# Surprisal Theory is Tautological (without Rational Grounding)

Source: [arXiv](https://arxiv.org/abs/2607.21574v1)  
Feed7 permalink: https://feed7.dev/p/2607-21574v1-0m2upel  
Published: 2026-07-23T17:54:37.000Z  
Trust: Needs Review (needs_review)

## Why Included

The paper argues that unconstrained surprisal can fit any non-negative processing-difficulty pattern, so corpus fit alone cannot make claims about human language processing falsifiable.

## Source Summary

The paper shows that, under mild conditions, **any non-negative difficulty measure** can be represented as an affine function of surprisal under some language model. Without constraining that model, surprisal theory therefore makes no falsifiable prediction.

## Practical Implication

Builders evaluating language models against human behavior should not treat better corpus likelihood as evidence of better cognitive fidelity. The proposed remedy is to derive the model from independent assumptions such as memory limits or processing goals.

## Agent-Ready Context

The paper shows that, under mild conditions, **any non-negative difficulty measure** can be represented as an affine function of surprisal under some language model. Without constraining that model, surprisal theory therefore makes no falsifiable prediction.

Builders evaluating language models against human behavior should not treat better corpus likelihood as evidence of better cognitive fidelity. The proposed remedy is to derive the model from independent assumptions such as memory limits or processing goals.

This is a theoretical argument, not a new agent evaluation or empirical benchmark. Its practical force depends on whether researchers can specify independently motivated comprehender models that produce testable predictions.

## Context Map

- Layer: benchmark
- Domains: research
- Topics: benchmark-integrity

## Uncertainty

- This is a theoretical argument, not a new agent evaluation or empirical benchmark. Its practical force depends on whether researchers can specify independently motivated comprehender models that produce testable predictions.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.