REVIEW 42 cited by
SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal
read the original abstract
Evaluating aligned large language models' (LLMs) ability to recognize and reject unsafe user requests is crucial for safe, policy-compliant deployments. Existing evaluation efforts, however, face three limitations that we address with SORRY-Bench, our proposed benchmark. First, existing methods often use coarse-grained taxonomies of unsafe topics, and are over-representing some fine-grained topics. For example, among the ten existing datasets that we evaluated, tests for refusals of self-harm instructions are over 3x less represented than tests for fraudulent activities. SORRY-Bench improves on this by using a fine-grained taxonomy of 44 potentially unsafe topics, and 440 class-balanced unsafe instructions, compiled through human-in-the-loop methods. Second, linguistic characteristics and formatting of prompts are often overlooked, like different languages, dialects, and more -- which are only implicitly considered in many evaluations. We supplement SORRY-Bench with 20 diverse linguistic augmentations to systematically examine these effects. Third, existing evaluations rely on large LLMs (e.g., GPT-4) for evaluation, which can be computationally expensive. We investigate design choices for creating a fast, accurate automated safety evaluator. By collecting 7K+ human annotations and conducting a meta-evaluation of diverse LLM-as-a-judge designs, we show that fine-tuned 7B LLMs can achieve accuracy comparable to GPT-4 scale LLMs, with lower computational cost. Putting these together, we evaluate over 50 proprietary and open-weight LLMs on SORRY-Bench, analyzing their distinctive safety refusal behaviors. We hope our effort provides a building block for systematic evaluations of LLMs' safety refusal capabilities, in a balanced, granular, and efficient manner. Benchmark demo, data, code, and models are available through https://sorry-bench.github.io.
Forward citations
Cited by 42 Pith papers
-
RefusalBench: Why Refusal Rate Misranks Frontier LLMs on Biological Research Prompts
RefusalBench shows strict refusal rates fail to rank frontier LLMs correctly on biological safety, with provider effects and partial-compliance patterns that binary metrics miss.
-
VoxSafeBench: Not Just What Is Said, but Who, How, and Where
VoxSafeBench reveals that speech language models recognize social norms from text but fail to apply them when acoustic cues like speaker or scene determine the appropriate response.
-
Agentic Abstention: Do Agents Know When to Stop Instead of Act?
LLM agents often fail to abstain at the right time in uncertain multi-turn tasks, and the CONVOLVE context engineering method raises timely abstention rates on WebShop from 26.7 to 57.4 without parameter updates.
-
Reliable to Expressive: A Curriculum for Rubric-Following Safety Judges
A reliable-to-expressive curriculum with dynamic rubrics trains a 12B safety judge to achieve 94%+ accuracy with only 0.76 cross-rubric variance on three different rubric prompts.
-
LLM-Based Persuasion Enables Guardrail Override in Frontier LLMs
LLM attackers persuade frontier LLMs to generate prohibited essays on consensus topics through multi-turn natural-language pressure, with success rates up to 100% in some model-topic pairs.
-
Sparse Tokens Suffice: Jailbreaking Audio Language Models via Token-Aware Gradient Optimization
TAGO performs sparse jailbreak optimization on audio LMs by retaining only high-gradient-energy tokens, preserving near-full ASR at 25% retention across three models.
-
Sparse Tokens Suffice: Jailbreaking Audio Language Models via Token-Aware Gradient Optimization
Sparse selection of high-gradient-energy audio tokens suffices for effective jailbreaking of audio language models with minimal drop in attack success rate.
-
FinSafetyBench: Evaluating LLM Safety in Real-World Financial Scenarios
FinSafetyBench shows that LLMs remain vulnerable to adversarial prompts that bypass financial compliance safeguards, with notably higher failure rates in Chinese-language scenarios.
-
IatroBench: Pre-Registered Evidence of Iatrogenic Harm from AI Safety Measures
Frontier models systematically withhold clinically necessary guidance from laypeople that they provide to physicians on identical facts, and LLM judges fail to detect that omission harm.
-
Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents
ASB is a new benchmark that tests 10 prompt injection attacks, memory poisoning, a novel Plan-of-Thought backdoor attack, and 11 defenses on LLM agents across 13 models, finding attack success rates up to 84.3% and li...
-
AISPA: User-Centric System Prompt Auditing for Large Language Model Applications
A human-in-the-loop audit of system prompts from 88 commercial AI products finds protective instructions nearly universal yet incomplete, with ~40% of products containing at least one user-harmful directive.
-
Architectural Backdoors in Vision-Language Model Supply Chains via Representation Steering
Trigger-gated additive steering vectors embedded in VLM architecture definitions create dormant backdoors that work across VQA, text-to-image, retrieval, and brand/safety biasing without data poisoning.
-
BioTIER: A Refusal Benchmark for Targeted Biological Risk Mitigation
BioTIER, a 542-prompt benchmark with three risk tiers, shows frontier AI models differ by 90 percentage points in refusing dangerous biological queries, with top refusers over-refusing benign topics at the boundary.
-
Not All Refusals Are Equal: How Safety Alignment Fails Cybersecurity at Scale
Domain-specific abliteration removes cybersecurity refusal on Kimi K2 (100%→7%) while preserving most other safety refusals; susceptibility depends mainly on safety-training method and architecture, not size.
-
Addressing Over-Refusal in LLMs with Competing Rewards
SEAR trains one LLM via adversarial process rewards to explore harmful reasoning paths but flip to safe outputs, reducing over-refusal while preserving safety.
-
Do Encoders Suffice? A Systematic Comparison of Encoder and Decoder Safety Judges for LLM Adversarial Evaluation
Fine-tuned ModernBERT-family encoders match LLM judges on F1, false negative rate, and precision-recall for harmful output detection across adversarial datasets and attack types while promising lower cost and latency.
-
What Do Safety-Aligned LLMs Learn From Mixed Compliance Demonstrations?
Safety-aligned LLMs treat benign and harmful compliance demonstrations differently in in-context learning, with preference optimization preventing benign examples from increasing harmful compliance and strong recency ...
-
Sch\"utzen: Evaluating LLM Safety in Bulgarian and German Contexts
Schützen is a German-Bulgarian LLM safety dataset showing pronounced cross-language differences in model safety behavior.
-
Efficient Safety Benchmarking via Item Response Theory
Item Response Theory enables adaptive and fixed-subset item selection that reduces safety benchmark costs by 80-99.9% while preserving high correlation with full rankings.
-
Chain-of-Models: Cross-Model Auditing for Bias-Robust LLM Judges
Per-bias selection of a cross-family LLM auditor lifts biased-judgment accuracy from 0.805/0.824 baselines to 0.884.
-
Before the Last Token: Diagnosing Final-Token Safety Probe Failures
Final-token probes miss distributed unsafe evidence in jailbreaks, but a PCA-HMM model on prefill trajectories recovers many misses without naive pooling's false positives.
-
Few-Shot Truly Benign DPO Attack for Jailbreaking LLMs
A truly benign DPO attack using 10 harmless preference pairs jailbreaks frontier LLMs by suppressing refusal behavior, achieving up to 81.73% attack success rate on GPT-4.1-nano at low cost.
-
Revisiting JBShield: Breaking and Rebuilding Representation-Level Jailbreak Defenses
JBShield is vulnerable to adaptive JB-GCG attacks (up to 53% ASR) because jailbreak representations occupy a distinct region in refusal-direction space; the new RTV defense using Mahalanobis detection on multi-layer f...
-
MultiBreak: A Scalable and Diverse Multi-turn Jailbreak Benchmark for Evaluating LLM Safety
MultiBreak is a large diverse multi-turn jailbreak benchmark that achieves substantially higher attack success rates on LLMs than prior datasets and reveals topic-specific vulnerabilities in multi-turn settings.
-
PERSA: Reinforcement Learning for Professor-Style Personalized Feedback with LLMs
PERSA combines RLHF with selective parameter-efficient updates to top transformer layers, raising style alignment scores from 35% to 96% on code feedback benchmarks while holding correctness near 100%.
-
Reasoning Structure Matters for Safety Alignment of Reasoning Models
Changing the internal reasoning structure of large reasoning models through simple supervised fine-tuning on 1K examples produces strong safety alignment that generalizes across tasks and languages.
-
IatroBench: Pre-Registered Evidence of Iatrogenic Harm from AI Safety Measures
AI models exhibit identity-contingent withholding, providing better clinical guidance on benzodiazepine tapering to physicians than laypeople in identical scenarios, with a measured decoupling gap of +0.38 and 13.1 pe...
-
Blind Refusal: Language Models Refuse to Help Users Evade Unjust, Absurd, and Illegitimate Rules
Language models refuse 75.4% of requests to evade defeated rules and do so even after recognizing reasons that undermine the rule's legitimacy.
-
MetaBreak: Jailbreaking Online LLM Services via Special Token Manipulation
Special tokens that structure LLM conversations can be injected and swapped for lookalike words to bypass both built-in safety and external content filters.
-
YouthSafe: A Youth-Centric Safety Benchmark and Safeguard Model for Large Language Models
Introduces YAIR, a youth-GenAI risk benchmark, and YouthSafe, a fine-tuned classifier with AUPRC 0.94 on it, though it compares a trained model to untrained baselines.
-
Beyond I'm Sorry, I Can't: Dissecting Large Language Model Refusal
Sparse autoencoders plus greedy filtering and factorization-machine interaction modeling identify minimal sets of features in Gemma-2-2B-IT and LLaMA-3.1-8B-IT whose ablation produces jailbreaks by flipping refusal to...
-
How Jailbreak Attacks Inform Safety Alignment: A Defender-Centric, Shapley-Based Evaluation of Jailbreak Contributions
Attacks that break LLMs best are not the ones that improve safety most; a Shapley- and greedy-based framework that selects attack subsets by downstream defender utility outperforms attacker-centric and attribution-onl...
-
PHANTOM: A Large-Scale Dataset of Multimodal Adversarial Attacks for Vision-Language Models
PHANTOM is a consolidated open-source dataset of 47,524 multimodal adversarial samples for VLMs, extending prior benchmarks across 10 high-level categories and 55 subcategories of harmful intents.
-
Fence: Specialized SLM Guardrails for LLM Applications
Small language models trained on GAN-style adversarial synthetic data outperform prompted large models on use-case-specific guardrails for off-topic and prompt-injection detection.
-
Before the Last Token: Diagnosing Final-Token Safety Probe Failures
Final-token safety probes miss jailbreaks whose unsafe evidence appears earlier in prefill; a clean-trained PCA-HMM trajectory model recovers many of those misses without catastrophic false positives.
-
A Validated Prompt Bank for Malicious Code Generation: Separating Executable Weapons from Security Knowledge in 1,554 Consensus-Labeled Prompts
The paper releases a 1,554-prompt consensus-labeled bank separating executable malicious code requests from security knowledge requests, validated by five-model majority labeling with Fleiss' kappa of 0.876.
-
MEDIC-AD: Towards Medical Vision-Language Model's Clinical Intelligence
MEDIC-AD adds anomaly-aware and difference tokens to a medical VLM, claiming SOTA lesion detection, temporal tracking, and visual grounding; the zero-shot claim is undermined by likely train/test overlap.
-
$C$-$\Delta\Theta$: Circuit-Restricted Weight Arithmetic for Selective Refusal
Selective refusal can be improved by editing only the small weight circuit found by EAP-IG, yielding offline checkpoints with low over-refusal.
-
Discriminatory Compliance: How LLMs Answer Queries from Protected Groups
State-of-the-art LLMs respond inconsistently to queries from protected-group personas, with some responses omitting key information that should be provided.
-
Opir: Efficient Multi-Task Safety Classification for Toxicity, Jailbreaks, Hate Speech, and Harmful Content
Opir introduces efficient multi-task encoder models trained on a 996-category safety taxonomy that match or exceed larger baselines on most safety benchmarks while using under 100M parameters for edge variants.
-
A Survey on LLM-as-a-Judge
A survey on LLM-as-a-Judge that reviews reliability strategies, proposes evaluation methods, and introduces a novel benchmark for assessing such systems.
-
LLMs-as-Judges: A Comprehensive Survey on LLM-based Evaluation Methods
A survey that organizes LLMs-as-judges research into functionality, methodology, applications, meta-evaluation, and limitations.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.