Small-model reasoning kept producing the right answer and then failing to select it, and an interface layer lifted GPQA Diamond accuracy from 5.6 to 19.2 percent
AspireAI community · arXiv preprint
Multi-candidate reasoning - generate several answers, then pick one - is the standard test-time scaling route for small models that cannot be upgraded. The paper identifies a bottleneck in the gap between those two steps that is easy to miss: a correct answer may be sitting in the candidate pool and still never get picked, because it is incomplete, unparseable, or in a different answer mode from what the fixed selector expects. The authors call this the hidden interface tax and formalize it as a coverage-conversion decomposition that separates raw oracle coverage from selectable coverage.
CACC, an interface layer that repairs fragments, aligns answer modes and strips selector-confusing noise without retraining either side, lifted GPQA Diamond accuracy from 5.6% to 19.2% (13.6 points), and to 25.8% (20.2 points) when combined with proposer-side strengthening. The honest counterweight is in the paper: on MMLU-Pro, raising oracle coverage this way actually lowered final accuracy by 0.85 points under the original selector - more correct candidates is not automatically more selected ones.
Caveat: the gain depends on the task regime, and the MMLU-Pro result is a regression the authors report rather than hide.
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.