ARGUS Agentic Framework Uses Deterministic Verifiers to Limit LLM Hallucinations on Regulatory Variants
arXiv preprint dated October 8 2026 introduces ARGUS, an evidence-constrained agentic system that separates deterministic biological computation from LLM reasoning to improve…
- ARGUS employs a hypothesis-directed loop with a planner that selects evidence sources according to uncertainty and a verifier that applies deterministic interpretation.
- On the example variant rs6983267, the planner generated divergent investigation paths for different transcription factors before reaching conclusions or abstaining.
- LLM-mediated planning produced identical verdicts to fixed-priority approaches but required fewer tool calls by skipping evidence that could not resolve the claim.
The Challenge of Noncoding Regulatory Variants
Over 90 percent of disease-associated variants identified by genome-wide association studies lie in noncoding regulatory regions.
Large language models prompted to interpret such variants routinely hallucinate transcription factor binding changes, fabricate experimental support, and assign biological significance to statistically negligible signals.
ARGUS Framework Design
The preprint presents ARGUS, short for Agentic Regulatory Genomics for an Uncertainty-aware Scientist.
It strictly separates deterministic biological computation from LLM-mediated reasoning inside a hypothesis-directed investigation loop.
A planner selects evidence sources based on current uncertainty while a verifier deterministically interprets each observation, allowing intermediate results to alter the investigation path.
Evaluation on Cancer Risk Variant
The authors illustrate the framework using variant rs6983267 at the 8q24 cancer risk locus.
The same planner produced divergent trajectories when applied to four different transcription factors.
All observations derive from real queries to external resources; none are simulated. This capture is limited to the preprint abstract.
Comparison of Planning Strategies
The work compares fixed-priority planning with LLM-mediated planning.
The LLM planner reached identical verdicts while using fewer tool calls by declining evidence that cannot resolve the specific claim under test.
Implications for Transparent AI Use
The reported separation of deterministic verification from LLM reasoning demonstrates how agentic constraints can reduce hallucinations in specialized scientific tasks.
Such techniques underscore the value of greater transparency and user control over model outputs when applying AI to complex domains.
