AI Paper Atlas

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

SocialHarmBench: Revealing LLM Vulnerabilities to Socially Harmful Requests

Punya Syon Pandey, Lê Hải Sơn, Devansh Bhardwaj, Zhijing Jin

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Safety alignment of LLMs

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
evaluate on sociopolitical prompts instead of generic harmopen-weight LLMs · SocialHarmBenchHarmBench ASR: Mistral-7B 62.50 on historical revisionismHurt
compare closed vs open-weight modelsClaude-Sonnet-4 vs Mistral-7B-Instruct · SocialHarmBench combinedHarmBench ASR: 27.71 -> 0.78Improved
apply weight-tampering attackopen-weight LLMs · SocialHarmBenchHarmBench score: above 0.90 across nearly all modelsHurt
apply latent-space and input-space attacks (AutoDAN, GCG, SoftOpt, LAT)open-weight LLMs · SocialHarmBenchHarmBench scoreMixed
prompts set in the 21st centuryLLMs · SocialHarmBenchHB / SR: HB=0.67, SR=0.50Hurt
prompts tied to Latin America, USA, UKMistral, Gemma, DeepSeek · SocialHarmBenchHB: HB 0.50-0.75Hurt

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