HyperDoc’s biggest gain comes from changing the retrieval unit—not scaling the vision model

HyperDoc reports a 53.2% generalized accuracy score on the full 1,091-question MMLongBench-Doc benchmark with Qwen3-VL-8B. That is 20.4 percentage points above MDocAgent’s 32.8%, the strongest non-HyperDoc accuracy result in the paper’s standardized 8B comparison. Generalized F1 rises from MDocAgent’s 35.6 to 42.3, a smaller but still material 6.7-point gain.

The mechanism matters more than the headline. HyperDoc does not simply retrieve more pages. It parses each PDF into page and block nodes, groups related figures, tables, captions and explanatory text into contextual hyperedges, and preserves page and section containment as structural hyperedges. At query time it predicts whether the request is localization, counting, enumeration, comparison or cross-evidence reasoning, then selects evidence groups under the model’s input budget.

The gain is not uniform across benchmarks

The authors reran all baselines with Qwen-family vision-language models on H100 GPUs, using PaddleOCR and ColPali v1.3 for common preprocessing. With Qwen3-VL-8B, HyperDoc reaches 61.7% generalized accuracy and 63.0 F1 on LongDocURL, versus 58.4 accuracy and 61.3 F1 for MDocAgent. On DocBench, HyperDoc scores 75.8% accuracy, compared with MDocAgent’s 73.7%. DocBench uses an LLM semantic judge based on GPT-4o, so that number is not directly interchangeable with the rule-based generalized metrics on the other two suites.

The sharper MMLongBench-Doc jump suggests structured evidence selection helps most where fixed page retrieval misses relations needed to assemble an answer. The paper’s ablation supports that interpretation: removing contextual hyperedges drops Qwen3-VL-8B accuracy from 53.2% to 43.0% on MMLongBench-Doc, from 61.7% to 47.2% on LongDocURL, and from 75.8% to 46.4% on DocBench. Removing structural hyperedges also hurts, but less severely.

Better answers still require an index

The improvement is not free. HyperDoc adds an offline hypergraph-construction stage taking 4.9–5.6 seconds per document, while its persistent index occupies 22.3–58.6 MB per benchmark. At inference, the full 8B configuration averages 14,300 tokens and 13.9 seconds per MMLongBench-Doc query. Removing hyperedge selection lowers answer quality and raises usage to 36,100 tokens and 26.4 seconds, showing that the selector is doing both quality and cost control.

The repository includes code, configurations, evaluation prompts and a small smoke-test subset, while full experiments require separately downloaded benchmark data and local Qwen checkpoints. This is stronger disclosure than a leaderboard-only claim, but the results are still author-run. The next useful evidence is an independent rerun that reproduces both the accuracy gains and the preprocessing/indexing costs.