Overview
Field: NLP Authors: Tadanobu Chuyo Kamijo, Ori Rottenstreich, Javier Conde Published: 2026-08-12 arXiv: 2508.05140
Abstract (translated)
Large language model evaluations typically focus on performance under nominal conditions, creating an illusion of capability where models comfortably walk a narrow, highly optimized generation corridor. In real-world deployments, however, complex system prompts, safety guardrails, and structural constraints continuously force models off this nominal path, driving a divergence between benchmark scores and deployed performance.
To address this issue, the authors introduce Decoding-Level Taboo, a zero-prompt diagnostic stress test that intervenes directly in logit space at runtime, forcing models out of their nominal paths. By dynamically masking primary candidate tokens at word boundaries, Taboo forces machine circumlocution — the model must find alternative ways to express the intended content.
Findings
- Evaluating Taboo across several open-weight model families reveals that off-path robustness is strongly influenced by parameter scale and post-training instruction alignment.
- Robustness generally improves with larger model scale and stronger alignment.
- Generating diverse synthetic datasets
- Stress-testing runtime safety guardrails
- Auditing model reliability before real-world deployment
Applications
Beyond the results presented in the paper, Taboo provides a novel primitive for: