UnityBench-Mini3D tested 21 models on 77 project-level Unity tasks and the average pass rate fell from 80.9 percent on easy tasks to 35.2 percent on hard ones
AspireAI community · arXiv (UnityBench-Mini3D authors)
*Game development stresses coding agents in ways general software engineering does not: correctness depends on how engine APIs, serialized scene state and runtime behaviour interact, not on source code alone, and existing benchmarks lean heavily on LLM judges or coarse task-level outcomes.* UnityBench-Mini3D is a benchmark of 77 project-level Unity tasks organised around seven expert-defined skill categories rather than tutorials or genres, across three difficulty tiers. Each task pairs a complete Unity project with natural-language requirements and is evaluated with 1,178 held-out, human-authored milestone tests covering scene structure, component configuration and runtime behaviour.
The in-engine evaluation gives deterministic partial credit, localises failures and separates partial implementations from complete solutions without a subjective judge. Across 21 models and 9 coding harnesses, the average pass rate reaches 80.9 percent on easy tasks but falls to 35.2 percent on hard tasks, with 57.7 percent overall - a gap that is invisible to any single aggregate number. The authors position the benchmark as a reproducible, diagnostic testbed for distinguishing systems across difficulty levels rather than a single headline leaderboard.
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.