XNews

NVIDIA's SoL-Pi reduces token usage by half for coding agents

NVIDIA's SoL-Pi reduces token usage by half for coding agents

NVIDIA researchers developed SoL-Pi, an automated pipeline that discovers harness-level optimizations for coding agents. It found four mechanisms that reduce token traffic by 49% while retaining 94% of task scores on EdgeBench. The optimizations are model-agnostic and transferable across tasks.

Key insights

SoL-Pi automates discovery of token-saving mechanisms

The system searched 152 research directions across six families, ran 3,000+ experiments with 60,000+ agent-environment interactions, and identified four surviving mechanisms. The search used independent validation to prevent evolutionary overfitting, with a holdout set evaluated only once after freezing candidates.

Four mechanisms target specific token waste patterns

Action Fusion merges edit and test into one call. Context Compact compresses only when sub-tasks complete, checking if savings cover cache rewrite cost. ObservationPack archives large outputs locally, giving the model only handles and excerpts. Evidence Reducer uses a cheap model to distill logs with verification and fallback.

SoL-Pi achieves 49% token reduction with minimal score loss

On EdgeBench 51 long-horizon tasks, SoL-Pi retained about 94% of scores while cutting token traffic by 49% and API cost by about one-third. The optimizations transferred to Opus 5 with 44.7% savings, showing they target waste patterns rather than model-specific habits. Cost per unit score dropped from $0.586 to $0.417.

Source posts