A progress critic scored robot manipulation about 65 percent above GPT-6-Sol on terminal success
AspireAI community · arXiv (Xu, Zheng, Zhu, Ren, Fu, Wang & Zhong)
Robotic manipulation needs dense, quantitative progress feedback for data curation, policy evaluation and reinforcement learning, and the usual source is sparse keyframe labels made by hand. NimCritic replaces that with action-aware quantification: a vision-language model decomposes long-horizon trajectories into object-level subtasks and assigns weighted action credits, giving the model both a progress signal and a record of what was executed.
The supporting data is an 800-plus hour training video set, and LOOM, a long-horizon progress benchmark with 19,618 human-calibrated evaluation frames. The critic itself is trained by Joint Video-Progress Learning, with Historical Progress Perturbation to counter sequential drift and a contrastive Video-Progress Alignment objective that grounds the visual dynamics against the denoised progress trajectory. It matches or beats state-of-the-art process reward models on the Video Progress Benchmark and LOOM, and on VPB it surpasses GVL (GPT-6-Sol) by roughly 65 percent in terminal success accuracy. On RoboArena it retrieves human-selected successful rollouts at about 97 percent overall recall.
The caveat is that the comparison is against process reward models on manipulation benchmarks, and the labelling pipeline depends on a vision-language model doing the subtask decomposition.
https://arxiv.org/abs/2610.32776
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.