MIRA: A Multimodal Image-Text Deep Research Agent for Controllable Report Generation

Title

MIRA: A Multimodal Image-Text Deep Research Agent for Controllable Report Generation

Meta Description

MIRA is a multimodal image-text deep research agent that controls report granularity and multimodal composition through generation contracts, achieving open-source SOTA on 7 benchmarks with traceable text and visual evidence.

Article Content

Deep research agents can already search the open web and synthesize evidence, but controllable report generation remains a weak spot in multimodal deep research. Such tasks require agents to handle image-text interleaved inputs, collect multimodal evidence through multi-step web searches, and output complete image-text reports according to user requirements. Existing systems have two limitations: at inference time, they often underutilize images discovered in the search trajectory; at generation time, they tend toward fixed-form answers with limited control over report granularity, multimodal composition, and evidence grounding. To address this, researchers propose MIRA—a multimodal image-text deep research agent oriented toward controllable reports.

Core Problem: Fragmented Image-Text Evidence and Lack of Control

Most current deep research agent workflows are built around textual evidence, with images often treated as auxiliary information or even ignored during reasoning. This leads to final reports lacking visual support and struggling to meet users' needs for content that combines text and images. On the other hand, the report generation stage is typically end-to-end: users can only specify a topic but cannot control the report's level of detail, the image-text ratio, or the evidence source behind each claim. This "black-box" generation limits the agent's usability in professional scenarios.

MIRA's starting point is precisely to thread textual and visual evidence throughout the entire research and writing process, and to translate user requirements into executable generation contracts.

Deep Research Agent refers to an intelligent agent system capable of autonomously performing multi-step web searches, integrating information across sources, and generating structured research reports. Unlike traditional retrieval-augmented generation (RAG), deep research agents emphasize "multi-hop" reasoning: they dynamically adjust subsequent search strategies based on intermediate findings rather than retrieving once and answering directly. Representative works include OpenAI Deep Research and Perplexity. The core challenges of such systems lie in search trajectory planning, evidence credibility assessment, and the ability to organize long reports. When tasks extend from pure text to image-text interleaved scenarios, the agent must also judge when to retrieve images and how to align visual evidence with textual arguments, which significantly increases the complexity of reasoning and generation.

MIRA-Harness: Connecting Four Stages with Generation Contracts

MIRA's core component is MIRA-Harness, which translates user requirements into a "generation contract" and connects four stages: evidence collection, report submission, feedback-driven revision, and final delivery. This design makes report generation no longer a one-way prompt-output process but an iterative, controllable one.

During the evidence collection stage, the agent performs multi-step web searches, simultaneously capturing textual and visual evidence and recording their sources, laying the foundation for subsequent traceability. In the report submission stage, the agent produces a first draft according to the generation contract, which specifies constraints such as granularity and multimodal composition. The feedback-driven revision stage allows the report to be adjusted based on user or automated evaluation feedback, and the final delivery stage ensures the output meets the contract requirements.

This staged, contract-driven approach enables the agent to maintain autonomy while exercising fine-grained control over key report attributes.

MIRA-Data and Model Post-Training: Building High-Quality Multi-Hop Trajectories

To enable the base model to operate within this workflow, the research team built MIRA-Data to generate high-quality multi-hop multimodal research trajectories. These trajectories simulate the real research process: starting from an initial question, going through multiple searches and evidence integration, and ultimately forming a report. Based on this data, the team post-trained the base model so that it learns to act within the process defined by MIRA-Harness.

The key to post-training is enabling the model to understand the constraints of the "generation contract" and to learn to actively leverage image evidence during search. This differs from traditional instruction fine-tuning in that it injects the structured knowledge of the workflow into the model.

Multi-hop trajectory refers to the complete sequence of actions an agent takes from an initial question through multiple searches, evidence filtering, and integration, gradually approaching the answer. Each step includes an action (e.g., issuing a query, opening a page, extracting an image) and an observation (the retrieved text or image content). In training, such trajectories can serve both as target sequences for supervised fine-tuning and for preference learning, enabling the model to distinguish high-quality from low-quality search paths. Compared with single-hop QA data, multi-hop trajectories can teach the model when to continue searching, when to stop, and how to establish connections across different sources—making them a key data form for training deep research agents. The value of MIRA-Data lies precisely in embedding the acquisition and use of image evidence into this trajectory.

Benchmark Performance: SOTA Among Open-Source Agents

Using Qwen3.6-35B-A3B as the base model, MIRA achieves state-of-the-art performance among open-source agents on 7 representative multimodal search and deep research agent benchmarks. These benchmarks cover multimodal search and deep research tasks, enabling comprehensive evaluation of the agent's evidence collection, synthesis, and report generation capabilities.

More importantly, the image-text reports generated by MIRA align with user requirements, and both textual and visual evidence are traceable. This means users can not only see the conclusions but also trace back to the original materials on which each conclusion is based, including image sources.

Significance and Limitations

MIRA's value lies in introducing "controllability" into multimodal deep research report generation. Through generation contracts and a staged workflow, it allows users to specify the report's granularity and multimodal composition while maintaining evidence traceability. This is particularly important for research scenarios that require rigorous citations.

However, the current results are based on the specific base model Qwen3.6-35B-A3B, and the generalizability of MIRA-Harness and MIRA-Data still needs to be verified on more models. In addition, the formulation of generation contracts may currently rely on manual effort or fixed templates; how to enable the agent to automatically understand and refine user intent remains a direction worth exploring.

Overall, MIRA provides a controllable report generation framework for multimodal deep research agents, and its staged design and evidence traceability mechanism offer useful ideas for future research.

Background Supplement

Generation contract is the core abstraction proposed by MIRA, which can be understood as a set of output constraints explicitly agreed upon before generation and continuously validated during generation. It typically includes dimensions such as report granularity (e.g., number of paragraphs, level of detail per section), multimodal composition ratio (how text and images interleave), and evidence grounding requirements (each claim must be accompanied by a source link or image citation). Unlike ordinary prompts, the contract is repeatedly referenced at multiple stages of the workflow: used for self-checking when submitting the first draft, used to judge whether feedback has been addressed during revision, and used for acceptance at final delivery. This mechanism transforms vague user preferences into executable, verifiable specifications, thereby achieving fine-grained control while maintaining agent autonomy and providing clear standards for automated evaluation.

One-line Summary

MIRA uses generation contracts to connect a four-stage workflow, enabling controllable and traceable multimodal image-text deep research report generation.

Paragraph Summary

MIRA is a multimodal image-text deep research agent oriented toward controllable report generation. To address the problems of existing systems underutilizing image evidence during reasoning and lacking granularity and multimodal control during generation, MIRA uses MIRA-Harness to translate user requirements into "generation contracts" and connects four stages: evidence collection, report submission, feedback-driven revision, and final delivery. The research team also built MIRA-Data to generate high-quality multi-hop multimodal research trajectories for post-training the base model to adapt to the workflow. Based on Qwen3.6-35B-A3B, MIRA achieves SOTA performance among open-source agents on 7 multimodal search and deep research benchmarks, and the generated reports align with user requirements, with both textual and visual evidence traceable.