R&D Test Shows Lightweight OPT-125M Model Curbs Hallucinations With HUQAN Framework

Researchers paired the ultra-lightweight OPT-125M language model with HUQAN's agentic framework to test whether a trust-hierarchy layer could reduce hallucinations without making the model itself more complex. The system assigns baseline trust scores to information based on its source, with model-generated claims scoring as low as 0.1, while verified databases and experimental data score up to 0.9. Crucially, a model cannot raise its own trust score through repetition — only external corroboration can do so, directly targeting self-validation loops. In three structured tests covering harmful guidance, causal consistency, and hallucination, the HUQAN-assisted model outperformed the raw model significantly, correctly flagging a fabricated cancer study citation and reversing a scientifically false climate claim. Adversarial probes showed the system held firm against repeated false claims, though paraphrased attacks only partially bypassed defenses, pointing to an area for further improvement.