Alem (Medium): leaderboard
Metric: Total return (%): each agent's cumulative reward divided by the maximum achievable total reward (376) and averaged across agents on the Medium difficulty, zero-shot homogeneous team (every agent driven by the same LLM) in alem, a JAX Craftax-like open-ended multi-agent world with procedurally generated coordination tasks (synchronous, all-agent and handover) and soft role specialisation; text interface with the rules, local observation and legal actions each step, scratchpad memory, broadcast communication and the last eight observations; reasoning enabled where available, proprietary models at high reasoning; mean over 20 seeds (10 for the API models); higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 13 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) (High) | 18.6 |
| 2 | Gemma 4 31B (IT) (Thinking) | 10.1 |
| 3 | Qwen 3.6 35B A3B | 7.7 |
| 4 | GPT-5.4 (High) | 7.4 |
| 5 | Qwen 3.5 122B A10B | 6.9 |
| 6 | Qwen 3.6 27B | 6.6 |
| 7 | Qwen 3.5 27B | 5.1 |
| 8 | gemma-4-26B-A4B-it (Thinking) | 4.9 |
| 9 | Qwen 3.5 9B | 3.6 |
| 10 | Qwen 3.5 35B A3B | 3.3 |
| 11 | Gemma 4 E4B (IT) (Thinking) | 2.7 |
| 12 | Llama 3.3 70B Instruct | 1.9 |
| 13 | Llama 3.1 8B Instruct | 0.8 |
Interactive version: theaggregate.ai/benchmark?slug=alem-medium · How It Works · Data refreshed daily, snapshot 2026-09-29.