The paper introduces Mr.LHDR, a benchmark for evaluating agents on long, interdependent chains of evidence. Each question requires an average of 12.1 intermediate conclusions with a dependency depth of 10.4, and includes multimodal inputs such as images, maps, and charts. The benchmark scores both final answers and intermediate conclusions.
Evaluating state-of-the-art systems, the best performer (GPT-5.5) achieved only 43.1% overall accuracy and 34.3% strict accuracy. Removing images reduced the dependency-aware checklist score by 12.6 points, showing the importance of non-text evidence. Strict accuracy also declined as reasoning chains lengthened.
These results identify sustained, dependency-consistent evidence integration rather than isolated fact retrieval as the main bottleneck for current deep research agents. The benchmark reveals that final-answer accuracy overestimates complete research success, highlighting the need for agents that maintain correct intermediate conclusions across long, multimodal chains.