Evolve the inference program. Keep the model frozen.

EvoVLM is a research project on LLM-guided, multi-objective evolutionary search for vision-language inference. It searches over prompts, image preprocessing, generation budgets, and answer extraction to improve the accuracy–latency trade-off on scientific chart understanding, without fine-tuning model weights.

中文：EvoVLM 在固定视觉语言模型权重的前提下，通过大模型引导的变异、实测评估和 Pareto 选择，迭代优化推理程序，在图表理解任务上探索准确率与推理效率的平衡。

Method · Results · Quick start · Technical report · Reproducibility

The implementation follows the evolutionary-algorithm pattern of mutation, evaluation, and selection, with a constrained program representation:

1. Start from manually designed inference programs for Qwen3-VL-2B-Instruct and Qwen3-VL-2B-Thinking.
2. Use a local Qwen model to rank a frozen catalog of typed, focus-specific mutations. A trusted renderer turns the selected configuration into Python.
3. Evaluate candidate programs on small search and selection subsets, recording correctness, latency, errors, and execution provenance.
4. Apply Pareto-based selection to retain useful accuracy–latency trade-offs and continue the next generation.
5. Export selected programs and verify their recorded evaluation evidence.

The recorded search is deliberately small: two generations, with at most six children per model population. It is not unrestricted code generation, weight training, or a claim to implement a named algorithm such as NSGA-II in full. The implementation's distinguishing choices are constrained LLM-guided mutations, executable inference programs, and fail-closed evidence verification.

Recorded results on Dev128 and a separate Speed32 timing protocol:

Relative to its manual seed, Evolved Instruct preserves the recorded accuracy with an 8.9% lower median latency. Evolved Thinking improves accuracy by 16.41 percentage points with a 10.6% lower median latency.

Numbers come from the sealed metric trace, not from a new run made for this GitHub release. Speed32 first takes the median across three repeats for each query, then the median across the 32 queries. The unmodified Thinking baseline was timeout-censored at 7,200 seconds; a completed Dev128 accuracy is not available for it.

Interpretation matters: Dev128 is a public validation subset, not an untouched test set. Two exact queries overlap the evolution-used set, and eight Dev128 queries share two figures with it. These are small-sample development results, not a full CharXiv leaderboard result or evidence of generalization. Latency is specific to the recorded hardware and software environment.

git clone https://github.com/zzzz7788990213-ops/EvoVLM.git
cd EvoVLM/research/evolving-vlm-inference-main

# CPU-only: replay recorded evidence, then run the final checks.
python3 -B -m evaluation.portable_reproduce
bash reproduce.sh check

# CPU-only: import the exported programs and verify their interface.
bash reproduce.sh smoke

The CPU checks do not require model weights and do not rerun inference. The historical directory name is retained because some verification routines bind to it. Keep research/evolving-vlm-inference-main intact.

Always specify a mode: bash reproduce.sh without a mode requests a fresh GPU reproduction. It is not the CPU verification command.

GPU inference requires PyTorch, Transformers, the other dependencies, and the external model snapshots. The recorded environment is documented in requirements.lock.txt; it includes Linux/CUDA-specific packages and is not a universal macOS lockfile.

See the runtime guide for pinned model downloads, local overrides, offline behavior, and example commands. Model weights are not included in this repository.

EvoVLM/
├── README.md
├── THIRD_PARTY_NOTICES.md
├── PUBLICATION_PROVENANCE.json
└── research/evolving-vlm-inference-main/
    ├── evolution/           # typed mutations, local provider, search, selection
    ├── evaluation/          # protocols, identity checks, semantic verification
    ├── tests/closeout/      # portability, evidence, and tamper tests
    ├── results/            # preserved experiment records and journals
    ├── data/               # fixed manifests and reproduction image assets
    ├── charxiv/            # upstream data, attribution, and selected images
    ├── best_*.py           # selected inference programs
    ├── reproduce.sh        # explicit CPU and GPU modes
    └── report.pdf          # archived technical report

The repository includes source code, packaged data subsets, selected image assets, and recorded evidence. Model snapshots and the execution environment are external dependencies. This is not a self-contained model distribution or the complete CharXiv image dataset.

Verification covers file hashes, critical model/processor identity, cross-bound attempt records, and recomputation of metrics from per-query evidence. Missing historical fields are recorded as legacy caveats rather than invented.

This public edition changes documentation and adds missing upstream license texts; it does not change inference code, model settings, selected programs, or recorded results. The original local archive is preserved separately. The publication provenance identifies the parent archive and every in-package publication change.

Historical evidence retains original run identifiers and server-path strings for auditability. Those paths are not local setup instructions. The archived report retains its original title and formatting; repository publication does not imply conference submission or acceptance.

EvoVLM uses Qwen model families and CharXiv data. Upstream authorship and bibliographic references are retained; their inclusion does not imply project endorsement or collaboration. See third-party notices.

No new repository-wide license is imposed by this publication. Existing third-party licenses continue to apply to their respective materials.