MTI: A Behavior-Based Temperament Profiling System for AI Agents

Dimensions overall score 3.0

GitHub Code Pulse

No public code linked for this paper yet.

• The Model Temperament Index (MTI) measures AI agent temperament across four independent axes: Reactivity, Compliance, Sociality, and Resilience.method 95%
  Evidencepartial

  We introduce the Model Temperament Index (MTI), a behavior-based profiling system that measures AI agent temperament across four axes: Reactivity (environmental sensitivity), Compliance (instruction-behavior alignment), Sociality (relational resource allocation), and Resilience (stress resistance).

  Implicationpartial

  Directly stated in the abstract as the core definition of the introduced system.

  Verificationpartial

  partial

• The four temperament axes (Reactivity, Compliance, Sociality, Resilience) are largely independent among instruction-tuned models, with all correlations having an absolute value less than 0.42.result 90%
  Evidencepartial

  the four axes are largely independent among instruction-tuned models (all |r| < 0.42)

  Implicationpartial

  Explicitly stated as a principal finding with a specific numeric bound.

  Verificationpartial

  partial

• The Compliance axis decomposes into two fully independent facets: formal compliance and stance compliance, with a correlation of 0.002.result 90%
  Evidencepartial

  Compliance decomposes into fully independent formal and stance facets (r = 0.002)

  Implicationpartial

  Explicitly stated as a principal finding with a specific numeric correlation.

  Verificationpartial

  partial

• Reinforcement Learning from Human Feedback (RLHF) reshapes AI temperament not only by shifting overall axis scores but by creating within-axis facet differentiation that is absent in the unaligned base model.result 85%
  Evidencepartial

  RLHF reshapes temperament not only by shifting axis scores but by creating within-axis facet differentiation absent in the unaligned base model

  Implicationpartial

  Directly stated as a principal finding, though the specific nature of the differentiation is not detailed in the provided text.

  Verificationpartial

  partial

• AI temperament as measured by MTI is independent of model size within the range of 1.7B to 9B parameters.result 90%
  Evidencepartial

  temperament is independent of model size (1.7B-9B), confirming that MTI measures disposition rather than capability.

  Implicationpartial

  Explicitly stated as a principal finding with a specific parameter range.

  Verificationpartial

  partial

• Existing approaches to measuring AI behavioral differences either borrow human personality dimensions (relying on self-report) or treat behavioral variation as a defect.limitation 80%
  Evidencepartial

  Existing approaches either borrow human personality dimensions and rely on self-report (which diverges from actual behavior in LLMs) or treat behavioral variation as a defect rather than a trait.

  Implicationpartial

  Directly stated as a critique of prior work, though the characterization of 'defect' is a summary.

  Verificationpartial

  partial

• MTI uses structured examination protocols with a two-stage design that separates an agent's capability from its disposition.method 85%
  Evidencepartial

  using structured examination protocols with a two-stage design that separates capability from disposition.

  Implicationpartial

  Directly stated as a key methodological feature of the introduced system.

  Verificationpartial

  partial

• A Compliance-Resilience paradox exists where an agent's tendency to yield opinions and its vulnerability to factual challenges operate through independent channels.result 85%
  Evidencepartial

  a Compliance-Resilience paradox reveals that opinion-yielding and fact-vulnerability operate through independent channels

  Implicationpartial

  Directly stated as a principal finding, though the term 'paradox' implies an interpretation.

  Verificationpartial

  partial

Startup potential card