curr_diamond

Model Card for Asymmetric-Executor-Diamond-Instruct

Model Summary

This model is a fine-tuned version of Qwen3-VL-8B-Instruct, specialized for the Diamond Domain. This domain requires handling strict sequential dependencies (Key -> Door -> Button -> Diamond) and high environmental stochasticity (randomized button/wall logic).

The model is trained to strictly follow the Reflection-Action protocol, allowing it to detect execution conflicts (e.g., "Key Mismatch") and signal the High-Level Planner for re-planning.

Model Details

Task: Sequential Interaction & Causal Discovery (Diamond Domain).

Key Capability: Conflict Detection (e.g., detecting when a key does not fit a door) to trigger the event-based planner.

Training: Utilizes a "Key Episode" dataset to handle sparse reward signals for subgoal completion.

Inference Format (System Prompt)

The model expects the following prompt structure:

Perception: "Where am I? What key objects do I see?"

Verification: "Have I completed the current subgoal? (YES/NO)"

Graph Transition: "Signal planner if Complete/Conflict."

Spatial Reasoning: "Where is the target relative to me?"

Action: "Atomic Action."

Performance

Diamond Domain Success Rate: ~92% (comparable to Gemini 3 Pro based SH-VLM).

Cost Efficiency: Reduces API calls to the high-level planner by 37x.