ARIA: Adaptive Reasoning through Difficulty-Graded Trace Compression
A reasoning model that learns to allocate thinking tokens by problem difficulty — 2.3× fewer tokens with a slight accuracy gain on GSM8K and MATH-500.
2026 · PDF
A reasoning model that learns to allocate thinking tokens by problem difficulty — 2.3× fewer tokens with a slight accuracy gain on GSM8K and MATH-500.
2026 · PDF