Recursive Self-Improvement Delivers New SOTA Coding Performance
a day ago
- #Recursive Self-Improvement
- #Model-Agnostic Optimization
- #AI Coding Benchmark
- Poetiq's Meta-System automatically creates and optimizes harnesses from scratch, improving performance on LiveCodeBench Pro (LCB Pro) without fine-tuning or special model access.
- LCB Pro is a coding benchmark with memory and runtime constraints, designed to mitigate data contamination and test complex problem-solving in C++.
- The optimized harness, initially built for Gemini 3.1 Pro, boosts its accuracy by 12.3% (from 78.6% to 90.9%), surpassing GPT 5.5 High and Gemini Deep Think.
- The same harness improves other models: GPT 5.5 High reaches 93.9% accuracy, and Gemini 3 Flash sees a 10-point increase, outperforming larger models like Claude Opus 4.7.
- Improvements are consistent across difficulty levels (Easy, Medium, Hard), with the harness outperforming base models in every category.
- Poetiq's approach is model-agnostic, applying learned harnesses to any LLM via recursive self-improvement, and has also shown success on reasoning and retrieval benchmarks like ARC-AGI and HLE.
- The system focuses on automating knowledge extraction for tasks involving reasoning, retrieval, and coding, aiming to enhance AI's economic impact and problem-solving capabilities.