AI Paper Atlas

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

OffTopicEval: When Large Language Models Enter the Wrong Chat, Almost Always!

Jingdi Lei, Varun Gumma, Rishabh Bhardwaj, Seok Min Lim, Chuan Li, Amir Zadeh, Soujanya Poria

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
turn direct OOD queries into adaptive OOD queries via prompt laundering20 open-weight LLM agents · OffTopicEvalOOD refusal rate: 12.24% failure on direct -> 70.72% failure on adaptiveHurt
enable thinking mode in Qwen-3 (ablation)Qwen-3 (0.6B-32B) · OffTopicEval Englishoperational safety (OS): 58.09 -> 44.17 averageHurt
enable thinking mode in Qwen-3 (30B) (ablation)Qwen-3 (30B) · OffTopicEval Englishoperational safety (OS): 65.10 -> 88.64Improved
add Q-ground (rewrite query minimally, then respond) suffixopen-weight LLM agents · OffTopicEval Englishoperational safety (OS): Llama-3.3 53.92 -> 77.25; Gemma-3 63.92 -> 57.50Mixed
add P-ground (focus on system prompt) suffixopen-weight LLM agents · OffTopicEval Englishoperational safety (OS): Llama-3.3 53.92 -> 94.99; Gemma-3 63.92 -> 51.19Mixed
add P-ground to Mistral (24B) on Chinese and HindiMistral (24B) · OffTopicEval Chinese/Hindioperational safety (OS): 84.02 -> 46.70 (Chinese); 79.05 -> 31.32 (Hindi)Hurt
add P-ground or Q-ground to closed-weight modelsGPT-5, Gemini 2.5 Flash-Lite · OffTopicEval Englishoperational safety (OS): GPT-5 89.04 -> 91.89; Flash-Lite 79.90 -> 71.44 (P-ground)Mixed
add activation steering using pos-sys vs neg-sys vectorMinistral 8B, Phi-4, Qwen3 4B · OffTopicEval English (bankhelper, medischeduler)operational safety (OS): +2.60 to +17.09Improved
LoRA SFT as ID/OOD discriminator on 2,000 samplesGemma-3 (27B), Qwen-3 (4B) · OffTopicEval Englishoperational safety (OS): Gemma 63.78 -> 58.20; Qwen-3 4B 50.92 -> 8.89Hurt
make system prompt longer (P1 to P8 cumulative paragraphs) (ablation)Llama-3.3 (70B) · OffTopicEval Englishdirect OOD refusal rate: 18.38 -> 70.69 (peak 84.64 at P6)Improved
scale Qwen-3 from 0.6B/1.7B to larger sizes (non-thinking) (ablation)Qwen-3 · OffTopicEval Englishoperational safety (OS): 17.40 (1.7B) -> 77.77 (235B)Improved
preceding K in-domain queries before an adaptive OOD query (ablation)Llama, Phi-4, Gemma, Mistral · OffTopicEval BankHelperOOD refusal rateHurt

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