Gemini 3.6 Flash shows language-dependent negotiation behavior in study
Researchers tested whether language affects how large language models negotiate by using identical Gemini 3.6 Flash agents to bargain in English and Korean. Across 12 scenarios, Korean negotiations succeeded 77.3% of the time versus 52.5% in English—a 24.8-point gap. The difference emerged later in talks, with Korean sellers conceding more and final prices lower, despite similar opening offers.…
Key points
- Gemini 3.6 Flash agents reached agreement 77.3% of the time in Korean vs. 52.5% in English, a 24.8-point gap
- Korean sellers conceded more and final prices were lower, despite identical opening offers in both languages
- Politeness in Korean dialogue reduced face-threatening acts by 11.7% versus 3.6% in English, per pragmatic analysis
The study, published on arXiv, used a 2×2×3 design with 432 negotiations, comparing English and Korean interaction languages, seller urgency levels, and buyer personas (cooperative, neutral, headstrong). A repeat with Qwen3.7 Flash supported the trend: Korean negotiations also closed more often, though the effect was smaller. The findings highlight how cross-lingual social-policy fidelity—whether a prompt’s behavioral intent holds across languages—varies by context.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Researchers introduce ResMem to turn retrieved knowledge into parametric memory · 1 src
- Circuit Oracle automates attribution graph analysis · 1 src
- arXiv study finds AI humanizers alter 35–65% of fact units in documents · 1 src
- Researchers propose gavagai to improve Vision-Language model alignment · 1 src
- Researchers outline framework for unifying embedding models across modalities · 1 src
Comments
via GitHub Discussions