Know Your Limits: A Survey of Abstention in Large Language Models

• LLMs can produce false, overconfident, incomplete, or harmful responses, motivating abstention under uncertainty.Source: §1 · p. 1
• Prior work shows abstention’s potential in constrained settings, while the survey covers scenarios faced by general-purpose chatbots in open-domain interactions.Source: §1 · p. 1
• The survey categorizes methods by pretraining, alignment, and inference, and reviews evaluation benchmarks and metrics.Source: §1 · p. 2
• The framework examines query answerability, model confidence, and human-value alignment of queries and responses.Source: §1 · p. 2
• The authors identify under-explored issues and future directions, and encourage robust abstention mechanisms demonstrated in real-world applications.Source: §1 · p. 2

2 Abstention in LLMs

• Abstention spans uncertainty, conflicting conclusions, harm-related refusals, and partial replies; expressions include inability, refutation, multiple perspectives, risk disclaimers, or uncertainty.Source: §2 · p. 2

2.1 Abstention Framework

• The framework considers query ambiguity or answerability, model knowledge and confidence, and whether queries or responses align with human values and norms.Source: §2.1 · pp. 2–3

2.2 Problem Formulation

• The formalization maps a query and response to refusal score r, combining answerability, model confidence, and human-value alignment; any perspective indicating full abstention triggers it, otherwise abstention may be partial.Source: §2.2 · p. 3
• A designer-chosen connector M integrates the three perspectives, with application-dependent options such as weighted averages, logical operations, or thresholds.Source: §2.2 · p. 3

2.3 Inclusion in this Survey

• The survey selects work under any framework perspective, organizes methods by the LLM lifecycle and perspective, reviews benchmarks and metrics, and discusses research outside method or evaluation categories.Source: §2.3 · p. 3

3.1 Pretraining Stage

• No existing research studies abstention during pretraining; proposed directions include refusal-aware data analysis and staged domain-aware, curricular, loss-based, or uncertainty-regularization methods.The proposed methods include adding domain and abstention-reason information, progressing from simple to complex scenarios, and balancing token prediction with abstention.Source: §3.1 · pp. 3–4

3.2 Alignment Stage

• Abstention-aware instruction tuning can improve abstention but cause over-abstention, while researchers disagree on whether it generalizes as a task-independent meta-capability.The section proposes more varied abstention expressions, definitions, and domains, potentially combined through ensembles such as mixtures of LoRA experts, to improve generalization.Source: §3.2 · p. 5
• Preference optimization can reduce instruction-tuning over-abstention by aligning answers and refusals with preferences, but reward models or preference data that overvalue safety can preserve that risk.Because abstention spans degrees of behavior, ranking-based preference optimization is proposed as an alternative to pairwise comparisons.Source: §3.2 · pp. 5–6

3.3 Inference Stage

• Inference methods operate at input, in-processing, or output stages, targeting query answerability and human values, model knowledge, or both, respectively.The section groups methods by when they are applied.Source: §3.3 · p. 6

3.3.1 Input-processing Approaches

• Input-processing methods assess query ambiguity or alignment with human values, but existing settings often overlook how earlier conversational context affects later judgments.The discussed approaches include predicting ambiguity and detecting malicious or harmful queries.Source: §3.3.1 · p. 6

3.3.2 In-processing Approaches

• In-processing methods use internal representations to assess knowledge boundaries and safety; uncertainty estimates can guide abstention, but probability and verbal-confidence measures may be miscalibrated, and calibration faces domain-shift challenges.Calibration methods can improve output reliability and abstention, but their effectiveness across datasets and models remains a research concern.Source: §3.3.2 · pp. 6–7
• Other in-processing methods estimate uncertainty by aggregating response consistency or prompting abstention, while interpretable instruction and demonstration design and the limits of in-context learning remain underexplored.Consistency approaches compare multiple responses; prompting approaches use examples, hints, and protective prompts.Source: §3.3.2 · pp. 7–8

3.3.3 Output-processing Approaches

• Self-evaluation methods ask LLMs to judge whether responses are certain or safe and abstain when they are not; Chen et al. (2023b) learn self-evaluation parameters with Soft Prompt Tuning.Source: §3.3.3 · p. 8
• Kim et al. (2024a) use iterative self-feedback to refine responses and improve safety in non-safety-aligned LLMs, while Wang et al. (2024d) report that harmful/harmless tagging may encourage over-abstention.Source: §3.3.3 · p. 8
• LLM collaboration supports abstention through test models that detect harmful queries or correct compliant answers, and larger systems that use expert models and judges (Wang et al., 2024b; Pisano et al., 2023; Feng et al., 2024b; Chen et al., 2023a; Zeng et al., 2024).Source: §3.3.3 · p. 8

4.1 Evaluation Benchmarks

• Benchmarks cover unanswerable or ambiguous queries and model knowledge gaps, including questions about current events and long-tail entities.Examples include SQuAD2, Natural Questions, and MuSiQue for query answerability, and RealTimeQA and POPQA for knowledge boundaries.Source: §4.1 · p. 9, Model knowledge-centric abstention… · p. 9

Model knowledge-centric abstention datasets

• Safety benchmarks range from toxicity-focused datasets to broad safety taxonomies, with WildGuard explicitly evaluating refusal performance.Examples include RealToxicityPrompts and ToxiGen, as well as Do-Not-Answer, XSafety, SALAD-Bench, and SORRY-Bench.Source: Model knowledge-centric abstention… · p. 9

4.2 Evaluation metrics

• No single statistical metric captures all performance aspects, and existing automated methods focus on full rather than partial abstention.The survey recommends balancing task and abstention performance, coverage, and accuracy-coverage trade-offs; partial abstention can be scored by weighting answering and abstention accuracy.Source: §4.2 · pp. 10–11
• Model-based evaluations focus particularly on safety, while a generalized framework spanning multiple perspectives remains a future need.GPT-4-level off-the-shelf judgments can agree well with humans but incur high financial and time costs.Source: §4.2 · p. 11
• Human evaluations examine perceptions of denial styles and response usefulness; participants generally prefer diverting denial, while first-person uncertainty phrases reduce confidence in reliability and acceptance.Current human evaluations focus on full abstention and strong expressions, leaving partial abstention and other expression types underexplored.Source: §4.2 · p. 11

Over-abstention

• Excessive abstention can make models refuse benign inputs, so abstention must be balanced with utility.The section cites self-check and abstention-focused instruction tuning as examples of over-caution.Source: §5 · p. 11

Vulnerability of abstention

• Prompt wording and manipulation can bypass safety-driven abstention, while some safety interventions can inadvertently undermine alignment.Reported tactics include persuasive or persona-based prompts, cipher-based communication, and low-resource-language translations.Source: §5 · pp. 11–12

Introducing biases

• Abstention varies across demographic groups and languages, raising concerns that some communities may be underserved.Examples include more abstention for African American English than White American English, and less abstention for election predictions about Africa and Asia.Source: §5 · p. 12

Following up after abstention

• After refusing, a system can seek information through user questions or retrieval, or reformulate an unanswerable question into one answerable from a document.Zhao et al. (2024a) study the latter task.Source: §5 · p. 12

Personalized abstention

• Personalized abstention could adapt to user profiles, tolerance for conservative responses, interaction history, and query-specific needs.The section notes that preferences differ and no one-size-fits-all solution exists (Kirk et al., 2023b).Source: §5 · p. 12

Future research agenda

• Future work could test abstention as a meta-capability across tasks and develop more generalizable evaluation and customization for user needs.The survey also calls for research into task- or domain-specific abstention abilities.Source: §6 · p. 12
• Abstention-aware designs could address privacy and copyright protections, and research could extend abstention beyond LLMs to vision, vision-language, and generative applications.Source: §6 · p. 12
• The survey calls for improved multilingual abstention because performance discrepancies exist between high- and low-resource languages.Source: §6 · p. 12

Conclusion

• The survey frames abstention around the query, the model, and human values, and reviews strategies across stages of LLM development.Source: §7 · p. 12
• Its review of literature, benchmarking datasets, and evaluation metrics identifies gaps and limitations in current methodologies.Source: §7 · p. 12
• The conclusion calls for broader applications and more dynamic contexts, with adaptive abstention balanced against helpfulness.Source: §7 · p. 12