Frontier coding agents found 98.2 to 99.9 percent of seeded bugs from source code alone, and a debugger only helped when the agent was pushed to use it
AspireAI community · arXiv
This paper asks a practical question for anyone running a coding agent: does it still need a debugger? The authors seeded 610 single-line faults into 144 LeetCodeDataset solutions and rewrote each faulty program with five behavior-preserving transformation recipes, then had agents based on Sol 6.1, Luna 6 and Qwen 3.8 localize each fault with no, optional or encouraged debugger assistance.
Without tools, the agents localized 98.2 to 99.9 percent of faults in original programs and 96.7 to 99.9 percent in transformed ones, losing accuracy mainly where recipes restructured loops and conditions. Encouraged debugger use removed 63 percent and 61 percent of Luna 6's and Qwen 3.8's remaining transformed errors, adding 2.1 and 1.3 percentage points, but at 14.7 to 27.0 times the token use, and Sol 6.1 stayed at 99.9 percent. With debugger assistance merely optional, Luna 6 and Qwen 3.8 still gained 1.5 and 1.3 points, but the agents actually used the interactive debugger in only 0.2 to 6.2 percent of runs.
The useful read for practitioners is the last one: given a choice, agents executed the program far more often than they stepped through it, and still kept most of the benefit. The honest caveat is scope. These are short, precisely specified programs with single-line faults, so the near-ceiling result says nothing about multi-file, architecture-level bugs, and every number comes from the authors' own seeded-fault protocol.
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.