A query split into four jobs
MulVec, a training-free image-retrieval system, led the reported results on CIRCO, CIRR and FashionIQ, according to a version-1 arXiv preprint. Its central move is to treat separate parts of a request as different jobs: define the full target, add content, preserve content and remove content.
MulVec compiles each query into four named roles. Global represents the full target; Desired identifies content to add; Preserve marks content to keep; and Forbidden identifies content to remove. Frozen encoders then produce a target vector and role-specific probe vectors, alongside global and local candidate evidence. A fixed weighted sum ranks the gallery in one pass, without task-specific parameter learning.
The lead held across all three tests
On CIRCO, MulVec was highest at every reported cutoff and backbone scale. The evaluation used frozen OpenCLIP towers at ViT-B/32, ViT-L/14 and ViT-G/14. At mAP@5, its margin over the best available comparator was reported as 23.0% with ViT-B, 22.4% with ViT-L and 19.7% with ViT-G.
On CIRR, MulVec led all reported full-gallery and subset recall measures at every backbone scale in the selected comparison. FashionIQ showed the same pattern: MulVec led the category-specific and averaged R@10 and R@50 metrics at every scale, topping all 24 reported metrics.
Matched controls separated the effects
The central matched test asked whether explicit role handling provided a gain beyond simply increasing the number of candidate vectors. In matched ViT-L/14 controls, role-aware dispatch had higher reported metrics than a shared read under both Global-only evidence and Global-plus-local candidate evidence. The paired gains were 7.08 and 5.58 points on CIRCO, 5.26 and 5.50 on CIRR, and 16.46 and 12.36 on FashionIQ. In each pair, the first number is the Global-only gain and the second is the Global-plus-local gain.
Adding local visual vectors raised the shared-read metric by 4.88 points on CIRCO, 2.63 on CIRR and 6.72 on FashionIQ. For role-aware matching, the corresponding gains were 3.38, 2.87 and 2.62 points. The local-vector comparison changed both reading strategies, with the size of the change differing by benchmark.
Component tests showed the largest single loss when Global was removed. CIRCO mAP@5 fell 15.15 points, CIRR Recall@1 fell 13.64 points and FashionIQ macro Recall@10 fell 14.34 points. Other operator changes were also unfavorable: role reassignment, removing reference continuity and changing Forbidden into a reward each lowered all three reported validation metrics relative to full role-aware matching.
The boundary was visible in the examples
Illustrative controls showed both a rescue and a hard miss. In one CIRCO control, role-aware dispatch moved the best-positive rank, the highest-ranked labeled match, from 5 to 2 with Global-only evidence and from 4 to 1 with Global-plus-local evidence. In another control, it missed all labeled positives within the displayed Top-50. Role-aware matching improved one illustrative retrieval but did not recover the other within that displayed range.
Sensitivity checks found the best tested temperature at τ = 0.02. At 0.01 and 0.04, reported scores were lower by 0.70 and 0.52 points on CIRCO, 0.96 and 0.24 on CIRR, and 0.62 and 0.28 on FashionIQ, respectively. Across 16 non-anchor role-weight settings, maximum drops were 1.56 points on CIRCO, 0.45 on CIRR and 0.55 on FashionIQ. Within the tested settings, those role-weight changes were modest.
The evaluation covers CIRCO, CIRR and FashionIQ with frozen OpenCLIP towers at ViT-B/32, ViT-L/14 and ViT-G/14, but dataset sizes and query counts are not reported. MulVec's ranking uses a fixed weighted sum in one pass, and the illustrative Top-50 miss shows that role-aware matching did not recover all labeled positives in that case.
The document is an arXiv version-1 preprint dated 26 Aug 2026. Its supplied front matter lists author and affiliation information but includes no funding statement.
Paper data and sources
Original title: MulVec: Fine-Grained Role-Aware Matching for Training-Free Zero-Shot Composed Image Retrieval
Authors: Zihao Zhang, Dayan Wu, Xinze Liu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text