Tag
Introduces CILN, a framework for generating instance-dependent label noise benchmarks through controlled input corruptions, enabling explicit control over ambiguity source and severity. Experiments show it produces realistic noise structures and exposes failure modes in popular noisy-label learning methods.
The author describes a problem where an LLM agent's memory graph gets corrupted by incorrect edges, and proposes using a declared ontology to validate writes and traversals. A test on 120 deliberately broken traversals caught all errors.
A technical guide on deliberately corrupting ZFS files using the 'zinject' tool or manual byte-level edits to understand ZFS's error handling and self-healing behavior.