MIRA: A Multimodal Illustrated Deep Research Agent for Controllable Reports

Key points

MIRA-Harness turns user requirements into a generation contract that sets acceptance conditions on report length, image composition, and evidence grounding, and connects four stages: evidence collection, report submission, feedback-driven revision, and final delivery.
Enabling the report gate raised fully compliant reports from 0 out of 32 to 25 out of 32 and increased the proportion of cited images from 23.5% to 92.6%.
The authors built MIRA-Data from about 4.6K Harness-Native trajectories generated by Kimi K2.6 inside MIRA-Harness and verified by a synthesis gate, combined with 33K general search trajectories.
MIRA-SFT, trained on MIRA-Data, improved over MIRA-Base with the same backbone by 5.50 points on average across eight benchmarks, including 10.80 points on VDR-Bench and 9.33 points on FVQA.
MIRA-SFT achieved the best results among the compared open agents on 7 of 8 benchmarks and surpassed the much larger Qwen3.5-397B-A17B on six benchmarks.
As limitations, the authors note that transient API errors and parsing failures from external search infrastructure can interrupt evidence collection, and hallucinations may arise when the agent misreads retrieved content or makes weakly supported claims.

External research summaries. These are not HDATF publications or measured product results.