An arXiv preprint reports higher clustering scores for a method called CloSeR in comparisons with SimGCD and SelEx. The reported differences were larger on fine-grained image benchmarks. The document is version 1, dated 26 Aug 2026.
The work addresses generalized category discovery, or GCD, where a model must deal with known and novel categories. Its central question is whether learning known-class recognition first and then transferring that closed-set knowledge can improve the task.
A two-stage route to category discovery
CloSeR uses two linked stages. In closed-set transfer learning, it inserts lightweight, block-wise adapters into a frozen foundation model and learns closed-set prototypes from labelled data only. During GCD training, its unified relational distillation, or URD, transfers global sample-to-prototype relations and local sample-to-sample relations through separate feature pathways. The approach supports both parametric and non-parametric GCD methods.
How the comparisons were scored
The evaluation used DINO and DINOv2 backbones on six image benchmarks: CIFAR-10, CIFAR-100, ImageNet-100, CUB, Stanford-Cars and FGVC-Aircraft. The reported metric was clustering accuracy, or ACC. Results were separated into All, Old and New categories, with averages provided across datasets; Old and New denote the known and novel parts of the task.
The biggest gaps came on fine-grained images
On the generic comparisons, DINO SimGCD's average All accuracy rose from 86.7 without CloSeR to 88.8 with it, a reported increase of 2.1 points. For SelEx, the corresponding scores were 87.1 and 89.0, a 1.9-point increase. With DINOv2, SimGCD with CloSeR recorded average All and New scores of 94.9 and 94.6, while the SelEx comparison went from 92.4 to 93.7.
The DINO fine-grained results included larger reported differences. On Cars, SimGCD's All score went from 53.8 without CloSeR to 63.0 with it, a 9.2-point increase. Its average All score across the fine-grained datasets rose from 56.1 to 62.6, described as an 11.6 percent relative increase. SelEx's average New score climbed from 58.8 to 69.4, an 18.0 percent relative increase, while its average All score moved from 63.1 to 72.0.
With DINOv2, SimGCD's average New score rose from 65.3 without CloSeR to 74.5 with it, a 9.2 percentage-point increase and a 14.1 percent relative increase. SelEx with CloSeR recorded an average All score of 85.0. On CUB New, the SelEx comparison went from 88.5 to 91.5.
Those figures are reported benchmark comparisons. The supplied results do not include confidence intervals or run-to-run variability, so the exact stability of the score differences is not established by the report.
Tests of the design
The paper also tested how the DINO model was adapted. In one transfer-strategy comparison, full fine-tuning scored 60.9 on Cars All, versus 46.7 when only the last block was fine-tuned. On Aircraft All, the figures were 54.8 and 45.8. Block-wise adapters reached 63.0 on Cars All, 56.6 on Aircraft All and 54.7 on Cars New, and were reported as the best overall transfer strategy.
Another ablation described global and local relations in URD as complementary. With feature decoupling, All accuracy was 82.8 on CUB versus 81.0 without it, 66.5 versus 65.8 on Cars, and 66.6 versus 64.9 on Aircraft.
The scaling test produced different reported patterns for SimGCD and CloSeR. Under DINOv2 SimGCD, scaling the student from ViT-B to ViT-L was accompanied by CUB All scores of 71.5 and 71.0, and Cars All scores of 71.5 and 71.3. In the CloSeR comparison with larger teacher and student models, Cars went from 79.9 to 83.8, CUB from 81.3 to 81.4 and Aircraft from 75.0 to 76.2.
What the benchmark results leave open
Attention maps offered a qualitative view of the comparisons. The authors judged that CloSeR focused on semantically diagnostic object parts and covered multiple parts more fully, while the baseline showed diffuse attention and background leakage, particularly for unseen categories. The comparison was descriptive, not a numeric effect estimate or a validated mechanistic measure.
The reported evaluation covered the six named image benchmarks and the DINO and DINOv2 backbones, along with transfer, relational-distillation and scaling ablations and qualitative attention visualizations. The implementation section says that additional dataset, optimization and hyperparameter details are provided in supplementary material.
Paper data and sources
Original title: CloSeR: Unified Relational Distillation from Closed-Set Teachers for Category Discovery
Authors: Yuanpei Liu, Zhenqi He, Jialu Tang, Kai Han
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text