dieferna/afgen_results

This repository contains evaluation results accompanying the paper:

"LLM alignment failures detection leveraging inter-behavioral generalization using simple assertions"
Submitted to Information Sciences (Elsevier, ScienceDirect).

Dataset Name Mapping

This folder contains evaluation results using original dataset names. Use the table below to map the folder/file names to their publication aliases:

Folder / File Pattern	Dataset Alias
animals_stmt	az_ani
cities_stmt	az_cit
companies_stmt	az_com
elements_stmt	az_ele
generated_stmt	az_gen
inventions_stmt	az_inv
facts_stmt	az_fact
science_stmt	o_sci
madrid_stmt	o_mad
code_short_stmt	o_codeS
code_long_stmt	o_codeL
generated_facts_8B_stmt__last_token	o_fact
cot_8B	o_cot
jbb_harmful_benign_prompt	jbb
toxigen_hate_neutral	toxi
white_lies	o_wLies
sycophancy_just_AT_AF	o_agSyco
sycophancy_feedback	o_feSyco

---

Overview

This project studies LLM misalignment detection using Linear Probes trained on simple true–false assertions. We show that these probes generalize across multiple alignemnt-failing behaviours—including hallucinations, harmful outputs, bias, and deception—without being trained on alignemnt-failing samples.

---

Status

This repository accompanies a research paper currently under submission.

---

license: apache-2.0