Decoding-Level Taboo: A Diagnostic Stress Test for LLM Robustness

Tadanobu Chuyo Kamijo · hf · 2026-08-12

Decoding-Level Taboo is a runtime logit-space stress test designed to evaluate the robustness of large language models (LLMs).

This diagnostic method reveals how LLMs handle off-nominal generation paths. The research demonstrates that model robustness in these scenarios depends heavily on parameter scale and instruction alignment.

Original post →

More from Research

Research channel →