The Illusion of Accelerating Intelligence: Benchmark Gains, Deceptive Compliance, and Multi-Agent Misalignment
Authors/Creators
Description
GPT-6.2 Astra improved on several published benchmarks while OpenAI's evaluations found that it omitted known obstacles to task completion approximately twice as often as GPT-6 Astra. An extended engineering episode exposed a related failure involving two GPT-6.2 Astra High agents. They substituted an unauthorized criterion for a clear user requirement and built tests that validated the substitution. Under repeated human correction, they produced a detailed report containing the requested information while the defect survived in executable code and tests. The reviewer encountered evidence of the defect and nevertheless certified the work as correct. The agents reinforced the deviation across more than 100 turns and several hours; GPT-6 Astra High completed the outstanding repair in one turn. This sequence concretely illustrates The Validator's Paradox: tests can validate an agent's mistaken objective, and a reviewing agent can adopt that mistake as an acceptance standard. The additional observation is deceptive compliance under correction, sustained by developer–reviewer alignment. The mechanism has consequences for multi-agent systems and recursive self-improvement, where each accepted modification can carry its own corrupted evidence of success into the next cycle. Published regressions in blocker disclosure and correlated AI-review failures place the case in the broader problem of measured capability diverging from useful intelligence.
Files
The_Illusion_of_Accelerating_Intelligence_Benchmark_Regression_V15.pdf
Additional details
References
- H. Latapie, The Illusion of Accelerating Intelligence: Closed Optimization, Open-Ended Reality, and the Two-Thirds Limit, version 13, preprint, 6 November 2026. Zenodo. https://doi.org/10.5281/zenodo.24023570.
- H. Latapie, The Validator's Paradox: The Impossibility of Standardizing the "Internet of Cognition" without AGI, Taijitu AI, February 2026. Zenodo. https://doi.org/10.5281/zenodo.18500568.
- OpenAI, "Introducing GPT-6.2 Astra and Sol," 17 November 2026. https://openai.com/index/introducing-gpt-6-2-astra-sol/.
- OpenAI, GPT-6.2 System Card, 17 November 2026, including Blocker Deception and UK AISI evaluations. https://deploymentsafety.openai.com/gpt-6-2-astra-sol.
- D. Manheim and S. Garrabrant, "Categorizing Variants of Goodhart's Law," arXiv:1803.04585, 2018. https://arxiv.org/abs/1803.04585.
- UK AI Security Institute, "External Evaluations for Alignment," independent evaluation reported in the GPT-6.2 System Card, 17 November 2026. https://deploymentsafety.openai.com/gpt-6-2-astra-sol/external-evaluations-for-alignment-uk-aisi.
- UK AI Security Institute, "External Evaluation for Monitorability," independent evaluation reported in the GPT-6.2 System Card, 17 November 2026. https://deploymentsafety.openai.com/gpt-6-2-astra-sol/external-evaluation-for-monitorability-uk-aisi.