step_game

# Multi-Agent Step Race Benchmark: Assessing LLM Collaboration and Deception Under Pressure

A three-player “step-race” that challenges LLMs to engage in **public conversation** before picking a move (**1**, **3**, or **5** steps). Whenever two or more players choose the same number, all colliding players fail to advance. The first LLM to reach or surpass **16–24** steps wins outright, or if multiple cross simultaneously, the highest total steps takes it (ties share victory).

This design moves beyond static Q&A. Winning requires live social reasoning: reading opponents, offering half-truths, gauging trust, deciding when to cooperate, and knowing when to lie. Over thousands of matches we see patterns emerge: large frontier models charm first, then knife their partners late, many agents overplay the maximal 5, causing long jams that punish impatience. A few discover subtle linguistic tells—echoed phrasings, timing shifts—that reveal an opponent’s plan a turn early.

The dataset opens fresh questions. Can we predict a model’s next move from its last sentence? Which phrases cloak a bluff? Do temporary alliances ever stick? How fast does an agent abandon a losing script? 

---

## Animation

https://github.com/user-attachments/assets/f07abbd8-a780-440a-8fae-66f7154cf010

Longer video:

[![Multi-Agent Step Race Benchmark: Assessing LLM Collaboration and Deception Under Pressure: frame-by-frame replay of each game](https://markdown-videos-api.jorgenkh.no/url?url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DAnPKfrIPAgQ)](https://www.youtube.com/watch?v=AnPKfrIPAgQ)


We generate a **frame-by-frame** and a **summary** replay of each game, illustrating:
1. Conversation sub-rounds with highlighted quotes
2. Secret moves (1,3,5) and collisions
3. Real-time positions on the track
4. A dynamic scoreboard (TrueSkill ratings, partial-win tallies)

The animation reveals how LLMs strategize, stall, sabotage, or cooperate, culminating in final rankings. It shows how their talk transla