A diagnostic framework is introduced that abstracts chain-of-thought reasoning as a trajectory through discrete latent reasoning states, yielding a positional taxonomy of six reasoning pathologies, each paired with a transition-grounded detection predicate and a matched inference-time controller that positions it as a reproducible diagnostic substrate for studying CoT failures.
It is argued that emergent reasoning in LLMs is a product of three factors: model size, prompting approach, and evaluation metric, and proposed implications for designing benchmarks and assessing capabilities are proposed.
N. Kuotsu· International Journal of Cre...· 0 citations
Large language models (LLMs) are usually evaluated by final-output correctness or safety, yet an endpoint label does not reveal how a failure develops throughout the reasoning process. We therefore ask whether ordinary reasoning failures and adversarially induced failures can be characterized using the same internal co...
Zi-Rui Ding, Xin-Xin Fan, Hao Yuan et al.· 0 citations
This work uses synthetic multi-hop lookup tasks to measure faithfulness causally at the activation level, specifically on self-generated reasoning, and aims to measure faithfulness causally at the activation level, specifically on self-generated reasoning.
Explicit reasoning in large language models (LLMs) is often associated with reliability, yet it can still pose risks when a model infers harmful goals under a benign context. We identify a distinct challenge: ensuring the safety of the \emph{multi-step reasoning process} itself, where locally coherent steps may still l...
Hao Yang, Ke Ma, Yangbangyan Jiang et al.· 1 citation
Ask a reasoning model a harmful question and it refuses; ask it again after a long, entirely benign reasoning chain and it answers. Across six open-weight models, average refusal falls from 94.6% to 12.7% behind an 8,192-token benign prefix. We name this failure Safety Dilution and make it measurable: refusal decays sm...
Dong-Xing Guo, Ji-Kun Wu, Siu-Ming Yiu· Proceedings of the 2026 ACM...· 0 citations
This framework organizes existing methods around five components: explicit state representation, transition structuring, validation and constraint enforcement, search and rollback, and uncertainty management, and yields a diagnostic hypothesis: interventions should be most effective when they target the error or uncert...
Jia-Zhang Cai, Tao Wang, Rui-Dong Zhang et al.· arXiv.org· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.