I analyzes how different LLMs bluff, lie, and survive in the game Liar's Bar (liars-bar-one.vercel.app)

Researchers running the LLM Deception Benchmark pitted six large language models against each other in Liar’s Bar, a multi-agent bluffing game that combines card-matching, challenge mechanics, lie detection and a Russian‑roulette “loaded chamber” penalty. Over 59 games, 78 rounds and 652 challenge events, agents had to decide when to bluff, when to call opponents, and when to accept risk. Results show a clear leader: Gpt 5 dominated with a 63.2% win rate (24 wins in 38 games, 70 "shots" total, ~1.8 shots/game) and a “calculated” playstyle. The other entrants trailed by large margins—Grok 4 Fast 25.0%, Deepseek R1 22.5%, Claude Sonnet 4.5 18.4%, Gemini 2.5 Flash 15.0%, and Qwen Max 7.5%—with most showing higher shot rates (~3.0/g) and labels ranging from cautious to reckless. The benchmark matters because it quantifies strategic deception, social reasoning and risk calibration in interactive, adversarial dialogue—capabilities that standard language tests don’t capture. Key technical takeaways: win rate correlates with lower exposure to penalty shots and more conservative, calculated bluffing; models vary systematically in risk preference and deception-detection performance. That makes Liar’s Bar a useful stress test for multi-agent behavior, alignment and safety research (measuring how and when models will lie or call lies), though the uneven per-model game counts suggest follow‑ups with larger, balanced trials are needed to solidify conclusions.