RobotAgentBench evaluates open-world robot autonomy across 50 household layouts, 90 atomic tasks and 8 long-horizon tasks, scored on discovery, reasoning, manipulation, recovery and robustness
AspireAI community · arXiv
RobotAgentBench, submitted to arXiv on October 26, 2026, argues that existing robotics benchmarks test isolated capabilities while real open-world deployment needs task discovery, contextual reasoning, physical execution and failure recovery in one loop. The benchmark builds 50 household layouts with interactive objects and physically grounded tasks through an agent-assisted, human-in-the-loop real-to-sim pipeline, spanning 90 atomic tasks, 18 short-horizon tasks, 8 long-horizon tasks and 4 open-ended scenarios. Evaluation runs at two levels: task-level scoring across five dimensions — task discovery, context reasoning, manipulation, failure recovery and robustness — and scenario-level scoring for whether an agent integrates them to fulfil a high-level need. The authors report it as enabling systematic comparison of standalone vision-language-action and world-action models against diverse robot-agent architectures under a unified planning-execution-verification framework. The honest limitation: this is a benchmark release rather than a model result, so it measures a starting point rather than a capability gain, and the benchmark is the authors' own construction, so absolute scores are not comparable to any other benchmark.
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.