A research team has built a multimodal search agent that outperforms open-source rivals by structuring training, environment, and rewards around a single shared knowledge graph.
The system, called SearchEyes, targets a specific failure mode in how AI search agents are trained: the pipeline components — training data, search environment, and reward signals — are typically built independently, which means useful structural metadata gets thrown away and rewards only land at the end of a reasoning chain rather than at each step. SearchEyes replaces that patchwork with a simulated search world built on a typed knowledge graph. Two new techniques power it: Perception-Knowledge Chains, which sample multi-hop reasoning paths across a visual-knowledge dataset derived from Wikidata5M, and Hop-Anchored Policy Optimization, which uses the metadata from each hop to assign credit at the step level without training a separate reward model. Tested across six multimodal knowledge-intensive benchmarks, the 27-billion-parameter version improved on the strongest open-source baseline by 6.2 points on average.
Most AI search research either handles text or images well — not both, and rarely across multiple reasoning steps. SearchEyes addresses the multi-hop gap directly, which matters as AI assistants are increasingly expected to answer questions that require chaining together visual and factual evidence. Tying the reward signal to intermediate steps rather than final answers is a meaningful architectural choice: it sidesteps the sparse-reward problem that has slowed reinforcement learning in complex reasoning tasks.
The results are on academic benchmarks, and benchmark leads in AI research have a habit of not surviving contact with real-world search traffic — worth watching whether this holds up outside the lab.