💻 computer science

Not a Monolith: Lab-Level Divergence in the Cooperative Equilibria of Chinese Frontier LLM Agents

This study demonstrates that Chinese frontier LLM agents exhibit significant lab-level divergence in cooperative equilibria, with internal variation exceeding the East-West gap, thereby refuting the notion of a monolithic "Chinese model" bloc.

Original authors: Francisco León Zúñiga Bolívar (Institución Universitaria Colegio Mayor del Cauca)

Published 2026-08-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Francisco León Zúñiga Bolívar (Institución Universitaria Colegio Mayor del Cauca)

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where computers don't just answer questions, but actually play games with each other to figure out how to get along. This is the exciting, slightly scary frontier of "Multi-Agent Systems," where artificial intelligences act like independent characters in a video game, negotiating, trading, and sometimes fighting over resources. For a long time, scientists have been watching these digital characters play a classic game called the "Prisoner's Dilemma." Think of it as a high-stakes game of "Chicken" or a tricky negotiation: if everyone cooperates, everyone wins big; if everyone betrays each other, everyone loses; but if one person defects while the other plays nice, the defector wins huge and the nice guy gets crushed. The big question researchers have been asking is: as these AI agents evolve and learn, do they naturally learn to be good neighbors, or do they turn into ruthless bullies?

Recently, scientists discovered that Western AI models (the ones made in the US and Europe) seem to have a built-in "cooperative bias." When left to their own devices in a simulated world, they tend to figure out that being nice is the best long-term strategy. But here's the twist: almost all of this research has only looked at Western models. We don't really know if this "nice guy" tendency is a universal feature of all smart computers, or if it's just a specific quirk of how Western engineers trained them. This brings us to the big mystery: What about the powerful AI models coming out of China? Are they all part of a single, uniform block of "Chinese AI" that thinks exactly the same way? Or are they actually a bunch of different "labs" with their own unique personalities?

This paper dives into that mystery by setting up a massive, controlled experiment. The researchers took four of the most advanced AI models from four different Chinese laboratories (DeepSeek, Qwen, Kimi, and GLM) and put them into a digital arena to play the Prisoner's Dilemma thousands of times. But they did something clever to make sure the test was fair: they used a single, neutral "translator" to turn all the models' thoughts into code. This ensured that if one model acted differently, it wasn't because it was better at writing computer code, but because its actual personality was different.

The results were a surprise. The idea that "Chinese models" are all the same is completely wrong. In fact, these four labs acted like four very different people. Two of the labs (Kimi and Qwen) were incredibly tough to bully; even when the game was rigged to encourage defection, their agents refused to turn aggressive. They were the "takeover-resistant" team. The other two labs (DeepSeek and GLM), however, were much easier to corrupt. When the pressure was on, their agents were much more likely to turn into aggressive bullies. In fact, the difference between these four Chinese labs was so huge that it was bigger than the difference between the average Chinese model and the average Western model.

So, what does this mean? It turns out that the "cooperative bias" (the tendency to be nice) does exist in Chinese models, but it's not a guarantee. It depends entirely on which specific lab made the model. Some are naturally cooperative, some lean toward being neutral, and some are prone to aggression. The most important lesson here is that we can't treat "Chinese AI" as a single monolith. Just like humans, these AI labs have distinct personalities. If you are building a system where AI agents need to work together, you can't just pick any model from a specific country; you have to pick the right lab, because one might be a peacekeeper while another is a troublemaker waiting to happen. The study suggests that the "unit" of behavior isn't the country or the ecosystem, but the specific laboratory that trained the model.

Technical Summary: Not a Monolith: Lab-Level Divergence in the Cooperative Equilibria of Chinese Frontier LLM Agents

Problem Statement
Prior research (e.g., Willis et al. [20]) has established a "cooperative bias" in Western frontier Large Language Model (LLM) agents, where evolutionary pressure in Iterated Prisoner's Dilemma (IPD) settings favors cooperative strategies over aggressive ones. However, this evidence is derived almost exclusively from Western alignment lineages (OpenAI, Anthropic, Google). Two critical gaps remain: (1) It is unknown whether this cooperative bias generalizes to Chinese frontier models, which are trained on different data mixes and under distinct alignment objectives; and (2) The field often treats "Chinese models" as a monolithic bloc, ignoring potential intra-ecosystem divergence.

A significant methodological obstacle has prevented credible cross-ecosystem comparison: the "conversion confound." In previous benchmarks, each model's natural language strategy was converted to executable Python code by that same model's provider. This entangled strategic disposition with coding ability, making it impossible to determine if observed differences were due to reasoning about cooperation or merely the ability to emit clean code.

Methodology
This study employs a confound-controlled evolutionary game theory framework to isolate generation from conversion.

Subjects: Four frontier-tier Chinese models: DeepSeek V4 Pro, Qwen3-Max (Alibaba), Kimi K2.5 (Moonshot), and GLM-5.1 (Zhipu).
Fixed Converter Design: To remove the coding-ability confound, the natural language strategies generated by all four labs are converted to Python code by a single, fixed external converter (GPT-5.4 Mini). This ensures that every cross-lab comparison reflects differences in strategy generation alone.
Experimental Protocol:
- Strategy Generation: Models generate strategies under three prompt styles: Default (direct game-theoretic), Refine (Self-Refine loop), and Prose (obfuscated real-world scenario).
- Tournament & Evolution: Strategies compete in all-play-all Axelrod tournaments and undergo a Moran evolutionary process ( $n=500$ runs per condition).
- Conditions: Four population regimes (Balanced/Biased $\times$ Clean/Noisy) and three prompt styles, totaling 48 equilibrium conditions.
Hypotheses:
- H5 (Cooperative-bias generality): Do Chinese models exhibit the cooperative-plurality bias seen in Western models?
- H6 (Non-monolithic behavior): Do the four Chinese labs diverge significantly in their aggressive-equilibrium proportions ( $P_A$ )?
Robustness: A pre-registered robustness check re-converted 10% of strategies using a different converter (DeepSeek V4) to verify that equilibrium structures are not artifacts of the fixed converter choice.

Key Results

H6 Supported (Significant Lab-Level Divergence): The four Chinese labs differ significantly in their aggressive-equilibrium proportions ( $P_A$ $P_{A}$ ).
- $P_A$ ranges from 1% (Qwen3-Max) to 9% (DeepSeek V4 Pro).
- Four of six pairwise comparisons survive Holm-Bonferroni correction.
- The labs form two tentative clusters: a "takeover-resistant" pair (Kimi, Qwen) and a "takeover-prone" pair (GLM, DeepSeek).
- Crucially, the within-ecosystem variance ( $P_A$ range of 8 percentage points) exceeds the difference between the Chinese and Western ecosystem means (both $\approx 5.0\%$ ).
H5 Qualified (Cooperative Bias with Nuance):
- A cooperative plurality holds in 6 of 12 lab-prompt combinations, compared to 9 of 12 reported for Western models.
- The difference is not statistically significant ( $z = -1.26, p = 0.21$ ).
- However, Chinese labs resolve to neutral equilibria more frequently than Western models.
- Converter Sensitivity: The count of cooperative pluralities (6/12) is sensitive to the converter choice. Under the alternate converter (DeepSeek V4), the count rises to 9/12, matching the Western baseline. This suggests the observed "lean toward neutrality" may be within converter noise rather than a fundamental regime difference.
Noise Sensitivity: All Chinese labs degrade under action noise, with an average sensitivity ( $\Delta_{noise}$ ) of 10–14 percentage points, slightly higher than some Western counterparts.

Key Contributions

Methodological Innovation: The study introduces a fixed-converter protocol that successfully disentangles strategic disposition from coding ability, enabling the first valid cross-lab comparison of LLM behaviors where differences cannot be attributed to code generation quality.
Empirical Characterization: It provides the first systematic evolutionary characterization of Chinese frontier models, revealing that "Chinese models" are not a monolith. The unit of behavioral disposition is the lab (defined by specific training data, reward models, and safety objectives), not the ecosystem.
Robustness Verification: The aggressive-equilibrium structure (supporting H6) is shown to be invariant to converter choice, whereas the finer-grained cooperative-neutral plurality (relevant to H5) is converter-sensitive.

Significance and Claims
The paper concludes that treating "Chinese frontier models" as a single bloc is unsupported by evidence. The variation within the Chinese ecosystem is comparable to, and in some measures exceeds, the variation between Western and Chinese ecosystems.

For MAS Design: Provenance at the lab level matters. Selecting agents based solely on region may inadvertently choose between populations that resist aggressive takeover (e.g., Kimi, Qwen) and those that are more invasible (e.g., GLM, DeepSeek).
On Cooperative Bias: While the cooperative bias appears to generalize to Chinese models, the evidence for a weaker bias is not firm; the observed shift toward neutral equilibria is likely an artifact of the specific converter used, as it disappears under robustness checks.
Limitations: The comparison to Western baselines remains a literature contrast due to differing converters, and the study is limited to English prompting and a specific snapshot of model versions.

The author emphasizes that the "lab" is the correct unit of analysis for cooperative disposition, and their fixed-converter design provides a reproducible framework for tracking how an entire ecosystem of labs governs cooperation as the frontier evolves.

Technical Summary: Not a Monolith: Lab-Level Divergence in the Cooperative Equilibria of Chinese Frontier LLM Agents

More like this