Paper claims 9B model can reach frontier reasoning via Parallel Power Tempering sampling, no RL post-training
dwgreyman · x · 2026-10-01
A paper proposing Parallel Power Tempering (PPT) argues a 9B model can approach GPT-5 / Opus-4.5-level reasoning through inference-time sampling alone. PPT runs multiple interacting replicas at different sharpening levels: low-power replicas explore diverse reasoning trajectories while high-power chains exploit high-likelihood answers, fixing truncation bias found in earlier power samplers.
- No parameter updates or RL post-training, avoiding costly optimization and jagged generalization
- Authored by the poster's intern's summer project; engineering work remains
- Authors believe small models contain far more intelligence than commonly assumed
Related event: Parallel Power Tempering Pushes 9B Models Near GPT-5 Without RL(2 posts)→