Atlas

Benchmarks

← All benchmarks

BALROG — TextWorld (LLM)

Games · 2024-11-20

Mean per-episode progress (0-100) on TextWorld interactive-fiction games (Treasure Hunter, The Cooking Game and Coin Collector), text-only, in BALROG's language-only track with a zero-shot agent. One of the six environments averaged into the BALROG leaderboard score; the episode count depends on the submission's protocol version.

Top models (higher is better)

ModelScore
Gemini 3.1 Pro Preview75.7
Claude Opus 571.6
GPT-5.6 Sol63.5
Grok 462.9
Gemini 3 Pro Preview60.2
Opus 4.559.0
GPT-6 Astra54.5
Gemini 3 Flash Preview50.2
Gemini 2.5 Pro Experimental 03-2549.2
Haiku 4.549.0
Claude 3.5 Sonnet (Oct 2024)42.1
GPT-4o (2024-05-13)39.3
Grok 3 (API)32.5
GPT-530.6
Gemini 2.5 Flash29.6
GPT-5.6 Terra27.8
DeepSeek-R121.8
GPT-5.6 Luna18.2
Haiku 3.518.0
Qwen2 VL 72B Instruct16.5
Llama 3.1 70B Instruct15.0
GPT-4o Mini12.3
Reka Flash 311.6
Llama 3.2 90B Vision Instruct11.2
Qwen2.5 72B Instruct11.2
DeepSeek-R1-Distill-Qwen-32B10.6
Llama 3.3 70B Instruct9.0
Llama 3.2 11B Vision Instruct6.7
Llama 3.1 8B Instruct6.1
Mistral NeMo Instruct 24074.5
Qwen2.5 7B Instruct3.9
Llama 3.2 3B Instruct3.5
Llama 3.2 1B Instruct3.3
Phi-42.5
Qwen2 VL 7B Instruct1.6
Gemini 1.5 Flash 0020.0
Gemini 1.5 Pro 0020.0
Loading Atlas data…