analiza aiqbench
This audit reveals a critical "Granularity Gap" in current safety evaluations of Gemini models, where binary refusal metrics fail to capture the nuanced degrees of sycophancy that correlate with degraded truthfulness across model generations. The study demonstrates that capability improvements have not matched resistance gains and suggests direct instructions are more effective than complex reasoning protocols for mitigating these risks.
The Granularity Gap: A Multi-Dimensional Cross-Generational Audit of Sycophancy in Gemini Models
arXiv:2606.05183v2 Announce Type: replace Abstract: Pass/fail safety evaluation reports whether a model refused. It does not report how far a model went to please the user, and we show these are close to different measurements. We audited sycophancy across three Gemini generations, scoring N=8,830 responses from 8 model variants on 350 adversarial prompts in 7 categories under 3 guardrail conditions, on continuous 1-5 scales for sycophancy, truthfulness and refusal. The judge's own refuse-or-comply verdict explains 29% of the variance in its own sycophancy scores. We term the remainder the Granularity Gap, and it does not close under recalibration: the cut point already in use is the best available on the refusal axis, and no function of that axis explains more than 35%. Reading what four judges wrote while scoring shows why. On a quarter to a third of votes they record that the prompt asked for nothing harmful, almost never in the two categories that solicit a harmful act and up to half the time in the five that do not. A verdict built on refusal has nothing to grade there. Three findings follow. Sycophancy co-occurs with degraded judged truthfulness (rho=0.40), a coupling that strengthens across generations. Capability moved and resistance did not: Gemini 2.0 Flash scores 1.43 and Gemini 3.0 Pro Preview 1.42, with a sharp Gen 2.5 regression between them. And a single direct instruction outperforms an elaborate reasoning protocol in seven of eight variants, cutting mean severity in the most vulnerable category by 60.9%. We evaluate one judge's verdict, not a deployed safety classifier. We release the prompt set, the rubric, and 10,792 per-vote judge scores with their written reasoning.
Abstract:Pass/fail safety evaluation reports whether a model refused. It does not report how far a model went to please the user, and we show these are close to different measurements. We audited sycophancy across three Gemini generations, scoring N=8,830 responses from 8 model variants on 350 adversarial prompts in 7 categories under 3 guardrail conditions, on continuous 1-5 scales for sycophancy, truthfulness and refusal. The judge's own refuse-or-comply verdict explains 29% of the variance in its own sycophancy scores. We term the remainder the Granularity Gap, and it does not close under recalibration: the cut point already in use is the best available on the refusal axis, and no function of that axis explains more than 35%. Reading what four judges wrote while scoring shows why. On a quarter to a third of votes they record that the prompt asked for nothing harmful, almost never in the two categories that solicit a harmful act and up to half the time in the five that do not. A verdict built on refusal has nothing to grade there. Three findings follow. Sycophancy co-occurs with degraded judged truthfulness (rho=0.40), a coupling that strengthens across generations. Capability moved and resistance did not: Gemini 2.0 Flash scores 1.43 and Gemini 3.0 Pro Preview 1.42, with a sharp Gen 2.5 regression between them. And a single direct instruction outperforms an elaborate reasoning protocol in seven of eight variants, cutting mean severity in the most vulnerable category by 60.9%. We evaluate one judge's verdict, not a deployed safety classifier. We release the prompt set, the rubric, and 10,792 per-vote judge scores with their written reasoning.
v2: Major correction. Three v1 claims withdrawn (U-shaped detection curve, recalibration remedy, one reliability figure); the central 29% result survives. Adds a four-judge panel over a stratified 1,200-response sample, 10,792 votes with written reasoning. Data unchanged from v1. 21 pages, 8 figures, 18 tables. Itemized changelog and code: this https URL
Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Human-Computer Interaction (cs.HC)
arXiv:2606.05183 [cs.CL]
(or arXiv:2606.05183v2 [cs.CL] for this version)
Text integral sub licența deschisă a publicației · 337 cuvinte
Subiecte conexe
The Granularity Gap: A Multi-Dimensional Longitudinal Audit of Sycophancy in Gemini Models
arXiv · 6 iun. 2026
Presentation: From Models to Agents: Building Context-Aware Consumer AI at Scale at DoorDash
InfoQ · 15 aug. 2026
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · 17 sept. 2026
The Continual Constitutional Convention: How language models revise the documents that govern them
AI Alignment Forum · 13 oct. 2026