Standard evaluations of large language models (LLMs) often paint an overly rosy picture by focusing on performance under ideal conditions. This creates an "illusion of capability," where models excel within a narrow, highly optimized generation corridor. However, real-world deployments frequently push models off this path through complex system prompts, safety guardrails, and structural constraints, leading to a gap between benchmark scores and actual performance.
To address this, researchers from Hugging Face introduce Decoding-Level Taboo, a zero-prompt diagnostic stress test that intervenes directly in logit space at runtime. By dynamically masking the primary candidate tokens at word boundaries, Taboo forces models into "machine circumlocution," compelling them to find alternative ways to express ideas when the most obvious tokens are unavailable.
Evaluating Taboo across several open-weight model families, the team found that off-path robustness is heavily influenced by both parameter scale and post-training instruction alignment. In general, larger models and those better aligned with instructions demonstrate greater robustness when forced off their nominal paths.
Beyond its diagnostic value, Taboo offers a novel primitive for generating diverse synthetic datasets, stress-testing runtime safety guardrails, and auditing model reliability before deployment. This work highlights the importance of evaluating LLMs under off-nominal conditions to ensure they perform reliably in real-world scenarios.