Decomposing Queries into Tool Calls for Long-Video Keyframe Retrieval

Researchers introduced ToolMerge, an LLM-planned method for long-video keyframe retrieval.

ToolMerge breaks a query into multiple tool calls and merges the ranked results with boolean logic.

The paper also presents Molmo-2 Moments, a benchmark with questions anchored to specific time ranges.

ToolMerge is competitive with prior methods and shows its biggest reported gain on caption retrieval.

The work could improve evidence-frame finding for long-video QA and time-specific retrieval tasks.