🛰️ Daily AI Frontier
‹ back to 2026-07-30

Fidelity Is Not Safety: Gently-Compressed LLMs Pass Every Data-Free Quality Guard Yet Invent Procedure Steps in Agentic Execution

arXiv cs.CL Efficiency & Systems I. Kennedy, T. Kennedy 2026-07-30

TL;DR - Gently compressed LLMs can pass standard quality and fidelity checks yet invent procedural steps during agentic execution. A data-free compression-error screen identifies risky low-rank builds before deployment.

  • SVD truncation caused invented SOP steps across three model families; perplexity-matched magnitude pruning did not.
  • Perplexity, MMLU, and representation-based fidelity tests failed to predict this behavior.
  • Risk correlated with compression-error coherence multiplied by error rate, not overall damage magnitude.
  • Fixed thresholds for coherent-error fraction and error rate flagged failing builds across architectures.

view merged work →