Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Activations as Features: Probing LLMs for Generalizable Essay Scoring Representations

Jinwei Chi, Ke Wang, Yu Chen, Xuanye Lin, Qiang Xu

TL;DR

This work investigates using intermediate activations from large language models (LLMs) to improve cross-prompt automated essay scoring (AES). By applying simple linear probes to per-head activations extracted from final-token representations, the authors demonstrate strong discriminative power for essay quality and trait evaluation, often outperforming established baselines. The study also analyzes how model choice, input content, and trait/prompt directions affect performance, showing that instructions and model scale influence results, while activation directions reveal adaptable evaluation perspectives across prompts and traits. Overall, activations-based probing offers a promising direction for robust cross-prompt AES without extensive fine-tuning. The findings have implications for deploying prompt-agnostic scoring systems and for understanding how LLMs internally represent writing quality and trait dimensions.

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.

Activations as Features: Probing LLMs for Generalizable Essay Scoring Representations

TL;DR

This work investigates using intermediate activations from large language models (LLMs) to improve cross-prompt automated essay scoring (AES). By applying simple linear probes to per-head activations extracted from final-token representations, the authors demonstrate strong discriminative power for essay quality and trait evaluation, often outperforming established baselines. The study also analyzes how model choice, input content, and trait/prompt directions affect performance, showing that instructions and model scale influence results, while activation directions reveal adaptable evaluation perspectives across prompts and traits. Overall, activations-based probing offers a promising direction for robust cross-prompt AES without extensive fine-tuning. The findings have implications for deploying prompt-agnostic scoring systems and for understanding how LLMs internally represent writing quality and trait dimensions.

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.
Paper Structure (24 sections, 4 equations, 4 figures, 4 tables)

This paper contains 24 sections, 4 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Automated Essay Scoring
  4. Probing
  5. Methodology
  6. Input Content
  7. Acquisition of Activations
  8. Probe
  9. Experiment
  10. Dataset
  11. Validation and Evaluation
  12. Baselines
  13. Results and Discussion
  14. Main Result
  15. Performance Under Different Probes
  16. ...and 9 more sections

Figures (4)

  • Figure 1: Input content template.
  • Figure 2: Visualization of activations and heatmaps of QWK scores. \ref{['P7CONT']} and \ref{['P4HOL']} represent the distributions of activations after t-SNE dimensionality reduction. \ref{['P3NAR']} presents two heatmaps of QWK scores under specific prompts and traits.
  • Figure 3: The score for each word in the top 8 attention heads. Those greater than 0.5 will be colored.
  • Figure 4: Average cosine similarity between different directions. LM stands for Llama-2-7b-chat-hf, DS for DeepSeek-R1-Distill-Llama-8B, and QW for Qwen3-8B.