Trustworthy AI systems for sensitive domains through traceability, interpretability, grounding, and provenance
IKnowFlow is my doctoral research framework for trustworthy AI systems in sensitive domains. It treats trustworthiness as a system-level property, so retrieval origins, selection rationales, generated claims, and attribution evidence remain auditable before outputs reach end users.
Stage 1: Retrieval
Interpretability. Uses bidirectional embedding modulation to expose human-readable query-document matching signals and improve retrieval recall.
Stage 2: Evidence Selection
Grounding. Replaces opaque score-based re-ranking with preference-tuned rationales, adaptive chunk selection, and verification.
Stage 3: Attribution and Verification
Traceability. Diagnoses citation and attribution failures across 20,000 scientific papers and 12 domains, showing errors persist even with retrieval augmentation.
Together, these projects move AI systems from loosely retrieved context toward auditable knowledge flow: why a source was retrieved, why evidence was selected, and how generated claims are attributed.
RQ1
How can source-attribution failures be diagnosed across model families and domains?
Progress: REASONS benchmarks 20,000 scientific papers across 12 domains, and related citation work shows post-hoc verification reaches 78% human-evaluated correctness versus 69% for generation-time citation.
RQ2
How can dense retrievers expose why a clinical, legal, or scientific document was selected?
Progress: IMRNNs adds lightweight adapters for query-document embedding modulation, yielding +7.14% recall across seven benchmarks while exposing readable matching signals.
RQ3
How can evidence selection use domain knowledge to distinguish relevant evidence from semantically similar but logically opposed passages?
Progress: METEORA improves answer accuracy by 33.34% with 50% fewer chunks, while Neurosymbolic Retrievers integrate knowledge-graph reasoning with neural retrieval.
RQ4
How should attribution preserve how evidence originates, evolves, validates, or contradicts claims over time?
Next step: Provenance-aware verification will use claim-level source graphs with typed relations such as extends, validates, contradicts, and summarizes.