Turning on reasoning lifted Qwen3-32B's HumanEval Pro pass@1 from 71.1 to 76.5 percent, but keeping only the first half of the trace cost nothing in accuracy
AspireAI community · arXiv
Benchmarks for code generation usually report correctness and ignore what the reasoning trace cost, so it is hard to tell whether longer thinking actually earned its tokens. This study runs 866 Python tasks from HumanEval Pro, MBPP Pro and BugsInPy under five model settings and measures the accuracy gain against the token overhead. Turning reasoning on in Qwen3-32B over five runs raised mean pass@1 on HumanEval Pro from 71.1 to 76.5 percent and mean tests pass on BugsInPy from 3.5 to 7.9 percent - but on MBPP Pro the move from 68.1 to 69.5 percent was not statistically significant, even though the model burned a median of over 2,000 thinking tokens per task. Across models, a longer median trace did not reliably mean higher accuracy. The most actionable finding is the truncation experiment: they cut stored traces offline and asked the same model with reasoning off to solve the task from what remained, and retaining the first half of a trace produced no statistically significant drop in correctness. Stopping when a trace became repetitive seldom triggered at all, and stopping as soon as solution code appeared helped some models and harmed others - plain fixed truncation beat both rules. The authors' conclusion is the practical one: more reasoning is not always better, and studies should report reasoning cost alongside correctness.
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.