When different AI models team up, do they all guess each other's moves by assuming 'they think like me'?
Do dissimilar AI models or families cooperate through the same similarity inference mechanism?
This explores whether AI models from different families, when they interact or team up, all work out what the other will do by the same process of assuming it is similar to themselves, or whether the process changes with the model.
This explores whether models from different families cooperate by the same route of inferring each other's behavior from similarity. The corpus has no note that tests that mechanism directly, so what follows is an inference from adjacent findings. Those findings point to no, and the reason is more interesting than a plain "no."
The first surprise is that models from different families are more alike than their labels suggest. A study of 70+ models on 26K open-ended queries found an "Artificial Hivemind": models built independently still produce strikingly similar or identical answers, because they share training data and alignment procedures (Do different AI models actually produce diverse outputs?). If that holds, a model that assumes "the other model probably thinks like I do" would be right more often than you'd expect across family lines. So similarity-based inference has some footing at the level of outputs.
The footing is weaker at the level of reasoning. An analysis of 22 LLMs in game-theory settings found distinct strategic styles: GPT-o1 leans on minimax reasoning, DeepSeek-R1 on trust-based reasoning, and GPT-o3-mini on anticipating the other player's beliefs. Performance tracked the type of game, not raw reasoning depth (Do large language models use one reasoning style or many?). Models that sound alike on open-ended prompts can therefore reach their decisions through different machinery, and no single similarity-based mechanism describes all three styles. Training regime seems to shape this too: reasoning and non-reasoning models stay apart however much inference compute the latter gets, because training sets up how extra tokens get used (Can non-reasoning models catch up with more compute?).
The notes on models helping each other suggest that cross-model cooperation usually runs on something other than similarity. A stronger model can nearly double a weaker one's Theory-of-Mind performance at test time by building it a harness. The harness moves unstable reasoning into deterministic code and adds task-specific routing, and it works because the models are not alike (Can a stronger model lift a weaker one at test time without retraining?). In the same way, small models learn function calling from a large teacher through explicit correct-versus-incorrect contrasts, not by resembling the teacher (Can small models match large models on function calling?). These notes cover transfer, not two agents negotiating in real time, so they only hint at how live cooperation would go.
The overall picture is that models look similar in what they say and differ in how they get there. The corpus does not yet contain a head-to-head test of cross-family cooperation, such as whether two models from different families predict each other better than same-family pairs do. That gap is where a direct answer would come from.
Sources 5 notes
INFINITY-CHAT analyzed 70+ models across 26K open-ended queries and found an "Artificial Hivemind" effect: models independently generate strikingly similar or identical responses due to overlapping training data and alignment procedures, undermining the diversity benefits of model ensembles.
Analysis of 22 LLMs across behavioral game theory reveals three dominant profiles: GPT-o1 uses minimax reasoning, DeepSeek-R1 uses trust-based reasoning, and GPT-o3-mini uses belief-anticipation. Performance correlates with game structure, not raw reasoning depth.
Reasoning models persistently outperform non-reasoning models regardless of inference budget because training instills a reasoning protocol that makes additional tokens productive. The gap is fundamentally about deployment mechanisms and training structure, not raw capability.
A stronger model built inference-time harnesses that nearly doubled weaker model performance on Theory-of-Mind benchmarks without retraining, primarily by moving unstable reasoning into deterministic code and task-specific routing rather than encouraging extended reasoning.
Small models fine-tuned via DPO on correct and incorrect function-calling examples from a large teacher model achieve high accuracy on logical and mathematical tasks. DPO's explicit negative examples directly target the rigid output format failures where SFT alone underperforms.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- ZebraLogic: On the Scaling Limits of LLMs for Logical Reasoning
- AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses
- Evaluating Theory of Mind in Reasoning Models: Robustness over Reasoning
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity
- On the Reasoning Capacity of AI Models and How to Quantify It
- LLM Strategic Reasoning: Agentic Study through Behavioral Game Theory
- Improving Small-Scale Large Language Models Function Calling for Reasoning Tasks
- Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)