GEPA Optimizer Lifts MATH Benchmark Performance from 67% to 93% Through Prompt Refinement Alone
New DSPy evolutionary optimizer achieves dramatic accuracy gains without fine-tuning, signaling major shift in prompt engineering efficiency.
AI-generated summary. This news item was written automatically by an AI pipeline from published reporting and was not reviewed by a person before publication. It may contain errors. Read the original source.
GEPA Achieves Dramatic Benchmark Gains Through Instruction Refinement
GEPA, an evolutionary prompt optimizer in the DSPy ecosystem, has lifted a chain-of-thought program on the MATH benchmark from a 67 percent baseline to 93 percent through instruction refinement alone, with no few-shot examples and no fine-tuning. The optimizer was presented as an ICLR 2026 paper.
In a real-world deployment, Shopify moved a single GPT-5 prompt to a small open model optimized with GEPA and reported a solution roughly 75 times cheaper and twice as reliable.
Chain-of-Thought Performance Varies Significantly Across Models
According to a Wharton Generative AI Labs report, forcing explicit chain-of-thought on reasoning models shows inconsistent results across different architectures. On o3-mini, the approach moved accuracy by 2.9 percent. On o4-mini, accuracy improved by 3.1 percent. However, on Gemini Flash 2.5, forcing explicit chain-of-thought reduced accuracy by 3.3 percent.
The same technique adds 20 to 80 percent to response time across reasoning models, per the Wharton report.
Anthropic Shifts Claude Opus 4.8 to Adaptive Thinking Only
On Claude Opus 4.8, the fixed budget_tokens approach for thinking has been removed; adaptive thinking is now the only thinking-on mode.
Source: Autocomplete AI / Substack
Irish pronunciation
All FoxxeLabs components are named in Irish. Click ▶ to hear each name spoken by a native Irish voice.