LoRA adapter fine-tuning Qwen/Qwen3-VL-8B-Instruct for RoboCasa tool-calling agent behavior with causal-consistent active-observation caching.

This is a follow-up to qwen3vl-8b-robocasa-v3-active-observation that fixes a training/eval consistency bug in how image observations were attached to training examples.

The bug: the original v3 training pipeline decided whether to attach the latest observation image to an example based on last_agent — effectively conditioning on which agent had most recently produced the acting example, which in practice let the model infer the correct next tool call from whether an image was present at all, rather than from genuine visual/task reasoning. This "cheat code" (image present → call a comm/physical tool; image absent → call get_image) is not available at inference time in the same closed-loop way it appeared during training, so it broke causal consistency between train and eval.

The fix: introduced a causal_single_cache mode that tracks an explicit active_observation_agent cache during example construction. The cached image owner only updates when a get_image call occurs, and resets to None whenever a different agent takes a non-communicate/non-task-complete action. Image attachment now depends only on this causally-prior state, not on which tool call comes next in the trajectory.

Same RoboCasa AgentSFT subset as the base v3 run — 46 kitchen tasks (see run_config.json for the full task list), tool-call SFT format.

8x NVIDIA B200 (single node), ~2h30m wall clock.

W&B run: qwen3vl-8b-v3-active-observation-causal