# Hydra: Unifying Document Retrieval and Generation in a Single Vision-Language Model

arXiv: https://arxiv.org/abs/2603.28554 · HTML: https://mlsanity.com/paper/2603.28554/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Document understanding and multimodal RAG, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 22 of its 33 references that are in the map.
- First posted on arXiv 30 Mar 2026.

## Placed in

- Multimodal LLMs / vision-language models / [Document understanding and multimodal RAG](https://mlsanity.com/mllm/document-understanding-and-multimodal-rag/)

## Builds on (most cited first, 22 of 22)

- [Visual Instruction Tuning](https://mlsanity.com/paper/2304.08485/) (2023)
- [Robust Speech Recognition via Large-Scale Weak Supervision](https://mlsanity.com/paper/2212.04356/) (2022)
- [InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks](https://mlsanity.com/paper/2312.14238/) (2023)
- [Towards VQA Models That Can Read](https://mlsanity.com/paper/1904.08920/) (2019)
- [ChartQA: A Benchmark for Question Answering about Charts with Visual and Logical Reasoning](https://mlsanity.com/paper/2203.10244/) (2022)
- [DocVQA: A Dataset for VQA on Document Images](https://mlsanity.com/paper/2007.00398/) (2020)
- [ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT](https://mlsanity.com/paper/2004.12832/) (2020)
- [Qwen2.5-Omni Technical Report](https://mlsanity.com/paper/2503.20215/) (2025)
- [MTEB: Massive Text Embedding Benchmark](https://mlsanity.com/paper/2210.07316/) (2022)
- [ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction](https://mlsanity.com/paper/2112.01488/) (2021)
- [InfographicVQA](https://mlsanity.com/paper/2104.12756/) (2021)
- [Scene Text Visual Question Answering](https://mlsanity.com/paper/1905.13648/) (2019)
- [Using the Output Embedding to Improve Language Models](https://mlsanity.com/paper/1608.05859/) (2016)
- [BigVGAN: A Universal Neural Vocoder with Large-Scale Training](https://mlsanity.com/paper/2206.04658/) (2022)
- [Generative Representational Instruction Tuning](https://mlsanity.com/paper/2402.09906/) (2024)
- [ColPali: Efficient Document Retrieval with Vision Language Models](https://mlsanity.com/paper/2407.01449/) (2024)
- [S-LoRA: Serving Thousands of Concurrent LoRA Adapters](https://mlsanity.com/paper/2311.03285/) (2023)
- [VDocRAG: Retrieval-Augmented Generation over Visually-Rich Documents](https://mlsanity.com/paper/2504.09795/) (2025)
- [ViDoRe V3: A Comprehensive Evaluation of Retrieval Augmented Generation in Complex Real-World Scenarios](https://mlsanity.com/paper/2601.08620/) (2026)
- [SV-RAG: LoRA-Contextualizing Adaptation of MLLMs for Long Document Understanding](https://mlsanity.com/paper/2411.01106/) (2024)
- [MeteoRA: Multiple-tasks Embedded LoRA for Large Language Models](https://mlsanity.com/paper/2405.13053/) (2024)
- [OneGen: Efficient One-Pass Unified Generation and Retrieval for LLMs](https://mlsanity.com/paper/2409.05152/) (2024)

## Abstract (arXiv)

Visual document understanding typically requires separate retrieval and generation models, doubling memory and system complexity. We present Hydra, a dual-head approach that provides both ColBERT-style late-interaction retrieval and autoregressive generation from a single vision-language model. A single LoRA adapter, trained only for retrieval, is toggled at inference: enabling it produces multi-vector embeddings; disabling it recovers the base model's generation quality, with 426 of 426 language-model weight tensors byte-for-byte identical to a freshly-loaded Qwen3.5-4B. We identify two failure modes that can silently break generation in retrieval-fine-tuned VLMs (attention-mode restoration and lm_head preservation) plus an efficiency requirement (KV-cache-aware decoding); Hydra sidesteps the first two structurally and addresses the third in the decode loop. We release two scales, Hydra-4B and Hydra-0.8B, sharing LoRA hyperparameters (r=32, alpha=32) and optimisation recipe; data mix and projection dim differ across scales. The single-model design cuts peak GPU memory from 28.85 GB to 10.77 GB at 4B (62.7% reduction) and from 5.79 GB to 2.37 GB at 0.8B (59.1%) relative to a co-resident two-model deployment. A controlled ablation finds GritLM-style joint training matches Hydra's retrieval-only training on the evaluated modes while its LoRA-on generation mode collapses. A proof-of-concept on Qwen2.5-Omni-3B preserves generation equivalence on a non-Qwen3.5 backbone and transfers image retrieval within 2-8 pp of Hydra-4B, with zero-shot audio retrieval emerging through the frozen Whisper encoder.