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.
More from Research
- New Paper Proposes Object Co-occurrence Framework to Interpret Semantic Representations via LLM Embeddings — TimKietzmann · 2026-08-12
- Training CFD Surrogate Models: Why Sampling Strategies Differ from LLMs and Images — capetorch · 2026-08-12
- A Game-Theoretic Framework for Responsible AI Release: Balancing the Capability Gap — dpaleka · 2026-08-12
- Reliable LLM Computation with ~400K Parameters: Discrete Execution Boundary — kostrubaty · 2026-08-12
- Trained Planaria Retain Memories Through Head Regeneration, Preprint Shows — drmichaellevin · 2026-08-12
- 360CityArena: A Photorealistic Urban Navigation Benchmark for Embodied Agents — hal-utokyo · 2026-08-12