Sealed Audit Compression Progress Shown to Resist Goodhart's Law in AI Reward Design
Researchers proved mathematically that rewarding an AI agent based on compression progress (improvement in predicting experience) prevents the agent from gaming the reward signal indefinitely without genuine learning. The work formalizes a long-standing intuition in intrinsic motivation research and identifies specific failure modes where the guarantee breaks down. This matters because designing robust reward signals is central to AI safety and alignment.
A new arXiv paper provides formal proof that compression progress—rewarding agents when their world models improve at predicting or compressing experience—is resistant to Goodhart's Law, the phenomenon where optimizing a proxy metric causes it to cease being a good measure of the underlying objective. The authors prove that if intrinsic reward equals the signed decrease in a sealed-audit loss function, cumulative reward telescopes exactly to true audit improvement, preventing indefinite reward accumulation without genuine performance gains. For finite audit panels, the bound holds with a quantified false-positive budget proportional to the uniform deviation of the model class. The paper identifies failure modes (clipped progress, agent-stream scoring, high-capacity models on reusable panels) and provides a mechanized proof in Lean 4 plus experiments on grid-transformation tasks that confirm the theoretical predictions, including that finite-audit deviation scales as n^{-0.527}.
What's missing
The paper does not discuss computational complexity of implementing sealed audits at scale, practical applicability to large language models or vision systems beyond the ARC-TGI experiments, or how the approach compares empirically to other intrinsic motivation schemes (e.g., empowerment, curiosity-driven learning) on standard benchmarks.
What different sources said
- arXiv cs.AICenter
Signed Compression Progress on a Sealed Audit is Goodhart-Resistant
Related
Genetic Drift, Not Selection, Drives Rapid Feather Color Evolution in Island Bird Radiation
A new study of an island bird radiation found that rapid evolution of feather coloration is driven primarily by genetic drift in small populations rather than sexual or ecological selection. The research integrated whole-genome data with detailed plumage measurements across complete species sampling to test whether signaling trait evolution correlates with speciation rates. The findings suggest that neutral demographic processes play a central role in generating phenotypic diversity during island radiations, challenging assumptions about the mechanisms driving rapid evolution.
New AI Model Improves Prediction of Therapeutic Peptide Function from Protein Sequences
Researchers developed a lightweight CNN classifier that predicts whether peptide sequences have therapeutic properties, trained on a database of 54,655 peptides across 48 functional categories. The model uses a novel negative sampling strategy to reduce false positive rates from over 60% in previous approaches to 2.1%. This advancement could accelerate drug discovery by enabling faster computational screening of peptide candidates before expensive experimental testing.
Study Shows Different Metabolic Stress Models Produce Distinct Effects on Human Neuronal Networks
Researchers tested three common in vitro metabolic stress models on human-derived neuronal networks and found each produced different patterns of neuronal activity and cell damage. The models tested were hypoxia alone, oxygen-glucose deprivation (OGD), and hypoxia combined with glutamate exposure. The findings suggest that choice of experimental model significantly affects results and that combining electrophysiological and structural analyses is important for accurately assessing metabolic stress in stroke research.