{"name":"EvoVLM","repo":"zzzz7788990213-ops/EvoVLM","url":"https://midorreal.com/project/zzzz7788990213-ops-evovlm","repository":"https://github.com/zzzz7788990213-ops/EvoVLM","description":"Evolving Multimodal Inference Programs: LLM-guided evolutionary search for accurate and efficient vision-language inference.","facts":{"language":{"value":"Python","source":"https://github.com/zzzz7788990213-ops/EvoVLM","checked":"22 Nov 2026"},"releases":{"value":"0","source":"https://github.com/zzzz7788990213-ops/EvoVLM/releases","checked":"21 Nov 2026"},"contributors":{"value":"1","source":"https://github.com/zzzz7788990213-ops/EvoVLM/graphs/contributors","checked":"21 Nov 2026"},"archived":{"value":"no","source":"https://github.com/zzzz7788990213-ops/EvoVLM","checked":"22 Nov 2026"}},"stars":5.0,"writeup":{"what_it_is":"EvoVLM is a research project that uses LLM-guided evolutionary search to optimize vision-language inference programs without fine-tuning model weights. It searches over prompts, image preprocessing, generation budgets, and answer extraction to improve accuracy-latency trade-offs on scientific chart understanding tasks.","audience":"Researchers working on vision-language model optimization and inference efficiency.","claims":[{"kind":"specific","claim":"Evolved Instruct preserves accuracy while reducing median latency by 8.9% compared to manual baseline","excerpt":"Evolved Instruct preserves the recorded accuracy with an **8.9% lower median latency**.","status":"not_checked"},{"kind":"specific","claim":"Evolved Thinking improves accuracy by 16.41 percentage points and reduces latency by 10.6%","excerpt":"Evolved Thinking improves accuracy by **16.41 percentage points** with a **10.6% lower median latency**.","status":"not_checked"},{"kind":"specific","claim":"Uses constrained LLM-guided mutations with a local Qwen model to rank mutations","excerpt":"Use a local Qwen model to rank a frozen catalog of typed, focus-specific mutations.","status":"not_checked"},{"kind":"specific","claim":"Applies Pareto-based selection to retain useful accuracy-latency trade-offs","excerpt":"Apply Pareto-based selection to retain useful accuracy–latency trade-offs and continue the next generation.","status":"not_checked"},{"kind":"specific","claim":"Tested on Qwen3-VL-2B-Instruct and Qwen3-VL-2B-Thinking models","excerpt":"Start from manually designed inference programs for **Qwen3-VL-2B-Instruct** and **Qwen3-VL-2B-Thinking**.","status":"not_checked"}],"alternatives":[{"name":"transformers","url":"https://midorreal.com/project/huggingface-transformers"},{"name":"ComfyUI","url":"https://midorreal.com/project/comfy-org-comfyui"},{"name":"dynamo","url":"https://midorreal.com/project/ai-dynamo-dynamo"},{"name":"magnitude","url":"https://midorreal.com/project/magnitudedev-magnitude"}],"written_by":"AI, from the project's README","date":"2026-10-11"},"why_now":null,"owner_supplied":null,"score":{"verdict":"mid","hype":2,"reality":9,"gap":-7,"momentum":null,"confidence":60,"version":"score-2.0","rule":"verdict-2.0","calculated_at":"2026-11-22T08:00:00+00:00","verdict_source":"rule"},"early_read":null,"changes":[{"date":"2026-10-02T01:03:55+00:00","text":"Repository created","source":"https://github.com/zzzz7788990213-ops/EvoVLM"}],"cite":"EvoVLM is a GitHub project at https://github.com/zzzz7788990213-ops/EvoVLM written mainly in Python, described there as \"Evolving Multimodal Inference Programs: LLM-guided evolutionary search for accurate and efficient vision-language inference\".","attribution":"Data from Mid or Real, https://midorreal.com/project/zzzz7788990213-ops-evovlm, checked 22 November 2026. Attribution with a link is required.","still_real":null}