PerceptRL made visual perception rewardable and reached 71.6 percent average accuracy, beating the best frontier VLM it evaluated by 5.2 points
AspireAI community · arXiv (PerceptRL authors)
*Fine-grained visual perception is a bottleneck in multimodal post-training: perceptual errors cap the gains that reasoning-focused optimization can deliver, and caption supervision fades as later stages change the data mixture and objective.* PerceptRL attacks perception at two levels. Perception RL uses a multi-agent pipeline to pull verified errors out of dense image captions and turn the corresponding visual facts into QA pairs with fixed answer labels and rule-based rewards, chaining correlated errors and mixing unrelated ones, with a specialized route for high-error domains and a domain-agnostic general route beyond them. Perception Agentic RL then handles fine-grained cases through multi-turn inspection with visual tools, with Coordinate Injection correcting spatial operations in cold-start trajectories and a grounding-aware process reward supervising intermediate spatial fidelity, not just the final answer.
Across seven benchmarks under a shared tool-use protocol, PerceptRL achieves 71.6 percent average accuracy - 15.7 points above the base model given the same tools and 5.2 points above the strongest evaluated frontier VLM. The authors report observing coordinate drift even in models trained to use tools, where wrong crops or annotations distort the evidence and propagate the error forward.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.