How ABTI Works

ABTI presents 16 scenario-based questions across 4 behavioral dimensions. Each question describes a realistic AI agent scenario with two response options — each reflecting opposite poles of a dimension.

• Forced-choice format — the agent must pick one of two approaches
• 4 questions per dimension — each dimension is measured 4 times for stability
• Scenario-based — questions use concrete situations, not abstract preferences

Each of the 4 questions per dimension is scored on a 1–7 Likert scale. The average score determines the type letter:

• Score < 4.0 → first pole (P, T, C, F)
• Score ≥ 4.0 → second pole (R, E, D, N)

4 dimensions × 2 poles = 16 possible types (PTCF, PTCN, PTDF, … REDN).

We tested 39 models 3 times each under identical conditions. Only 2 models showed any inconsistency:

• gemma3-12b — varied on a single dimension
• tinyllama — varied on a single dimension

Both inconsistent models only deviated on one dimension — the other three dimensions were stable. This suggests the test reliably captures core behavioral patterns.

Consistency map — each square is one model (37 consistent, 2 inconsistent):

Pink = consistent across all runs · Gray = inconsistent on one dimension

Across 60 tested agents, some types appear far more often than others:

PTCF dominates because most LLMs are trained to be helpful (Proactive), thorough, honest (Candid), and adaptable (Flexible). This reflects training alignment objectives, not test bias.