MIRA: A Multimodal Illustrated Deep Research Agent for Controllable Reports

Bar chart of scores on eight benchmarks. MIRA-SFT scores 80.7 on MMSearch, 38.9 on MMSearch-Plus, 58.4 on BrowseComp-VL, 79.0 on SimpleVQA, 82.0 on LiveVQA, 79.3 on FVQA, 29.8 on VDR-Bench and 42.6 on MMDeepResearch, compared with Qwen3.6-35B-A3B, WeAgent-MMSearch, OpenSearch-VL, DeepSeek-V4.1-Flash and Kimi-K3. — MIRA-SFT versus open agents and frontier models on eight benchmarks.

Overview

Deep research agents can search the open web, but they remain poorly suited to delivering what users often need: an illustrated report with the requested depth, images, and traceable sources. MIRA turns these requirements into generation contracts. MIRA-Harness gathers textual and visual evidence, checks each submitted draft against the contract, and revises until it is satisfied; MIRA-Data turns this workflow into training trajectories. Built on Qwen3.6-35B-A3B, MIRA-SFT is the strongest of the compared open agents on seven of eight benchmarks.

Highlights

Controllable reports. Users set report depth, visual composition, and evidence grounding.
Harness + data. Contracts drive research, writing, and revision; the same workflow produces training trajectories.
Strong open agent. Best among compared open agents on 7 of 8 benchmarks.

From Search Answers to Illustrated Reports

One multimodal request about CAM photosynthesis answered three ways: a traditional short answer, a MIRA Concise report and a MIRA Detail report with more images and sources. — One request, two controllable deliverables: Concise and Detail reports.

MIRA-Harness

MIRA-Data

Trajectory synthesis: a query generator builds report requests and rubrics, a teacher model runs inside MIRA-Harness, and a synthesis gate checks rubric coverage and trajectory integrity before trajectories join MIRA-Data. — Trajectories synthesized inside MIRA-Harness, verified by a gate, then mixed with general search data.

Main Results

Model	Params	MMSearch	MMSearch-Plus	BrowseComp-VL	SimpleVQA	LiveVQA	FVQA	VDR-Bench	MMDeepResearch
Large-Scale Frontier Models
Kimi-K3	2.8T (104B)	89.26	60.00	60.32	83.33	91.23	82.46	22.86	54.32
Qwen3.5	397B (17B)	69.01	36.01	58.40	77.33	84.33	75.33	14.40	39.54
DeepSeek-V4.1-Flash	552B (8B/16B)	88.33	59.91	72.86	81.33	88.67	85.00	32.20	51.23
Lightweight Multimodal Search Agents
OpenSearch-VL	30B (3B)	68.70	28.94	41.10	74.90	67.40	73.20	33.50	23.21
REDSearcher-MM	30B (3B)	72.90	26.60	57.20	–	79.30	–	–	–
Vision-DeepResearch	30B (3B)	69.60	28.50	53.70	66.33	77.60	74.20	37.80	12.78
WeAgent-MMSearch	30B (3B)	77.78	37.82	–	77.73	81.05	79.33	35.68	–
Qwen3.6 w/o Harness	35B (3B)	30.41	3.86	28.91	58.00	56.67	34.11	1.80	26.78
Qwen3.6 w/ WeAgent Harness	35B (3B)	66.28	29.40	–	65.79	78.20	66.09	15.57	–
Qwen3.6 w/ Vanilla Harness	35B (3B)	67.84	28.94	51.01	72.24	73.67	65.67	15.00	37.95
Ours
MIRA-Base	35B (3B)	73.10	33.44	56.64	74.33	78.33	70.00	19.00	41.90
Δ (MIRA-Base − Vanilla)		+5.26	+4.50	+5.63	+2.09	+4.66	+4.33	+4.00	+3.95
MIRA-SFT	35B (3B)	80.70	38.91	58.40	79.00	82.00	79.33	29.80	42.61
Δ (MIRA-SFT − MIRA-Base)		+7.60	+5.47	+1.76	+4.67	+3.67	+9.33	+10.80	+0.71

Bold: best outside the large-scale frontier models. “–”: not reported.

Report Examples

Two requests, each delivered in both modes. Reports are shown in full, with the paper's annotations, numbered images, and image sources.

Swipe or use ← →; click a page to open it full size.

BibTeX

@article{mira2026,
  title   = {MIRA: A Multimodal Illustrated Deep Research Agent for Controllable Reports},
  author  = {Wu, Hongrui and Tuo, Kaiwen and Wang, Congcong and Li, Gen and Dong, Shuai and
             Yu, Xinlei and Xu, Yuqi and Hou, Haowen and Peng, Zelin and Wang, Jiaqi},
  year    = {2026}
}