Activations as Features: Probing LLMs for Generalizable Essay Scoring Representations

TL;DR

This paper investigates the use of intermediate layer activations from large language models to improve automated essay scoring across diverse prompts, revealing their strong discriminative power and adaptability.

Contribution

It demonstrates that LLM activations can serve as effective features for cross-prompt essay scoring, highlighting their potential beyond output-based methods.

Findings

Activations have strong discriminative power for essay quality evaluation.

LLMs can adapt evaluation perspectives to different traits and essay types.

Activations effectively handle diversity in scoring criteria.

Abstract

Automated essay scoring (AES) is a challenging task in cross-prompt settings due to the diversity of scoring criteria. While previous studies have focused on the output of large language models (LLMs) to improve scoring accuracy, we believe activations from intermediate layers may also provide valuable information. To explore this possibility, we evaluated the discriminative power of LLMs' activations in cross-prompt essay scoring task. Specifically, we used activations to fit probes and further analyzed the effects of different models and input content of LLMs on this discriminative power. By computing the directions of essays across various trait dimensions under different prompts, we analyzed the variation in evaluation perspectives of large language models concerning essay types and traits. Results show that the activations possess strong discriminative power in evaluating essay quality and that LLMs can adapt their evaluation perspectives to different traits and essay types, effectively handling the diversity of scoring criteria in cross-prompt settings.