SceneEvolve evolved the harness around a 27B model instead of the model itself, and beat GPT-5.6-Sol at 3D reconstruction by 32.3 percent while cutting deployment cost 59.8 percent
AspireAI community · arXiv
SceneEvolve represents 3D scenes as executable code and improves reconstruction by evolving the harness around a small open-source agent rather than training the agent harder. It pairs SceneAct, a compositional semantic action interface over parameterized primitives with hierarchical composition and spatial relations, with SceneHarness-RSI, where a stronger teacher agent iteratively evolves the harness using feedback from development-set reconstructions. Throughout, model parameters, action semantics and the execution backend stay fixed.
The payoff is that the evolved harness becomes a reusable procedure the small model can apply to unseen scenes without parameter updates or teacher help. With Qwen3.8-27B it improved all four evaluated reconstruction metrics across two in-domain datasets and one out-of-domain dataset, cut mean inference time 47.0 percent against the base model, achieved 32.3 percent higher reconstruction quality than the much larger proprietary GPT-5.6-Sol, and cut mean equivalent deployment cost by 59.8 percent. The harness also generalized to Kimi-K3, reaching reconstruction performance comparable to GPT-6-Astra.
The honest limitation is that the comparison depends on a stronger teacher being available during harness evolution, even if it is not needed at inference time. The results are also the authors' own measurements on their own datasets, and the headline claim of beating a frontier model by 32.3 percent comes with a specific baseline configuration.
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.