- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper studies whether SVG reinforcement learning benefits from constraining changes in rendering behavior rather than relying only on token-space constraints. It combines a frozen SVG-finetuned LLM encoder, a variational representation trained to reconstruct DINOv2 rendering features, and a metric adapter trained using equivalent and corrupted SVG pairs. The resulting behavior-distance penalty complements token-KL regularization, with a posterior-based adaptive radius. Validation includes representation probes, SVG generation/refinement benchmarks, and ablations against reward patching, a DINOv2 render anchor, and a non-variational metric. The method improves reported generation quality and collapse-detection rates, although generalization to held-out transformation families is weaker.
Critical strengths
- Useful empirical analysis of representation geometry. The probe-pair experiments demonstrate that raw frozen LLM representations can place rendering-equivalent rewrites farther apart than visually disruptive edits. This provides useful evidence that representations used for behavioral constraints require explicit validation, rather than assuming task-finetuned hidden states already provide an appropriate distance metric.
- Meaningful controls and improvements beyond the proposed failure detectors. The paper includes reward-patching, DINOv2 render-anchor, and non-variational controls, with three-seed results. It also reports improvements in established generation metrics: for example, FID improves from 11.43 for token-KL alone to 9.37 for Full. These results provide evidence of practical benefits beyond merely reducing the paper’s own collapse statistics.
Critical weaknesses
1. The central reward-hacking claims are not sufficiently disentangled from the reward design and detector definitions.
Length collapse is defined solely by , without requiring any accompanying visual degradation. This definition may therefore conflate legitimate compression with an actual failure mode. Moreover, is constant once the generated length falls below this threshold, so continued shortening cannot be explained simply by progressively increasing efficiency reward. The use of GT length as the reference is also sensitive to the serialization of the ground-truth code and requires paired code-length information to be available during training.
For viewBox collapse, reproducible rollout examples and more precise rendering/preprocessing details are needed to establish the proposed exploit mechanism. In addition, the Fixed-reward control changes both the rendering setup and the length penalty, making it difficult to attribute the observed effect to either factor individually. The manuscript also does not report a matched ablation that removes while keeping the remaining setup fixed, nor does it quantify the incremental benefit of Full under the patched setting.
As a result, the generality of these failure modes and the causal explanation offered for them remain unclear.
2. The necessity of the additional representation-learning machinery is not fully established.
Render-anchor directly uses DINOv2 features, whereas Full additionally introduces a frozen code encoder together with trained representation- and metric-learning components. This comparison changes both the representation source and the amount of task-specific training, so it does not isolate whether the gains specifically require code-side encoding, or whether similar improvements could be obtained by calibrating the visual representation using the same supervision.
The existing -only control is informative and already performs close to Full. However, the incremental contributions of the VAE, posterior uncertainty, and adaptive radius would benefit from more tightly matched ablations.
This is particularly important because Full introduces additional training complexity and also incurs higher rollout-time cost than the simpler render-anchor baseline.
Questions for authors
1. Clarification and validation of the failure detectors.
Among the outputs classified as length collapse, what fraction actually exhibit loss of necessary visual content, as opposed to being valid but more compact programs? Please provide an independent visual validation of the detector, together with sensitivity analyses over the collapse threshold.
For viewBox collapse, please provide representative RL rollouts including the generated SVG code, rasterization dimensions, target-image preprocessing, rendered outputs, and individual reward components. In particular, does the failure mode persist when the evaluation resolution is fixed independently of the generated SVG content?
2. Evidence needed to establish benefits beyond the specific reward design.
What happens when only is removed, while all other reward weights and training settings are kept unchanged? Does Full continue to improve output quality under both this setting and the patched reward/rendering setup? Please also provide the exact formulation of the patched length reward.
These comparisons are important for determining whether the proposed method addresses a broader optimization issue or primarily compensates for failure modes induced by the particular reward configuration used in the current experiments.
3. Clarification of the GT-length assumption.
Why is one-half of the length of a single GT program an appropriate efficiency threshold, particularly in Text-to-SVG, where multiple programs of substantially different lengths can produce equally valid outputs? Are the reward values and collapse labels stable if the GT program is replaced by a rendering-equivalent but compressed or expanded version?
Please also clarify how the proposed training setup applies when paired GT SVG code is unavailable.
4. Evidence needed to isolate the representation contribution.
For Render-anchor, please specify the exact distance function, reference construction, penalty coefficient, and radius-calibration procedure. How does a DINOv2-based metric head trained with the same pairwise supervision compare?
Given the strong performance of the -only variant, please clarify which matched experiments isolate the incremental contributions of the VAE and the uncertainty-based radius. It would also be helpful to report whether these incremental gains are statistically reliable across runs.