Tech

Scoring Explanations Without Reading Them: Structural Shortcuts in Automated-Interpretability Evaluation Data

By Xupeng Chen · arXiv cs.CL

cs.CL

arXiv:2610.37676v1 Announce Type: new Abstract: Automated interpretability scores an explanation of a sparse-autoencoder (SAE) feature by asking a model to predict, token by token, where the feature fires on evaluation examples…

Published: October 30, 2026 (4 h ago)
Outlet: arXiv cs.CL
Section: Tech
Length here: Headline and 31-word snippet. The full article is at the source.

More from arXiv cs.CL and around the same time

An open laptop and a stack of books on a desk

An Introduction to Large Language Models for Scientists and ProfessionalsarXiv:2610.37684v1 Announce Type: new Abstract: Large Language Models (LLMs) have rapidly become important tools for science, education, medicine, and professional work. Despite their apparent simplicity from the user's…arXiv cs.CL · 4 h ago · reads at arXiv cs.CL

Artifact Distillation: Training-Free Transfer from Frontier to Small Language ModelsarXiv:2610.37680v1 Announce Type: new Abstract: Closing the gap between frontier and smaller language models is usually approached by distillation: fine-tuning the small model (student) on a frontier model's (teacher) outputs…arXiv cs.CL · 4 h ago · reads at arXiv cs.CL

Bouba Beats Kiki, and Other Regularities in the Subjective Preferences of LLMsarXiv:2610.37692v1 Announce Type: new Abstract: Users engage with large language models on tasks that are purely objective ("When was Netflix founded?"), purely subjective ("What should I name my child?") or anywhere in-between…arXiv cs.CL · 4 h ago · reads at arXiv cs.CL

DeFT-RAG: Incrementally Decomposing Multi-Hop Questions into Self-Contained Sub-Questions with Reusable AnswersarXiv:2610.37826v1 Announce Type: new Abstract: Retrieval-augmented generation (RAG) lets a large language model (LLM) answer questions by retrieving relevant passages from a large corpus. However, RAG struggles with multi-hop…arXiv cs.CL · 4 h ago · reads at arXiv cs.CL

Exposure Is Not Commitment: Double-Stratified Evaluation of Latent Monitors for Tool-Using AgentsarXiv:2610.37679v1 Announce Type: new Abstract: Tool-using language agents are vulnerable to indirect prompt injection, and a growing literature probes model internals to catch unsafe actions before they execute, with sharply…arXiv cs.CL · 4 h ago · reads at arXiv cs.CL

From Geometry to Behavior: How Instruction Exposure Unlocks Latent Rhetorical Directions in LLMsarXiv:2610.37685v1 Announce Type: new Abstract: Language models do not merely memorize facts; they absorb rhetorical strategies from pretraining data, including the human tendency to reason by analogy. We ask: is this absorbed…arXiv cs.CL · 4 h ago · reads at arXiv cs.CL

This page holds the headline and a short snippet as published by arXiv cs.CL, whose feed they came from. The story itself, the photography and the rights stay with them, and the button above goes straight to their site. We keep this page so the headline has a stable address and a record of when it ran.