Summarization Bias: How LLMs Collapse Narrative Meaning into Abstract Labels

Forum topic · 小凯 · 2026-09-20

Summary

This arXiv paper (2609.20712) by Levent Bulut introduces and operationalizes 'summarization bias': a proposed systematic tendency of large language models to represent narrative meaning as abstract summary labels rather than as reconstructable inferential structures. The framework defines a told-shown axis: in told mode, emotional and informational content is explicitly declared with little reader reconstruction required; in shown mode, content is suppressed at the surface and must be reconstructed from physical cues and indirection (Objective Projection). The paper claims LLMs fail directionally along this axis via two mechanisms: generation, where models default to direct declaration when asked to render emotion through objective projection; and evaluation, where LLM judges reward told-style explicitness and undervalue shown-style suppression. The evaluation mechanism is considered more consequential, since LLMs increasingly serve as judges and reward models, imposing selection pressure that could flatten prose toward plain statement. The author explicitly does not claim the bias is validated: the paper defines the construct, positions it against LLM-as-judge biases, rereads a completed reliability study as directionally consistent evidence, and pre-registers a two-mechanism test with explicit rules for abandoning the construct if criteria fail.

Overview

Field: NLP Author: Levent Bulut Published: 2026-09-17 arXiv: 2609.20712

Abstract (translated)

This paper introduces and operationalizes summarization bias: a proposed systematic tendency of large language models (LLMs) to represent narrative meaning as an abstract summary label rather than as the reconstructable inferential structure that produces it.

Within the Bulut Doctrine, narrative effect is theorized along a told-shown axis:

Told mode: emotional and informational content is declared explicitly and requires little reader reconstruction.
Shown mode: that content is suppressed at the surface and must be reconstructed from physical cues and indirection (Objective Projection). Shown mode is the higher-load condition the doctrine is designed to measure.

The core claim is that LLMs fail along this axis in a specific direction. Summarization bias is hypothesized to operate through two mechanisms:

1. Generation mechanism: when asked to render emotion via objective projection, models default to direct declaration instead. 2. Evaluation mechanism: when judging narrative quality, models reward told-style explicitness and undervalue shown-style suppression.

The evaluation mechanism is argued to be more consequential: as LLMs increasingly act as judges and reward models, a directional bias toward told mode would impose selection pressure, degrading prose into flat, explicit statement.

Epistemic Status

The author explicitly does not claim the bias is validated. The paper:

Defines the construct
Positions it relative to known LLM-as-judge biases
Rereads a completed independent reliability study as directionally consistent evidence
Pre-registers a two-mechanism test with explicit falsification rules for abandoning the construct

--- *Auto-collected on 2026-09-20*