AI Paper Atlas

Research map / Label-efficient and robust learning / Test-time adaptation

Duplex Rewards Optimization for Test-Time Composed Image Retrieval

Haoliang Zhou, Feifei Zhang, Changsheng Xu

AAAI 2026 · Label-efficient and robust learning › Test-time adaptation · approach: Test-time adaptation of vision-language models · 1 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace CMS with top-k sampling (ablation)TT-RLDR · FashionIQ avg / CIRRR@10 / R@1: 26.62 -> 24.89 (FashionIQ R@10 without CMS)Improved
use only reference rewards (ablation)TT-RLDR · FashionIQ avgR@10: 25.67 -> 26.62 (adding endogenous rewards)Improved
use only endogenous rewards (ablation)TT-RLDR · FashionIQ avgR@10: 25.67 -> 23.42 (ER-only vs RR-only)Hurt
use larger reference reward model (ablation)TT-RLDR · FashionIQ avgR@10: 26.62 -> 28.26 (C-G14 reward)Improved
use lighter reward model CLIP-B16 (ablation)TT-RLDR · FashionIQ avgR@10: 26.62 -> 25.15Hurt
use BLIP2 or ALBEF backbone (ablation)TT-RLDR · FashionIQ avgR@10: 26.62 -> 22.34 (BLIP2), 20.48 (ALBEF)Hurt
use CLIP-L/14 backbone instead of B/16 (ablation)TT-RLDR · CIRRR@10 / R@50: 71.98 -> 72.07 R@10; 91.89 -> 91.74 R@50No clear effect
TT-RLDR (C-B16) versus other TTA methodsTT-RLDR · CIRRR@1: 24.54 -> 27.91 (vs RLCF)Improved
TT-RLDR (C-B16) versus RLCF on COCOTT-RLDR · COCOR@10: 31.85 -> 34.01Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy