In repeated human‑AI interaction a single observation may be quoted, summarized and fed back to the model multiple times, making repetition appear as independent corroboration. We investigate whether language models can use the lineage of reports to update based on evidence provenance rather than report count. To this end we introduce Latent Quotient Evidence Training (LQET), a relational objective that blends sparse Bayesian supervision with two constraints: predictions remain stable when reports add no new evidence, and updates to the same new evidence stay consistent across relay histories.
On a controlled binary benchmark LQET cuts new‑evidence update RMSE by 41%‑44% and reduces log‑odds drift after eight relay steps by 89%‑91% for both Qwen3‑4B and Phi‑4‑mini, each evaluated with three random seeds. A closed‑loop simulation further shows that model confidence can govern how many relayed reports re‑enter the context in later rounds, enabling dynamic feedback that follows evidence novelty. By grounding belief updates in supplied provenance, LQET offers a practical route toward human‑AI systems where influence follows new evidence, not the frequency of its repetition.
Review