More Capable, Less Cooperative? When LLMs Fail At Zero-Cost Collaboration

The project is a classic academic research repository (8 days old, 0 stars, 3 forks) accompanying a paper. Its value lies in the Zero-Cost Collaboration (ZCC) methodology and the counter-intuitive finding that higher capability does not equate to higher cooperativeness. From a competitive standpoint, this is a research artifact rather than a product. It has a low defensibility score (2) because the evaluation methodology is easily reproducible by any lab or evaluation startup. Frontier labs (OpenAI, Anthropic) face a 'high' risk profile here because they are the primary entities concerned with LLM alignment and agentic behavior; they are likely to internalize these specific failure modes into their internal safety and alignment evals (e.g., OpenAI Evals). The 'displacement horizon' is short (6 months) because the field of LLM benchmarking moves extremely quickly, and these specific scenarios will likely be absorbed into larger, more comprehensive multi-agent benchmark suites like Sotopia or AgentBench.