Same evidence, repeated: LQET stops models from treating relays as new corroboration
Synopsis
On a controlled binary benchmark with supplied report lineage, LQET makes models update beliefs by evidence provenance rather than report count, cutting new-evidence update RMSE by 41-44% and log-odds drift after eight relay steps by 89-91% relative to task-only training.
Interpretation
A model can learn to update its belief according to the evidence a report carries, not the number of times a report appears. Pointwise supervision assigns each transcript its own target, so it does not directly enforce the cross-transcript relation that the same evidence should produce the same update across relay histories. On a controlled binary benchmark, LQET lowers new-evidence update RMSE by 41.25% on Qwen3-4B and 44.45% on Phi-4-mini, and depth-8 relay drift by 91.36% and 89.38%, with three paired training seeds per backbone.
The training objective translates visible relay relations into consistent belief updates. LQET adds four provenance-invariance constraints, three new-evidence update-consistency constraints, and a presentation-order constraint to sparse soft-target supervision, using 13 training views where task-only uses seven pointwise-supervised views. Each backbone trains on 4,000 families with 400 validation families for 500 updates, then is evaluated on 400 IID families, 800 controlled-shift families, and 200 recursive-relay families.
The stability comes from selectively ignoring relays, not from flattening all updates. In a matched comparison that holds report author and displayed outcome fixed while varying supplied provenance, copy-versus-new RMSE drops from 1.0145 to 0.4383 on Qwen and from 0.9818 to 0.4630 on Phi, while new-root RMSE stays close to task-only. The zero-intercept copy-versus-new slope in the same comparison rises from 0.027 to 0.782 on Qwen and from 0.067 to 0.769 on Phi, against an ideal value of one.
Perspective
The result applies to a controlled binary setting in which parent-report links are supplied and the source-reliability model is specified, which allows exact tests of evidence accounting. Evaluation covers two 4B-class backbones and three paired training seeds, and conclusions stay within that scope. The source explicitly lists transfer to larger models, naturalistic tasks, and incomplete or unreliable provenance as further study.
The forced-choice YES-NO score assesses binary preference, while free-form probability estimation and downstream decision quality require separate evaluation. Under held-out prior and calibration supports, lower update RMSE accompanies a supervised-view KL increase of about 0.013 nats, leaving numerical generalization a remaining challenge. The confidence-dependent feedback simulation isolates report recirculation under a prescribed trust rule, and the source states that studies with human participants are needed to assess how the mechanism operates in practical human-AI interaction.