Metacognition as Reward: Reinforcing LLM Reasoning via Knowledge and Regulation Signals

The core bet here is that a model's reasoning process contains enough self-referential signal to serve as its own reward, which sidesteps the question of who defines quality. That's a different philosophical move than automating rubric construction, and it's worth separating the two.

This paper sits in direct conversation with the ARES coverage from the same day. ARES attacks the rubric bottleneck by automating rubric synthesis from raw documents, scaling instance-level supervision without human annotation. MaR attacks the same bottleneck from the opposite direction: instead of building better rubrics, it argues you may not need rubrics at all if you can extract reward signal from the model's own metacognitive process. Together they represent two competing bets on how RL-based reasoning training escapes the annotation trap. Practitioners evaluating RL pipelines now have a genuine architectural choice to make, not just an engineering one.

Watch whether MaR's process-level rewards hold up on tasks where reasoning chains are short or highly constrained, such as formal math or code, since those are the domains where ARES-style rubrics have the clearest advantage and where metacognitive signal may be too sparse to generalize.