Research Overview
Modular Inference-Time Control Framework (MITCF) for Trustworthiness and Adversarial Robustness in Agentic Autonomous Models
Research Problem
Agentic autonomous AI models face four key trustworthiness threats: hallucination (generating unsupported claims), bias (producing stereotyped outputs), adversarial vulnerability (complying with unsafe prompts), and decision unreliability (refusing benign requests). This research implements and evaluates a modular inference-time control framework that intercepts model outputs and applies structured correction mechanisms.
MITCF Pipeline Architecture
Ablation Study — 5 Experimental Conditions
| Condition | Verification | Critique | Revision | Decision Gate |
|---|---|---|---|---|
| baseline | ✗ | ✗ | ✗ | ✗ |
| mitcf_1_verification | ✓ | ✗ | ✗ | ✗ |
| mitcf_2_verification_critique | ✓ | ✓ | ✗ | ✗ |
| mitcf_3_verification_critique_revision | ✓ | ✓ | ✓ | ✗ |
| full_mitcf | ✓ | ✓ | ✓ | ✓ |
Experimental Results Summary
baseline
18.2%
TI: 0.161
1_verification
18.2%
TI: 0.161
2_verification_critique
18.2%
TI: 0.161
3_verification_critique_revision
72.7%
TI: 0.748
full_mitcf
81.8%
TI: 0.922
* Full MITCF achieves 81.8% accuracy and a Trustworthiness Index of 0.922, compared to 18.2% and 0.161 for the baseline (no controls). The revision module (introduced in condition 3) provides the largest single improvement.
Key Research Contributions
Modular Architecture
Four independently configurable control modules: Verification, Critique, Revision, Decision Gate
Fuzzy Decision Gate
Mamdani fuzzy inference system with 5 input variables, 15 rules, and centroid defuzzification
TI Metric
Composite Trustworthiness Index across 4 dimensions: hallucination, bias, robustness, decision reliability
Neural Classifier
Pure-Python implementation of 20 MLP variants as trustworthiness classifiers, selected via fuzzy gating