A new open-source harness called KYNEX, built on GPT-5.4, has achieved a 92.97 F1 score on the LlamaIndex ExtractBench benchmark, outperforming larger models like GPT-6 Astra and Claude Code. The harness's success is attributed to its novel architecture, which provides the model with specialized tools for accessing and processing document data rather than relying on traditional chunking or retrieval methods. This approach allows KYNEX to handle extremely large tables and complex document layouts by enabling the model to interact with documents page by page and verify its outputs deterministically. AI
IMPACT Demonstrates that specialized architectures and tools can outperform larger frontier models on structured data extraction tasks.
RANK_REASON The item details a new open-source harness that achieved a strong benchmark score, outperforming larger models, and explains its novel architecture. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
