Training an agent to drive heterogeneous visual experts outperformed general-purpose models on reasoning segmentation, dense detection and referring counting
AspireAI community · arXiv
Vision-language experts built for detection, segmentation and counting are usually optimized in isolation, so a general multimodal model cannot easily borrow their precision when a task needs both dense small-object perception and real language understanding. VisHarness trains a visual agent that separates high-level perception, reasoning and decision-making from low-level task execution, so it learns a generalizable policy for driving a set of heterogeneous visual experts rather than training a task-specific model of its own. The authors add dynamic visual memory archiving so the visual-token overhead that accumulates during multi-turn expert interaction does not swamp the context, which is what makes on-policy reinforcement learning in a live environment practical. On four benchmarks spanning reasoning segmentation, generalized referring expression segmentation, dense small-object detection and referring expression counting, VisHarness substantially outperformed existing general-purpose models and was competitive with or better than task-specific ones, and the code, weights and training data are public. The limitation is structural: the approach only helps where a set of good specialist visual experts already exists, so it trades a single general model for a portfolio of models plus a trained router.
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.