analiza aiqbench
This research reframes unexpected failures in large language models as a predictable dynamical process driven by many-body token interactions rather than inherent unpredictability. By modeling these tipping points through finite-layer thresholds and attention disorder, the study provides engineers with actionable metrics to assess foreseeable risks across diverse ChatGPT-like architectures. The findings carry significant weight for legal liability assessments regarding AI-generated harm, suggesting that such errors are engineering challenges solvable via system design adjustments.
Many-body Tipping Dynamics of ChatGPT-like AIs
arXiv:2607.25279v1 Announce Type: new Abstract: Why do ChatGPT-like AIs, despite major architectural and training differences, unexpectedly tip to undesirable content (e.g. harmful, misleading, repetitive) even under deterministic greedy decoding? We show that a broad class of such tippings is caused by the many-body interactions between tokens (spins) as they cross the finite-layer system. Tipping emerges as a dynamical first passage process between competing output basins. Attention disorder controls the transport toward, away from, or along the basins' boundary. A few-basin reduction yields a closed finite-layer threshold, whose coarse-grained predictions show good agreement across ChatGPT-like families. These results suggest that a broad class of AI failures represents 'foreseeable engineering risk' rather than inherently unpredictable behavior, with important implications for legal and societal assessments of AI harm.
Abstract:Why do ChatGPT-like AIs, despite major architectural and training differences, unexpectedly tip to undesirable content (e.g. harmful, misleading, repetitive) even under deterministic greedy decoding? We show that a broad class of such tippings is caused by the many-body interactions between tokens (spins) as they cross the finite-layer system. Tipping emerges as a dynamical first passage process between competing output basins. Attention disorder controls the transport toward, away from, or along the basins' boundary. A few-basin reduction yields a closed finite-layer threshold, whose coarse-grained predictions show good agreement across ChatGPT-like families. These results suggest that a broad class of AI failures represents 'foreseeable engineering risk' rather than inherently unpredictable behavior, with important implications for legal and societal assessments of AI harm.
Artificial Intelligence (cs.AI); Disordered Systems and Neural Networks (cond-mat.dis-nn); Mathematical Physics (math-ph); Adaptation and Self-Organizing Systems (nlin.AO); Physics and Society (physics.soc-ph)
arXiv:2607.25279 [cs.AI]
(or arXiv:2607.25279v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2607.25279 arXiv-issued DOI via DataCite (pending registration)
From: Neil F. Johnson [view email] [v1] Tue, 28 Jul 2026 04:32:54 UTC (3,118 KB)
Text integral sub licența deschisă a publicației · 171 cuvinte
Subiecte conexe
Links #5: 2026/07
LessWrong · 31 iul. 2026
Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Latent Space · 13 oct. 2026
Dr. Jill Lepore on why AI backlash is vital for the future
The Verge · 20 iul. 2026
Our minds aren’t equipped to handle AI
The Verge · 5 oct. 2026