Tag
This paper shows that a deployed self-improving science agent already computes an MDL compression ratio that can unsupervisedly separate law-grade from statistical-grade scientific knowledge, recovering canonical laws with high significance while revealing 86% of its own admitted discoveries compress nothing — the first measured instance of a discovery agent auditing its own ledger and finding itself over-claiming.
This paper introduces a recursion formula for efficiently computing the stochastic complexity of vectors with cluster structure using the Normalized Maximum Likelihood model, reducing time complexity from polynomial to linear.
This paper analyzes integer-sequence benchmarks for language models using minimum description length, revealing that these benchmarks often measure memorisation rather than inductive reasoning, and introduces a new difficulty measure.
SkillZip is a new method for compressing the accumulated skills of self-evolving agents without evaluation rollouts, by finding a minimal faithful structural explanation that reuses repeated rules and procedures while preserving rare exceptions.