HyperEyes

HyperEyes is a parallel multimodal search agent developed by Xiaohongshu (rednote-hilab) and the University of Cambridge. It addresses the fundamental bottleneck of existing multimodal agents: the "N-run bottleneck", whereby tool calls are issued one entity at a time. For queries involving multiple independent entities — e.g., six people in a photograph — this produces cascading interaction rounds, growing token consumption, and compounding error risk.

The core innovation is UGS (Unified Grounded Search) — an atomic action that fuses visual grounding with multi-entity retrieval in a single step. Rather than "searching deeper" (more sequential rounds), HyperEyes "searches wider": one call covers all target entities concurrently.

Training proceeds in two stages: (1) cold-start via a Parallel-Amenable Data Synthesis Pipeline with Progressive Rejection Sampling, (2) Dual-Grained Efficiency-Aware Reinforcement Learning. TRACE (Tool-use Reference-Adaptive Cost Efficiency) operates at the trajectory level — rewarding correct answers with a penalty proportional to tool-call steps, with the efficiency threshold tightened monotonically across epochs. OPD (On-Policy Distillation) operates at the token level — activating exclusively on failed rollouts to inject dense corrective signals from a 235B teacher model via KL divergence.

The authors also introduced IMEB (Image Multi-Entity Benchmark) — 300 human-curated instances spanning humanities and science, each requiring simultaneous identification of multiple image entities. On IMEB, HyperEyes-30B outperforms the second-best model by 64.0%. The CAS (Companion Assessment Standard) metric measures useful information returns per unit compute; HyperEyes-30B achieves 7.6× better information efficiency than sequential models.