Daily Paper Briefing

Abstract Overview

VeriFine is presented as an agent-harness framework for scaling verification in embodied reasoning by co-evolving the policy, the training curriculum, and the judge. The framework couples a Policy Improvement Loop, which uses a reference-free rubric judge to diagnose recurring failures and adapt training data, with a Judge Improvement Loop, which is triggered when verification becomes a bottleneck and selectively incorporates human guidance through coactive calibration. The paper studies this setup in embodied settings where evaluation must account for spatial grounding, causal reasoning, and safety-aware decisions, focusing on autonomous driving and robot navigation. Experiments are reported for both reinforcement fine-tuning in driving and supervised fine-tuning in navigation, demonstrating continuous improvement in both policy quality and judge capability across rounds.

Novelty

The core contribution is treating verification as a dynamic, co-evolving process rather than a static evaluator, explicitly updating the policy, curriculum, and judge in tandem. The framework introduces a reference-free rubric judge that evaluates free-form embodied reasoning directly from physical context and refines verification via selective human queries and coactive calibration.

Results

In autonomous driving reasoning, the policy score under a fixed reference-based evaluator increased from 60.56 to 71.61 across three rounds (an 18.2% relative improvement), while the distilled student judge attained a 0.82 correlation with humans and 0.17 MAE. In test-time candidate selection across six candidates, the evolved judge chose outputs achieving a 78.6 human-rated reasoning score. On robot navigation, the policy reasoning score improved from 68.09 to 77.59 (a 14% increase), with judge-human correlation approaching 0.85.

Key Points

1. The framework operates via two coupled loops: a policy improvement loop for failure-driven curriculum selection and optimization, and a judge improvement loop that refines verification through coactive human calibration when progress plateaus.
2. A reference-free rubric judge serves as the primary verification substrate, generating structured diagnostic scores directly from physical context rather than relying on human-written reference answers.
3. Empirical evaluations confirm sustained gains in policy performance and judge alignment across driving reinforcement fine-tuning and robot navigation supervised fine-tuning, as well as in test-time candidate selection.

References

• arXiv: https://arxiv.org/abs/2610.08761v1
• Fugu-MT: https://fugumt.com/fugumt/paper_check/2610.08761v1
• Hugging Face Papers: https://huggingface.co/papers/2610.08761