Inference and communication in the game of Password

Yang Xu, Charles KempNeurIPS 2010

This paper, Inference and communication in the game of Password by Yang Xu and Charles Kemp, explores the hidden mechanics of human cooperation through the lens of a classic TV game show. The researchers use computational modeling to demonstrate that efficient communication isn't just about being clear—it is a sophisticated, two-way dance of mental modeling where both parties actively work to minimize their collective "cognitive load."

By analyzing data from the game show Password, the authors found that humans communicate efficiently by actively modeling each other's inference processes. We don't simply pick the "easiest" words to say; we balance our choices to share the mental effort with our partner. Interestingly, this cooperative strategy only breaks down when we are forced to operate under severe time pressure.

The Communication Bottleneck

Communication is inherently a bottleneck. As the paper notes, human thought is fast, but phonetic articulation (speaking) is slow. To be efficient, we leave "inferential gaps"—we say less and expect the listener to fill in the blanks.

However, this creates a complex "game-theoretic regress." If I am the speaker, I have to think: "If I say X, will they infer Y?" Meanwhile, the hearer thinks: "Did they say X because they thought I’d infer Y?" This paper asks how humans actually navigate this complexity in real time. Do we behave selfishly, minimizing our own effort, or do we coordinate our strategies to make the search for meaning easier for both parties?

Modeling the Game of Password

The authors model the game of Password, where a speaker gives a one-word clue to help a hearer guess a secret password. They focus heavily on the asymmetry of word associations.

Consider the "Shovel-Snow" trap: A "selfish" speaker might pick "Snow" as a clue for "Shovel" because it is the first word that comes to mind (a strong forward association). But a "considerate" speaker realizes that "Snow" is a poor clue because the hearer will likely guess "Flake" or "Cold" before they ever guess "Shovel" (a weak backward association).

Word association strengths for clues and guesses in the game of Password.

To formalize this, the researchers developed three hypotheses:

  1. Considerate: Both parties take the other's perspective into account (e.g., the speaker avoids clues with weak backward associations).
  2. Calibrated: Both parties make accurate assumptions about the other’s strategy.
  3. Collaborative: Both parties share the cognitive burden equally, adhering to a "Principle of Least Collaborative Effort."

They used a mixture model to represent the probability of a clue or guess as a weighted combination of forward and backward association strengths. By fitting these weights to real-world game data, they could measure exactly how much "perspective-taking" was happening.

Datasets

The study relied on two primary data sources:

  • Password Game Data: 100 "normal" rounds and 100 "lightning" rounds (time-pressured) sampled from the 1962–1967 TV show hosted by Allen Ludden.
  • USF Word Association Database: A massive lexicon of 5,016 words and their association strengths, generated by over 6,000 participants.

Key Findings

  • Symmetry in Effort (Collaboration): In normal play, speakers and hearers weight forward and backward associations almost equally (roughly 50/50). This supports the idea that the cognitive burden is shared equally; we are naturally wired to meet our interlocutors halfway.
  • High Calibration: In normal rounds, the mixture weights for speakers and hearers were remarkably similar, suggesting that players make highly accurate assumptions about their partners' strategies.
  • The Fragility of Cooperation: When the "lightning round" starts, the cooperative system breaks down. Speakers, under time pressure, stop being "considerate" and revert to "greedy" forward associations—picking the easiest word for them. Because the hearer doesn't always adjust to this sudden shift in strategy, communication fails much more frequently.
  • The Cost of Consideration: Speakers took significantly longer to choose clues in normal rounds than in lightning rounds (approximately 4 seconds vs. 1 second). This suggests that "considerate" communication is not an automatic reflex; it requires active, time-consuming computation.

Limitations and Future Directions

The authors acknowledge that their mixture models are relatively simple and may not fully capture the step-by-step psychological processing of the contestants. They suggest that future work could explore more complex "bounded rationality" models. For instance, a speaker might mentally sample a few words with high forward strength and then evaluate that small subset to pick the one with the best backward strength.

Additionally, the current study assumes a "uniform prior"—meaning it assumes all passwords are equally likely. Real-world communication involves much richer context and varying expectations, which could be integrated into future iterations of the model to better reflect everyday human interaction.

Summary updated

BibTeX
@inproceedings{xu2010inference,
  author = {Yang Xu and Charles Kemp},
  title = {{Inference and communication in the game of Password}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2010},
}
Figure