A new AI research paper says systems that pool evidence from many sources are quietly doing bad math, and it has a fix.
Researchers behind the paper argue that most systems asking a language model to reach a conclusion from many sources dump everything into one prompt. That conflates two different jobs: reading and interpreting a single source (which rewards a big model with lots of context) and combining those interpretations into a verdict (which needs fixed arithmetic and comparable scores across sources). The authors propose separating the two steps with a four-field evidence tuple, made up of a hypothesis, a reliability bucket, a rationale, and provenance, passed between them. They also name a specific failure mode, "count-scale drift": thresholding a sum of unweighted scores behaves like probability thresholding, but the threshold quietly shifts depending on how many sources get consulted and how reliable the reader is.
This is not just a language-model quirk. The authors say the same drift shows up anywhere a system tallies weighted votes, including triage engines, diagnostic panels that count positive results, and other additive multi-signal detectors, and that pooling calibrated log-likelihood ratios rather than summing raw scores avoids it. Tested on a longitudinal medical-style corpus, a small sequence encoder paired with a tree ensemble reached 0.921 AUPRC, against 0.805 for a hand-crafted baseline.
It is a useful reminder that when a model-based system "gets it wrong," the culprit is often the arithmetic bolted on around the model, not the model's read of any one source.