| Challenge: | Several prior studies test the safety promises of language models based on their effectiveness in blocking malicious requests. |
| Approach: | They create a benchmark based on benign concepts and ground them in a knowledge graph to evaluate abstention techniques. |
| Outcome: | The proposed framework causes models to abstain with over 80% abstention rates, but not as effective for descendants of the target concepts, where abstraction rates drop by 19%. |
Similar Papers
Knowing When to Abstain: Medical LLMs Under Clinical Uncertainty (2026.eacl-long)
Copied to clipboard
Sravanthi Machcha, Sushrita Yerra, Sahil Gupta, Aishwarya Sahoo, Sharmin Sultana, Hong Yu, Zonghai Yao
| Challenge: | Current evaluation of large language models prioritizes accuracy, but abstention is vital for trustworthy deployment. |
| Approach: | They propose a benchmark and evaluation protocol for abstention in medical multiple-choice question answering . they integrate conformal prediction, adversarial question perturbations, and explicit abstraction options. |
| Outcome: | The proposed protocol improves reliability of medical multiple-choice question answering models by providing explicit abstention options. |
Post-Abstention: Towards Reliably Re-Attempting the Abstained Instances in QA (2023.acl-long)
Copied to clipboard
| Challenge: | Despite remarkable progress made in natural language processing, even the state-of-the-art systems often make incorrect predictions. |
| Approach: | They propose to use selective prediction to enable models to abstain from answering when their predictions are likely to be incorrect. |
| Outcome: | The proposed method improves performance on 11 QA datasets and in- and out-of-domain settings. |
StealthGraph: Exposing Domain-Specific Risks in LLMs through Knowledge-Graph-Guided Harmful Prompt Generation (2026.acl-long)
Copied to clipboard
| Challenge: | Domain-specific datasets of harmful prompts are scarce and often rely on manual construction. Existing efforts to improve domain knowledge and reduce harmful prompt generation are lacking. |
| Approach: | They propose a framework that transforms domain knowledge into actionable constraints and increases the implicitness of generated harmful prompts. |
| Outcome: | The proposed framework yields high-quality datasets combining strong domain relevance with implicitness, enabling more realistic red-teaming and advancing LLM safety research. |
Answering the Wrong Question: Reasoning Trace Inversion for Abstention in LLMs (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models have been shown to have worse abstention abilities than reasoning models . a new class of abstraction methods is developed to improve absttention performance . |
| Approach: | They propose a class of abstention methods that generate reasoning trace and reconstruct most likely query from it. |
| Outcome: | The proposed method beats baselines in 33 out of 36 settings. |
When to Speak, When to Abstain: Contrastive Decoding with Abstention (2025.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) demonstrate exceptional performance across diverse tasks by leveraging pre-trained (parametric) and external (contextual) knowledge. |
| Approach: | They propose a training-free decoding method that allows LLMs to generate responses when relevant knowledge is available and to abstain otherwise. |
| Outcome: | The proposed method can generate responses when relevant knowledge is available and abstain otherwise. |
Do-Not-Answer: Evaluating Safeguards in LLMs (2024.findings-eacl)
Copied to clipboard
| Challenge: | a dataset evaluating harmful capabilities in large language models is available at https://github.com/Libr-AI/do-not-answer. |
| Approach: | They collect an open-source dataset to evaluate the safeguards in large language models . they find that simple BERT-style classifiers can achieve results comparable to GPT-4 . |
| Outcome: | The proposed dataset compares the safety of six popular LLMs to GPT-4 on automatic safety evaluation. |
Interpretation Meets Safety: A Survey on Interpretation Methods and Tools for Improving LLM Safety (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing surveys focus on interpretation or safety, but safety and understanding are core motivations for interpretation research. |
| Approach: | They propose a framework that connects interpretation methods, enhancements they inform, and tools that operationalize them. |
| Outcome: | The proposed framework summarizes nearly 70 studies at their intersections and concludes with open challenges and future directions. |
Deconstructing NLG Evaluation: Evaluation Practices, Assumptions, and Their Implications (2022.naacl-main)
Copied to clipboard
| Challenge: | Evaluating natural language generation systems is difficult, as there are many ways to express similar things in text. |
| Approach: | They combine interviews with NLG practitioners to examine ethical considerations and their implications for NLG evaluation. |
| Outcome: | The findings of the study surface goals, community practices, assumptions, and constraints that shape NLG evaluations, and examine their implications and how they embody ethical considerations. |
Can Knowledge Graphs Make Large Language Models More Trustworthy? An Empirical Study Over Open-ended Question Answering (2025.acl-long)
Copied to clipboard
| Challenge: | Existing benchmarks for integrating Knowledge Graphs with Large Language Models focus on closed-ended tasks, leaving a gap in evaluating performance on more complex, real-world scenarios. |
| Approach: | They propose a benchmark to evaluate LLMs augmented with KGs in open-ended, real-world question answering settings. |
| Outcome: | The proposed benchmark reflects practical complexities through diverse question types and incorporates metrics to quantify both hallucination rates and reasoning improvements in LLM+KG models. |
SMATCH++: Standardized and Extended Evaluation of Semantic Graphs (2023.findings-eacl)
Copied to clipboard
| Challenge: | Existing graph-alignment metrics that measure graph distances are not reliable, we show . metric is spread out and does not provide upper bounds for extended tasks. |
| Approach: | They propose a metric to measure a distance between graphs by aligning nodes and counting matching graph triples. |
| Outcome: | The proposed method reduces search space and improves scoring by reducing the number of errors. |