Can 1B LLM Surpass 405B LLM? Optimizing Computation for Small LLMs to Outperform Larger Models

Abstraction: Test-time scaling lets small LLMs outperform much larger models

Key points:

• Compute-optimal Test-Time Scaling (TTS) allows a 1B model to outperform a 405B model on MATH-500 and AIME24 benchmarks
• TTS improves efficiency by up to 256x compared to majority voting and boosts reasoning 154.6% over standard Chain-of-Thought
• Process Reward Models (PRMs) guide answer selection; on-policy PRMs yield more accurate rewards than offline ones
• A 7B PRM can effectively supervise a 72B policy model — "weak-to-strong" supervision dynamic
• Reward-aware TTS strategy integrates reward signals into compute allocation for better scaling
• Research from Shanghai AI Laboratory, Tsinghua University, Harbin Institute of Technology, and BUPT

Connections: Shanghai AI Laboratory · Deepseek R1 · Large Language Models · Test Time Compute · Chain Of Thought

Source: https://www.marktechpost.com/2025/02/13/can-1b-llm-surpass-405b-llm-optimizing-computation-for-small-llms-to-outperform-larger-models/