VERGE

VERGE: Verified Early-exit Reasoning with Grounded Evidence is a unified GRPO training project for medical multiple-choice QA, visual QA, and normalized bounding-box grounding. It combines the validated Step-QA path from stepEasyR1 with the bounded-IoU Step-Grounding path from stepGrounding.

The central rule is simple: forced early-exit completions are judges, not PPO trajectories. VERGE verifies that an original reasoning prefix already commits to a correct answer, then redistributes token-level advantage only on the original sampled trajectory. A task-normalized auxiliary loss teaches the model to emit </think> at verified cut points.

See VERGE.md for the algorithm and implementation map.

Mixed training

The main experiment uses:

• 3,000 MedMCQA examples;
• 3,000 PMC-VQA examples;
• 6,000 grounding examples;
• 300 validation examples from each dataset family.

Run the two-GPU functional smoke test first:

bash examples/smoke_verge_mixed_gpu2_3.sh

Then start the four-GPU experiment:

bash examples/run_verge_mixed12k_2epoch_val5_gpu2_5.sh

Both launchers start from /opt/data/private/rby/weights/Qwen3-VL-4B-Thinking.