curr_diamond_8T

Model Card for Asymmetric-Executor-Diamond-Thinking

Model Summary

This model is a fine-tuned version of Qwen3-VL-8B-Thinking, designed for the Diamond Domain. It leverages the "Thinking" backbone's enhanced reasoning capabilities to handle the complex causal discovery required in the Diamond environment (e.g., inferring which button controls which wall based on observation history).

Model Details

• Task: Logical Puzzle Solving & Manipulation (Diamond Domain)
• Training: Fine-tuned with LoRA on a curriculum of progressively harder reasoning steps (Q1 $\to$ Q5).

Intended Use

Best suited for deployment in the Diamond Domain or similar logic-heavy grid worlds where the agent must perform self-correction and verify complex state transitions (e.g., "The wall disappeared, so the button press was successful").

Performance

• Diamond Domain Success Rate: ~94% (State-of-the-Art for this architecture).
• Task Progress: Achieves 98% task progress, indicating high robustness in completing intermediate subgoals even in failure cases.