ICSE 2027: LLM-refined code ran faster but burned up to 886% more energy than the human version
AspireAI community · arXiv (University of Augsburg, King's College London)
Using LLMs to optimize code for speed quietly assumes that faster execution means lower energy use. A study from the University of Augsburg and King's College London tested that assumption across 11 open-source models and four decoding configurations on all 1,000 problems of the EffiBench performance benchmark, collecting per-process hardware-level energy measurements throughout. Inference energy spanned 0.79 to 8.68 MJ across model-variant pairs, roughly equivalent to the energy of 250 to 2,800 GPT-4.1 queries, and capping output tokens had the largest effect, cutting energy use by up to 69 percent against the default configuration. On the execution side, 42 of 44 configurations produced code consuming more energy per problem than the canonical human reference, with overheads from +4 percent to +886 percent. The two exceptions gained only marginally, with break-even points on the order of 10^6 to 10^7 executions. Mixture-of-Experts architectures consumed substantially less inference energy than equivalently sized dense models without giving up correctness, which the authors argue makes architecture a more consequential sustainability lever than decoding parameters. The honest caveat: the measurements are per-process energy on the benchmark's problems, and the break-even framing depends on how many times the optimized code actually runs in production.
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.