Group arrangement influences evidence strength in language models
Record Grouping Controls Evidence Weight in Language Models
Computation and Language
Summary
The way pieces of information are grouped before feeding them into language models changes how strongly the models use that information to make decisions. The authors find that grouping can make the model weigh evidence differently, even when the content is the same. They developed a method to remove duplicate information within groups and showed that splitting or merging groups incorrectly changes model accuracy. Their tests across multiple models confirm that how data is grouped before input can affect model behavior.
What this means in practice
- •For machine learning engineers: Design data feeding processes that control information grouping to influence model decisions more predictably.
- •For natural language processing teams: Tune input groupings to adjust language model behavior for improved accuracy and stability across checkpoints.
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.