C2c Eval

---
name: c2c-eval
description: Evaluates the strategic coordination and negotiation capabilities of language model agents in a long-horizon, mixed-motive multi-agent environment with asymmetric objectives and private information. Use when the user wants to benchmark on C2C, or asks about evaluating this task. Reports win rate.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.25088
  bibtex_key: [REDACTED]
  confidence: high
---

# c2c-eval

> Cooperate to Compete: Strategic Coordination in Multi-Agent Conquest — O'Neill et al. (2026) (arXiv:2604.25088, 2026)

## What this evaluates

Evaluates the strategic coordination and negotiation capabilities of language model agents in a long-horizon, mixed-motive multi-agent environment with asymmetric objectives and private information.

## Datasets

- **C2C** — total 162; splits: test (162); repo https://github.com/abbykoneill/negotiationgames

## Metrics

- `win rate` **(primary)** — range: percent
  - Percentage of games won by an agent or team across a set of starting positions. Calculated as (number of wins / total games played) * 100.

## Input / output format

**Input**: Game rules, starting positions, and real-time game state provided via a web-based interface for humans or a prompt-driven agentic framework for LMs.

**Output**: Negotiation proposals, alliance offers, and in-game actions within the C2C environment.

## Scoring recipe

```python
def compute_win_rate(outcomes):
    wins = sum(1 for o in outcomes if o == 'win')
    total = len(outcomes)
    return (wins / total) * 100 if total > 0 else 0.0
```

## Common pitfalls

- Participants were blinded to the backbone LMs of their opponents to prevent bias.
- AI-only games require random assignment of models to starting positions to average out position-specific advantages.
- Non-binding agreements mean negotiation reliability does not directly correlate with final game outcomes.

## Evidence (verbatim from paper)

> Prompt-based interventions targeting negotiation aggression, alliance-seeking, and deception improve AI win rates from 22.2% to 32.7%, revealing that coordination flexibility and behavioral alignment are critical for real-world deployment in complex social settings.

## Citation

```bibtex
@misc{oneill2026cooperate,
  title={Cooperate to Compete: Strategic Coordination in Multi-Agent Conquest},
  author={O'Neill et al. (2026)},
  year={2026},
  note={arXiv:2604.25088}
}
```

- arXiv: 2604.25088