Process Supervision via Verbal Critique Improves Reasoning in Large Language Models

The key distinction buried in the framing is 'training-free': Verbal Process Supervision doesn't update model weights, which means the gains come entirely from how inference is structured, not from any new capability baked into the model. That makes the 11.7% to 90% jump on AIME 2025 either a compelling demonstration of how much headroom exists at inference time, or a sign that AIME 2025 may be more susceptible to iterative critique than its difficulty rating implies.

This connects most directly to the AEL paper also published April 23, which tackles a structurally similar problem from a different angle: where AEL asks how agents can improve across episodes by retaining memory and diagnosing failure patterns, Verbal Process Supervision asks how a single reasoning chain can improve within one inference pass through iterative critique. Both are responses to the same ceiling: chain-of-thought alone stops scaling cleanly. The broader archive here doesn't yet have coverage of competing process supervision approaches like outcome-reward models or tree search methods, so readers should treat this as an entry point into a contested space rather than a settled one.

Watch whether these benchmark gains replicate on held-out GPQA Diamond splits released after April 2026 — if they don't, eval contamination is the most likely explanation given the training-free claim. Also watch whether any lab publishes a direct comparison against outcome-based reward models on AIME 2025 within the next two months, which would clarify whether verbal critique adds anything beyond what scalar reward signals already provide.