Training a multimodal model to connect visual evidence rather than to see more of it raised 27B accuracy to 60.5 percent, ahead of GPT-5.6-Luna, GLM-5.3-Flash and DeepSeek-V4.1-Flash
AspireAI community · arXiv
*Spatial perception and multimodal long-context understanding look like separate problems, but both reduce to finding the right visual facts, binding them to the right entities and composing them into an answer - and that capability turns out to be far cheaper to synthesise training data for than either task itself.* The EvidenceConnect framework generates questions by cropping related regions inside a single image and asking for comparison, aggregation or chaining across them, then filters the output with cross-model answer anchoring and drop-one verification, and the resulting Zooming without Zooming v2 models improve over their base model on all 25 benchmarks at every scale, with overall averages of 5.39, 4.96 and 4.99 points at 4B, 9B and 27B. The 27B model reaches 60.51 percent overall accuracy against 54.02 percent for GPT-5.6-Luna, 57.24 percent for GLM-5.3-Flash and 55.47 percent for DeepSeek-V4.1-Flash, and gains concentrate exactly where the framing predicts: 5.2 to 6.3 points on fine-grained and spatial perception, and 5.5 to 8.1 on embodied spatial reasoning. The data efficiency result is the more interesting one: EvidenceConnect reaches 53.86 percent average from 73K examples where the earlier ZwZ set reaches 49.86 from 20.4K, and it surpasses ZwZ on 10 of 11 benchmarks while trailing only on MileBench by 0.13 points, with the largest margin on MindCube-Tiny where ZwZ actually loses 13.8 points and EvidenceConnect gains 5.4. The limitation is that the comparison models were evaluated under the authors' own setup, and the multi-image limitation noted in the text-conditioned hashing line applies here too.
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.