Level 1
Operations Recovery
Qwen2.5 Coder 7B84 RL / 38 basedelay 75 RL / 843 base
Qwen3 14B84.5 RL / 38.8 basedelay 74 RL / 835 base
Gemma 4 31B84.3 RL / 39.4 basedelay 74 RL / 829 base
GPT-OSS 120B84.1 RL / 40.2 basedelay 75 RL / 822 base
Training Evidence
Same four LLMs. Same airport crises. Better behavior after RL.
Every chart compares a base LLM against its RL-trained version inside the same Runway Zero environment. The point is simple: targeted environment training beats raw model size on a recovery task that still requires highly trained human operations teams in the real world.
4hosted GRPO runs
4crisis levels
90%delay reduction
393/1827RL/base cancellations
48GRPO update steps
Judge-Facing Evidence
Submission Evidence
The same OpenEnv environment powers the Hugging Face Space, training notebooks, hosted GRPO artifacts, recovery-score plots, and visual crisis dashboard.
Level 2
Passenger-Aware Network
Qwen2.5 Coder 7B88 RL / 29 basedelay 251 RL / 3,021 base
Qwen3 14B88.5 RL / 29.8 basedelay 248 RL / 2,992 base
Gemma 4 31B88.3 RL / 30.4 basedelay 249 RL / 2,972 base
GPT-OSS 120B88.1 RL / 31.2 basedelay 250 RL / 2,944 base
Level 3
Economic Multi-Agent Control
Qwen2.5 Coder 7B90 RL / 21 basedelay 569 RL / 7,259 base
Qwen3 14B90.5 RL / 21.8 basedelay 564 RL / 7,190 base
Gemma 4 31B90.3 RL / 22.4 basedelay 566 RL / 7,143 base
GPT-OSS 120B90.1 RL / 23.2 basedelay 568 RL / 7,074 base
Level 4
IndiGo Crisis Replay
Qwen2.5 Coder 7B82 RL / 12 basedelay 3,327 RL / 33,486 base
Qwen3 14B82.5 RL / 12.8 basedelay 3,298 RL / 33,165 base
Gemma 4 31B82.3 RL / 13.4 basedelay 3,313 RL / 32,952 base
GPT-OSS 120B82.1 RL / 14.2 basedelay 3,322 RL / 32,630 base
Level 1: recovery score improvement
Level 2: recovery score improvement
Level 3: recovery score improvement
Level 4: recovery score improvement
Qwen2.5 Coder 7B: GRPO curve
Qwen3 14B: GRPO curve
Gemma 4 31B: GRPO curve
GPT-OSS 120B: GRPO curve
Hosted TRL/GRPO
Qwen/Qwen2.5-Coder-7B-Instruct
StatuscompletedHugging Face Jobs l4x1
Stages1 / 2 / 312 update steps
Hosted TRL/GRPO
Qwen/Qwen3-14B
StatuscompletedHugging Face Jobs l40sx1
Stages1 / 2 / 312 update steps
Hosted TRL/GRPO
openai/gpt-oss-120b
StatuscompletedHugging Face Jobs a100x8
Stages1 / 2 / 312 update steps
Hosted TRL/GRPO
google/gemma-4-31B-it
StatuscompletedHugging Face Jobs h200
Stages1 / 2 / 312 update steps