PreG Slashes Reasoning Token Costs

🔥 What happened Researchers from MIT and Stanford dropped PreG: instead of fully expanding every reasoning path, the method generates short prefixes, ranks them by internal confidence, and only completes the most promising candidates. 💡 Why it matters Parallel test-time scaling currently burns compute on paths that are obviously dead after a few tokens. PreG exploits those early signals and beats standard TTS on token efficiency. For anyone budgeting inference costs, this is a direct lever. ⚡ Our take Finally someone treats confidence scores as an active compute throttle, not just a diagnostic. If you're still blindly expanding every path in 2027, you're paying for garbage.