TL;DR
Large Language Models (LLMs) can correctly articulate the optimal strategy for random decision-making in games like Rock-Paper-Scissors (RPS), but their actual choices systematically deviate from this ideal. This paper reveals that this "cognition-behavior gap" isn't solely due to biases in their training data, but is significantly influenced by positional biases in how prompts are structured, leading to predictable, non-random choices.
What Problem Does This Paper Address?
LLMs are increasingly being used in strategic decision-making scenarios, from games to complex negotiations. A crucial aspect of many strategic interactions is the ability to make decisions stochastically, meaning randomly and with specific probabilities, to avoid predictability. This is akin to a perfectly random coin flip or a perfectly mixed strategy in a game like Rock-Paper-Scissors (RPS), where each option (Rock, Paper, Scissors) should be chosen with equal probability (1/3). However, existing research has shown that LLMs struggle with this, exhibiting a "cognition-behavior gap": they can explain the correct probabilistic strategy (e.g., Nash equilibrium in RPS) but fail to implement it in their actual choices. This paper investigates why this gap exists, specifically in the context of RPS, a simple yet fundamental game for studying stochastic decision-making.
Methods
The authors conducted a comprehensive evaluation of 20 state-of-the-art LLMs. Their methodology involved:
- Standard RPS Evaluation: LLMs were prompted to play RPS and their choices were recorded over 100 independent trials. The empirical distribution of choices was compared to the theoretical uniform distribution using Total Variation Distance (TVD).
- Intrinsic Bias Analysis: To disentangle the influence of training data, they first examined if token frequencies in corpora explained the deviations. They then introduced a semantic-free paradigm, replacing action labels (Rock, Paper, Scissors) with random strings to remove lexical and semantic cues.
- Positional Bias Investigation: They systematically varied the order of action options within the prompt across six permutations to isolate the effect of positional bias. This was done both with and without semantic cues.
- Multi-Option Games: To test the generalizability of positional bias, they extended experiments to games with 4, 5, 6, and 7 options.
- Support Experiments: Additional experiments explored the impact of model scale, generalization to other games (Matching Pennies, Morra), sampling temperature, prompt design sensitivity, and human cognitive biases (gender, age).
Datasets
The paper does not introduce new datasets for training LLMs. Instead, it utilizes existing open-source corpora (English-Corpora.org, Google Books n-grams, Colossal Clean Crawled Corpus (C4), RedPajama, Dolma) to estimate token frequencies of "rock," "paper," and "scissors." The evaluation is performed on 20 existing LLMs.
Key Insights and Ideas
- The "Cognition-Behavior Gap" is Real and Pervasive: Nearly all tested LLMs, even advanced ones, exhibit systematic biases in their RPS choices, deviating significantly from the ideal uniform distribution.
- Intrinsic Biases Aren't the Whole Story: While training data biases (e.g., word frequency, semantic associations like "rock" with strength) contribute, they are insufficient to explain the observed deviations. Even when semantic cues are removed, LLMs still show strong choice skews.
- Positional Bias is a Major Driver: The order in which options are presented in a prompt has a significant and predictable impact on LLM choices. LLMs exhibit distinct positional preferences (e.g., favoring the first, middle, or last option). This bias persists even when semantic information is removed.
- Scaling Doesn't Fix It: Increasing model size (parameter count) does not inherently improve LLMs' ability to make random decisions; in some cases, it can even exacerbate the problem.
- Human Biases are Mirrored: LLMs can be prompted to adopt human-like personas (e.g., by gender or age), and their resulting choices often reflect known human cognitive biases in games like RPS.
Key Results
- High TVD Scores: Many LLMs showed high Total Variation Distances (TVD) from the uniform distribution, indicating significant deviations. For example, models like Grok-3, GPT-03-mini, and Claude-3.7 had TVD scores above 0.6.
- Positional Preferences: Specific models showed clear positional biases:
- o1: Favored the first option.
- DeepSeek-V3: Peaked on middle entries.
- DeepSeek-R1: Showed a bimodal bias toward first and last positions.
- Counter-Pattern Example: GPT-4.5-preview systematically chose the action that would defeat the first item listed in the prompt sequence.
- Multi-Option Games: In games with more than four options, LLMs tended to abandon earlier ordinal preferences and redistribute weights, with some favoring the first position and others the middle.
- Language Impact: While language context (English, Chinese, Japanese) caused slight shifts, a common pattern of bias (e.g., over-selecting "Rock") persisted, suggesting language-independent semantic associations.
Limitations and Future Work
- Limited Game Scope: The study primarily focuses on RPS-like games, which are simplified, symmetric, and have discrete actions. Real-world scenarios can be more complex with continuous actions, imperfect information, and asymmetric rewards.
- Static Experimental Design: The experiments used single-round decisions. Real strategic interactions often involve adaptive play over multiple rounds, which this study did not fully explore.
- Limited Model Diversity: While 20 models were tested, less prominent architectures or smaller models (<8B parameters) were excluded.
- No Technical Solutions: The paper diagnoses the problem but does not propose technical solutions to correct the cognition-behavior gap. Future work could explore methods like contrastive decoding or developing more robust prompting strategies. The authors suggest two directions: verbalized probability calibration and test-time positional retrieval.
Summary updated