Grouping of records changes how AI models weigh evidence
Record Grouping Controls Evidence Weight in Language Models
Computation and Language
Summary
When AI language models make decisions, they look at pieces of information called records. This paper shows that how these records are grouped before the AI sees them can change the model’s conclusions. The authors developed a way to group records to remove duplicates but keep important details, and they measured how splitting or merging these groups affects the model’s answers. Their tests across many models found that small changes in grouping can shift the AI’s decisions in noticeable ways. This means that how information is prepared for AI models is an important factor that can be controlled and studied.
language modelsrecord groupingevidence weightingdata partitioningcontent deduplicationmodel checkpointdecision shiftspre-generation representationpartition errornatural language processing
Authors
Zhongxuan Liu, Sicheng Zhou, Hongzhi Wang
Abstract
Retrieved records are presentation units; a supplied partition determines which records enter a language model as one evidential contribution. We characterize the invariant group-content state that removes within-group copies while retaining complementary canonical content, show that equal group counts can encode different evidence states, and derive a sharp content-aware partition-error bound. Given a supplied partition, our pre-generation representation deduplicates and aggregates content within groups and bounds each group's contribution. Across 104,402 trials and 6 public checkpoints, a central natural-text intervention finds that content-fixed false splits add 10.27-32.66 percentage points and false merges remove 9.13-31.79 points; a matched six-slot control retains the positive direction in all 16 cells. In a new 48-item controlled campaign panel, changing the supplied partition produces measurable, checkpoint-dependent decision shifts across all four models, and the balanced mirror design exposes substantial order interactions. Together, the theory and experiments establish the supplied partition as a controllable pre-generation representation variable and characterize its checkpoint-dependent behavioral effects.