An arXiv preprint reports that DARS, a reinforcement-learning method for instruction-based image editing, achieved the top reported score in all five benchmark regimes. Against a matched Joint RL + Adaptive Curriculum baseline, it scored higher on every benchmark. The supplied analysis presents these as reported point scores, without confidence intervals or estimates of variation across repeated training seeds.
A planner with a clearer job
DARS replaces a free-form planner trace with four fields: Modify, Preserve, Overall and Tips. It then assigns rewards to those slots and reweights the learning signal for individual tokens.
A second layer separates reward variation between different plans from variation among renderings of the same plan. DARS uses those signals for soft routing between the planner and renderer, while a reward-based hardness estimate helps schedule an adaptive curriculum.
Training used 10,000 examples: 5,000 from the RL split of THINKEDIT-140K and 5,000 from UniREdit-Data-100K. Each input used four plans and four renderings per plan, for 16 rollouts.
The controlled baseline matched DARS on its backbone, data, reward model, rollout budget and adaptive curriculum. It used free-form reasoning instead of DARS's structured planner and did not use the method's dual-level credit assignment.
The score gap was uneven
DARS reported scores of 80.72 on KRIS-Bench, 27.50 on RISE-Bench, 4.39 on ImgEdit-Bench and 7.86 on GEdit-Bench-EN. On PICA-Bench, it scored 64.19 with simple prompts and 72.75 with detailed prompts.
Compared with the matched baseline, the differences were +8.57 points on KRIS-Bench, +1.80 on RISE-Bench, +0.19 on ImgEdit-Bench, +0.03 on GEdit-Bench-EN, and +0.64 and +0.53 points on PICA-Bench's simple and detailed settings. The reported ranking is therefore a benchmark comparison, not a causal estimate of superiority.
What changed inside the system
In the RISE-Bench/GEdit-Bench-EN ablation, Adaptive Curriculum scored 27.50/7.86, compared with 24.56/7.15 for Static Curriculum and 18.89/6.87 for No Curriculum. Soft Routing also scored 27.50/7.86, ahead of Hard Routing at 24.86/7.49 and No Routing at 20.05/7.12.
The Full Structured Answer scored 27.50/7.86, versus 23.94/6.98 for Free-form Reasoning. Prefix-Gated reward also scored 27.50/7.86, compared with 24.92/7.28 for Flat Average and 26.31/7.57 for Weighted Sum.
A reported stability comparison found that free-form joint RL without a curriculum diverged. The structured planner remained trainable without a curriculum, although its performance was worse.
The routing signals were compared with majority-voted GPT-5 pseudo-labels, which are model-generated reference labels rather than ground-truth failure labels. AUROC, a measure of how well a signal separates two groups, was 0.930 for Planner versus Renderer, 0.918 for Planner versus Mixed, and 0.923 for Renderer versus Mixed. The corresponding accuracies were 86.0%, 84.7% and 85.3%, while Macro-F1 scores were 85.8%, 84.1% and 84.9%.
A costly result with open questions
DARS requires M × K rollouts for each training example, making it more expensive than single-path updates. Its variance-based routing and post-render planner scores also depend on the reward model, and failures are harder to disentangle when the planner and renderer fail at the same time.
The supplied document is an arXiv preprint dated 20 Aug 2026. The work was supported by KlingAI Research.
Paper data and sources
Original title: DARS: Dual-Level Credit Assignment RL with Structured Reasoning for Instruction-Based Image Editing
Authors: Haoxiang Cao, Jiajiong Cao, Xuanpu Zhang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text