of Language Models in Multi-Agent Systems (arXiv:2408.15971)
notes: Distinct from MultiAgentBench: BattleAgentBench uses a fixed seven-sub-stage, three-difficulty-level progression (single agent -> paired agent -> multi-agent) with explicit competitive scenarios, rather than milestone KPIs across coordination topologies. discriminates left empty: abstract does not detail a specific collaboration-pattern mechanism (e.g. debate, reputation-elo) beyond general cooperation/competition capability.

axis: multi-agent
verification: mixed
whatItTests: Fine-grained evaluation across seven sub-stages of three difficulty levels, testing single-agent navigation, paired-agent task execution, and multi-agent collaboration AND competition capability, evaluated on four closed-source and seven open-source models.
saturation: open
patternLabFit: The staged single-to-paired-to-multi-agent difficulty progression is a candidate structure for graduated Pattern Lab collaboration tests that isolate where cooperation breaks down as agent count and adversarial pressure increase.
source: BattleAgentBench: A Benchmark for Evaluating Cooperation and Competition Capabilities of Language Models in Multi-Agent Systems (arXiv:2408.15971)
notes: Distinct from MultiAgentBench: BattleAgentBench uses a fixed seven-sub-stage, three-difficulty-level progression (single agent -> paired agent -> multi-agent) with explicit competitive scenarios, rather than milestone KPIs across coordination topologies. discriminates left empty: abstract does not detail a specific collaboration-pattern mechanism (e.g. debate, reputation-elo) beyond general cooperation/competition capability.