Verbal Process Supervision Boosts LLM Reasoning Without Training

There's this new method called Verbal Process Supervision (VPS) that boosts how well large language models think by introducing critiques from a more sophisticated model, and it doesn't need any extra training. For instance, on the GPQA Diamond benchmark, GPT-5.4 achieved an impressive 94.9% accuracy with a round budget of R=4, beating the previous best of 94.1% without any gradient tweaks. At AIME 2025, VPS really helps less powerful models, raising their performance from 11.7–26.7% to an amazing 63.3–90.0%, which is a jump of up to 63.3 points. When looking at similar computing power, VPS outperforms Reflexion by up to 12.1 points and Self-Consistency@5 by 5.0 points on GPQA, plus 8.3 points on LiveCodeBench V6.