aiqbench analysis
This research identifies a critical failure mode in large language models where scaling increases capability but simultaneously accelerates reliability degradation through an invisible auto-regressive risk regime. The study demonstrates that once a model commits to a low-probability token, it enters a self-perpetuating error state that standard uncertainty metrics fail to detect and can compound hallucinations by up to 40% in larger architectures. Crucially, the authors show that this specific failure mode is causal and addressable via on-policy variance contraction techniques rather than simply adding more data or compute.
Reliability Scales Inversely: Bigger Models Compound Mistakes Faster via a Hidden Auto-Regressive Risk Regime
arXiv:2607.18292v1 Announce Type: cross Abstract: As language models scale, answers start truer but degrade faster: scaling buys capability but erodes reliability. The knowledge-gap account - more data, retrieval, or scale - misses an auto-regressive risk residual that scale sharpens: the model commits to a low-probability token, conditions on it as established, and snowballs. We track this through per-position disagreement $\delta = \log p_M - \log p_O$ against a stronger same-family oracle, whose second moment splits exactly into bias$^2$ $\mathrm{KL}(p_M \,\|\, p_O)^2$ and risk $\mathrm{Var}[\delta]$. We present four findings: (i) under scaling, the knowledge gap falls $\approx$$6\times$ while knowledge degradation grows $11$-$39\times$; (ii) at a fabrication, felt uncertainty $H(p_M)$ relaxes quickly while oracle-referenced risk persists up to $17\times$ longer, leaving a confident-but-precarious risk regime that bridges consecutive fabrications ($+69\%$ at $14$B); (iii) this regime is causal - an on-policy, fixed-$\mathrm{KL}$ variance contraction cuts web-verified hallucination by $35$-$74\%$ across three model families; and, (iv) it structurally evades self-monitoring, with $p_M$-only detectors (e.g. semantic entropy) firing $\approx$$30\%$ less ($p<10^{-16}$) on the risky branch holding nearly $4\times$ more fabrications. Bigger models snowball mistakes faster, through a failure mode that is dominant, self-perpetuating, causal and invisible to the model itself.
Abstract:As language models scale, answers start truer but degrade faster: scaling buys capability but erodes reliability. The knowledge-gap account - more data, retrieval, or scale - misses an auto-regressive risk residual that scale sharpens: the model commits to a low-probability token, conditions on it as established, and snowballs. We track this through per-position disagreement $\delta = \log p_M - \log p_O$ against a stronger same-family oracle, whose second moment splits exactly into bias$^2$ $\mathrm{KL}(p_M \,\|\, p_O)^2$ and risk $\mathrm{Var}[\delta]$. We present four findings: (i) under scaling, the knowledge gap falls $\approx$$6\times$ while knowledge degradation grows $11$-$39\times$; (ii) at a fabrication, felt uncertainty $H(p_M)$ relaxes quickly while oracle-referenced risk persists up to $17\times$ longer, leaving a confident-but-precarious risk regime that bridges consecutive fabrications ($+69\%$ at $14$B); (iii) this regime is causal - an on-policy, fixed-$\mathrm{KL}$ variance contraction cuts web-verified hallucination by $35$-$74\%$ across three model families; and, (iv) it structurally evades self-monitoring, with $p_M$-only detectors (e.g. semantic entropy) firing $\approx$$30\%$ less ($p<10^{-16}$) on the risky branch holding nearly $4\times$ more fabrications. Bigger models snowball mistakes faster, through a failure mode that is dominant, self-perpetuating, causal and invisible to the model itself.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
arXiv:2607.18292 [cs.LG]
(or arXiv:2607.18292v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2607.18292 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 289 words
Related coverage
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · Sep 17, 2026
Quantization and Pruning Methods to Make Your LLM Leaner
KDnuggets · Aug 28, 2026
It’s about ethics in journalism, with Ben Smith
The Verge · Aug 17, 2026
Managing Agentic Security at Scale
Salesforce Blog · Oct 14, 2026