The Convention Gap: Towards Measuring Implicit Comm... | AI Research

The Convention Gap: Towards Measuring Implicit Communication in Cooperative AI Evaluation
Cooperative AI agents are often evaluated by how well they perform when paired with other AIs. However, human cooperation relies heavily on "implicit conventions"—shared, unspoken protocols that allow partners to understand meaning beyond what is literally communicated. This paper argues that current AI benchmarks may be measuring how well an AI works with other machines, rather than how well it can cooperate with humans. To address this, the authors introduce the "convention gap," a new metric that measures the difference between what a player should know based on literal information and how often they actually succeed in a task. The ai search story also surfaces in Google AI Releases TimesFM 3 for..., adding another angle.

Measuring the Unspoken

The researchers use the card game Hanabi as a testbed. In Hanabi, players hold cards facing outward, meaning they cannot see their own cards and must rely on limited, literal hints from their partners. Because the game’s rules and deck composition are finite and deterministic, the authors can mathematically calculate the exact probability of a player making a mistake based solely on the literal information they have received. The "convention gap" is the difference between this calculated failure probability and the actual observed failure rate. A positive gap indicates that players are succeeding more often than literal information alone would predict, suggesting they are successfully using implicit conventions to bridge the communication divide.

Key Findings Across Settings

By analyzing over 101,000 play actions, the study reveals a clear divide between humans and AIs:

• Human-Human Play: These pairs showed a large positive gap (+26.2 percentage points), proving that humans rely heavily on implicit communication to succeed.

• AI-AI Play: These pairs showed a gap near zero (-0.7 percentage points), indicating that they largely rely on literal information and lack the shared conventions found in human play.

• Human-AI Play: These pairs fell in the middle (+16.4 percentage points). Interestingly, the study found that the AI partners that elicited the largest "convention gap" from their human partners actually resulted in the fewest human failures. The same ai evaluation question is explored in Efficient Test-Time Adaptation through Human-AI Interaction, which adds a research perspective.

Why This Matters for AI Development

The research suggests that traditional AI-AI performance scores are not reliable predictors of how well an AI will work with a human. While game scores depend on the specific composition of the players involved, the convention gap is a more stable metric that separates human-like cooperation from machine-like performance at the agent level. By using this metric, developers can better identify which AI agents are truly "convention-compatible," potentially leading to more effective human-AI teams without needing to retrain the underlying models.

Validation Through Control

To ensure the metric is accurate, the authors tested it against "Off-Belief Learning" agents. These are specialized AIs designed with varying levels of convention-based communication. The convention gap successfully tracked these levels, rising from a near-zero gap for convention-free agents to a high gap for those designed to use complex conventions. This confirms that the metric is effectively isolating the intended signal—the ability to communicate beyond literal instructions. The ai search story also surfaces in OpenAI Says AI Found Possible Navier–Stokes..., adding another angle. as detailed in the full paper on Arxiv