Today for AI
HOT RADAR
agentHEAT 16.5°

AI CURATED HOT STORY · 10/24/2026

Meta and CMU Study: RL-Trained Small Models Outperform Llama-4 in Tool Selection

1 reports archived1 independent sourcesupdated 10/24/2026, 08:28:33
Synthesis & Latest Updates
1 Sources Cross-Validated

A new paper from Meta and Carnegie Mellon University demonstrates that an RL-trained Qwen3-8B small model outperforms a prompted 402B Llama-4-Maverick in tool selection tasks. By decomposing requests into steps and using keyword search, the small model achieved 41.5% accuracy on ToolBench, compared to 30.6% for the large model's best setup. This suggests that specialized small models can effectively replace expensive large models for decision-making in scenarios with extensive tool catalogs.

LATEST/A new paper from Meta and Carnegie Mellon University demonstrates that an RL-trained Qwen3-8B small model outperforms a prompted 402B Llama-4-Maverick in tool selection tasks. By decomposing requests into steps and using keyword search, the small model achieved 41.5% accuracy on ToolBench, compared to 30.6% for the large model's best setup. This suggests that specialized small models can effectively replace expensive large models for decision-making in scenarios with extensive tool catalogs.

HEAT TRENDHourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

TIMELINECoverage timeline

Total 1 reports · Latest first
  1. Rohan Paul (@rohanpaul_ai)T2·78 pts
    • RL-trained Qwen3-8B achieves 41.5% tool selection accuracy on ToolBench, surpassing the 30.6% of 402B Llama-4-Maverick.
    • The 'request decomposition + keyword search + single-step tool selection' workflow effectively handles context limits in massive tool catalogs.
    • The small model beats the large model on most new benchmarks without retraining, with performance improving further when paired with stronger retrievers.
Meta and CMU Study: RL-Trained Small Models Outperform Llama-4 in Tool Selection | Today for AI