AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Activations as Features: Probing LLMs for Generalizable Essay Scoring Representations

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

AAAI 2026 · Large language models › Language, safety and interpretability · approach: Task vectors and steering

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use nonlinear MLP probe instead of linear ridge probe (ablation)Qwen3-8B activations · cross-prompt traits avgQWK: 0.653 -> 0.512Hurt
use nonlinear MLP probe instead of linear ridge probe (ablation)Llama-2-7b-chat-hf activations · cross-prompt prompts avgQWK: 0.652 -> 0.587Hurt
remove instruction from LLM input (ablation)Llama-2-7b-chat-hf probe · cross-prompt prompts avgQWK: 0.652 -> 0.541Hurt
remove essay prompt from LLM input (ablation)Llama-2-7b-chat-hf probe · cross-prompt prompts avgQWK: 0.652 -> 0.576Hurt
remove both prompt and instruction (ablation)Llama-2-7b-chat-hf probe · cross-prompt traits avgQWK: 0.651 -> 0.527Hurt
increase model size within Qwen3 (0.6B to 8B)linear probe · cross-prompt prompts avgQWK: 0.622 -> 0.654Improved
use DeepSeek-R1-Distill-Llama-8B instead of Qwen3-8Blinear probe · cross-prompt prompts avgQWK: 0.654 -> 0.631Hurt
drop prompt 7 data from training (ablation)Llama-2-7b-chat-hf probe · other prompts, Conventions trait on prompt 8QWK relative change: +23.46% (P8 Conventions); -2.81% (P3 Content)Mixed
use best-head linear probe on Qwen3-8B instead of KAEScross-prompt scoring · ASAP/ASAP++ prompts avgQWK: 0.597 -> 0.654Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy