PaperlayerView exact revision on arXiv

Source-linked AI summary

AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions

Polina Kirichenko, Mark Ibrahim, Kamalika Chaudhuri, Samuel J. Bell

arXiv:2506.09038v1cs.AI

TL;DR

LLMs need to know when not to answer, but abstention lacks a systematic evaluation framework for diverse uncertain queries. AbstentionBench addresses this gap by benchmarking 20 frontier models across 20 datasets and finds that abstention remains unsolved, with reasoning fine-tuning reducing abstention by 24% on average. The results motivate research explicitly targeting abstention and uncertainty reasoning.

  • Problem

    Abstention is understudied despite being necessary for handling underspecified, ambiguous, and unanswerable queries reliably.

  • Method

    AbstentionBench evaluates 20 frontier LLMs across diverse datasets, including curated scenarios and underspecified math and science benchmarks.

  • Results

    Abstention remains an unsolved problem, while reasoning fine-tuning causes a 24% average abstention drop compared with non-reasoning counterparts.

  • Takeaways & Limitations

    Improving reliability may require post-training methods that explicitly target abstention and reasoning about uncertainty.

  • Takeaways & Limitations

    Current LLMs, including reasoning models, still struggle with abstention across the benchmark’s uncertainty scenarios.

Abstract

from arXiv · show

For Large Language Models (LLMs) to be reliably deployed in both everyday and high-stakes domains, knowing when not to answer is equally critical as answering correctly. Real-world user queries, which can be underspecified, ill-posed, or fundamentally unanswerable, require LLMs to reason about uncertainty and selectively abstain -- i.e., refuse to answer definitively. However, abstention remains understudied, without a systematic evaluation framework for modern LLMs. In this work, we introduce AbstentionBench, a large-scale benchmark for holistically evaluating abstention across 20 diverse datasets, including questions with unknown answers, underspecification, false premises, subjective interpretations, and outdated information. Evaluating 20 frontier LLMs reveals abstention is an unsolved problem, and one where scaling models is of little use. While recent reasoning LLMs have shown impressive results in complex problem solving, surprisingly, we find that reasoning fine-tuning degrades abstention (by $24\%$ on average), even for math and science domains on which reasoning models are explicitly trained. We find that while a carefully crafted system prompt can boost abstention in practice, it does not resolve models' fundamental inability to reason about uncertainty. We release AbstentionBench to foster research into advancing LLM reliability.

1 Introduction

Abstention is essential for reliable LLMs because many real-world questions are ambiguous, underspecified, or unanswerable. AbstentionBench evaluates this capability across diverse scenarios and finds that reasoning interventions can worsen abstention.

  • Motivation: Reliable LLMs must recognize when available information is insufficient and avoid giving definitive answers.Appropriate responses may express uncertainty, ask for clarification, or state that the answer is unknown.
  • Motivation: Existing abstention research lacks a holistic benchmark covering the diverse uncertainty conditions found in open-ended dialogue.These include vague or underspecified questions, unknown answers, and false premises.
  • Benchmark: AbstentionBench combines 17 curated datasets spanning six scenarios with modified math and science benchmarks containing underspecified questions.The benchmark also evaluates 20 frontier LLMs using quality-verified automatic scoring.
  • Findings: 24% average abstention drop is observed for reasoning models compared with non-reasoning counterparts.The degradation persists in math and science domains and can involve hallucinating missing context despite uncertainty in reasoning traces.
  • Implications: AbstentionBench identifies abstention as a fundamental gap and motivates post-training methods that explicitly target uncertainty.The authors suggest incorporating diverse uncertain scenarios into post-training and reasoning fine-tuning.

2 Related work

Prior work studies abstention through narrow datasets, uncertainty elicitation, hallucination mitigation, and reasoning benchmarks. AbstentionBench differs by evaluating direct abstention across diverse scenarios rather than focusing only on safety, factuality, or correctness.

  • Abstention evaluation: Existing abstention datasets typically target a single problem type, limiting evaluation across diverse uncertainty scenarios.Examples include unanswerable questions, missing-answer multiple-choice items, and underspecification.
  • Uncertainty elicitation: Verbalized uncertainty has shown limited performance and generalization, although prompting and fine-tuning can improve related calibration or uncertainty expression.These approaches differ from AbstentionBench’s direct, out-of-the-box evaluation of abstention.
  • Scope distinction: Compliance research mainly evaluates refusals grounded in policy, safety, or copyright, whereas this work targets questions that cannot be answered definitively.The benchmark also expands the scale relative to the 1k-prompt CoCoNot dataset.
  • Reasoning and mathematics: Reasoning research has emphasized correctness on narrow domains with clear answers, leaving reasoning about uncertainty less understood.Most mathematical evaluations likewise focus on answerable problems, while newer work examines unanswerable or unsolvable cases.

3 AbstentionBench: Benchmarking LLM Abstention

AbstentionBench is a large-scale benchmark that evaluates whether LLMs abstain appropriately across diverse scenarios, datasets, models, and evaluation measures. It combines curated general, math, and science datasets with automatic judging to support systematic comparisons.

  • Benchmark definition: AbstentionBench defines abstention broadly as refraining from directly answering through uncertainty, caveats, lack of knowledge, or partial answers.The benchmark includes responses such as “I don’t know” as well as detailed answers that address only answerable aspects.
  • Dataset construction: It adds unanswerable variants of GPQA, GSM8K, and MMLU-Math by removing context required to answer math and science questions, alongside UMWP.The resulting datasets mix answerable and unanswerable questions for reasoning-heavy evaluation.
  • Scenario coverage: The benchmark curates datasets spanning answer-unknown, false-premise, stale, subjective, underspecified-context, and underspecified-intent scenarios.These scenarios are explicitly described as neither exhaustive nor mutually exclusive.
  • Automatic evaluation: A Llama 3.1 8B Instruct judge labels abstention and non-abstention responses, achieving 88% accuracy on a manually annotated sample.The same judge is also used to assess correctness for non-abstaining responses, while recall, precision, and F1 measure abstention performance.
  • Analysis dimensions: The benchmark examines average and scale-related performance across models, as well as relationships between abstention recall and response correctness.Figures compare open and proprietary models, increasing Llama scale, Qwen’s scenario distribution, and dataset-level accuracy–abstention correlations.

4 Experiments

Experiments show that abstention remains difficult across frontier models, with scale offering little benefit. Post-training and reasoning interventions can improve accuracy while degrading abstention, although carefully crafted prompts provide practical gains without resolving uncertainty reasoning.

  • 4.1 Abstention is an open challenge for language models: Abstention remains challenging across most datasets, with performance ranging from near-perfect on Known Unknowns to near-zero on MediQ.No model consistently outperforms the others across all datasets.
  • 4.1 Abstention is an open challenge for language models: Increasing model scale has almost no effect on mean abstention performance across datasets.Large closed models perform relatively close to smaller Qwen 2.5 32B and Llama 3.1 8B models.
  • 4.1 Abstention is an open challenge for language models: Improved response accuracy does not necessarily imply improved abstention, with the relationship varying by dataset and uncertainty type.Accuracy correlates positively with abstention on underspecified math but with degraded abstention on outdated-information questions.
  • 4.2 Post-training instills select abstention capabilities: Post-training generally improves accuracy and abstention recall, but worsens abstention on underspecified contexts, especially during SFT.The RLVR stage produces a further degradation in abstention recall after training on math and verifiable instruction-following.
  • 4.3 Reasoning degrades abstention: Reasoning models exhibit worse abstention than their underlying instruction-tuned models, including on math and science reasoning datasets.For s1, accuracy improves while abstention declines, and models often hallucinate missing problem context.
  • 4.3 Reasoning degrades abstention: Increasing reasoning budgets improves accuracy while leaving abstention unchanged or worsening it, whereas system prompts can boost abstention without significantly reducing precision.Reasoning traces contain uncertainty expressions, but models still produce definitive final answers; prompt gains appear practical rather than fundamental.

5 Discussion

The discussion identifies a fundamental limitation in contemporary LLMs: they do not reliably know when not to answer. AbstentionBench frames abstention as a reliability goal beyond accuracy and motivates post-training methods that explicitly target uncertainty reasoning.

  • 5 Discussion: Contemporary LLMs do not reliably know when not to answer, despite strong performance on other capabilities.The paper presents this as a limitation of today’s best models.
  • 5 Discussion: AbstentionBench establishes abstention across diverse scenarios as a reliability goal beyond conventional accuracy evaluation.The benchmark systematically tests cases where models should abstain rather than respond.
  • 5 Discussion: Reasoning models can boost accuracy while degrading abstention, suggesting correctness-focused reward signals are insufficient for advancing reliability.The paper argues that explicitly teaching models to reason about evidence may be needed to determine when not to respond.

Supplementary materials outline

The supplementary materials provide limitations, broader implications, implementation details, additional results, and a fast benchmark subset.

  • Supplementary materials outline: Appendix A discusses study limitations, while Appendix B covers broader implications.
  • Supplementary materials outline: Appendix C documents models, hyperparameters, dataset review and filtering, benchmark implementation, and LLM judges.
  • Supplementary materials outline: Appendices D and E provide additional experimental results and a fast subset of the benchmark.

A Limitations

The study’s conclusions are bounded by its dataset coverage, possible benchmark leakage, finite model selection, and imperfect LLM-based judging.

  • A Limitations: The benchmark may omit abstention scenarios beyond those mined, filtered, or newly created, and it covers only English datasets.The authors identify broader scenario coverage and other languages as future directions.
  • A Limitations: Potential leakage between evaluation benchmarks and training data can confound abstention results.CoCoNot overlaps with Tülu post-training data, creating especially relevant overlap for OLMo and Tülu evaluations.
  • A Limitations: Findings are restricted to a finite selection of 20 models and therefore do not cover all contemporary model capabilities.The selected models span varied training paradigms and scales.
  • A Limitations: The LLM judge used to evaluate open-ended abstention and correctness is imperfect, despite high performance relative to human annotations.

B Broader impacts

AbstentionBench is an empirical evaluation of frontier LLM abstention, but its public-data design and LLM-judge reliance constrain interpretation and future benchmark validity.

  • Benchmark scope: AbstentionBench evaluates abstention across diverse scenarios using curated public datasets and automated LLM judging.The benchmark includes samples labeled for whether models should abstain, with optional reference answers for answerable questions.
  • Limitations: Because AbstentionBench uses only public data without a private or gated test set, dataset reuse may inflate performance estimates over time.The authors identify a private or gated test set as a possible direction for future work.
  • Dataset construction: The dataset-selection process searched 183 papers, shortlisted 82 datasets, and retained datasets after in-depth author review.Selection required public availability and abstention being appropriate for at least some samples.
  • Scenario coverage: The benchmark covers ambiguous moral choices, unanswerable questions, invalid assumptions, missing paper information, underspecification, and changing facts.It also includes false premises, unknown answers, ambiguous pronouns, unsupported questions, and missing medical context.
  • Evaluation procedure: LLM-judge examples did not improve judge quality, and annotated samples were split evenly between validation and testing.Final consensus labels served as ground truth for evaluating judge predictions.

C.3.4 Abstention detection judge evaluations

The benchmark uses LLM judges to evaluate abstention and correctness, finding strong apparent precision but substantial dataset variation in the relationship between abstention and correctness.

  • Judge selection: Llama 3.1 8B Instruct was selected as the abstention judge because it had the best recall while matching the other judges in overall accuracy.The choice prioritized efficiency and cost after evaluating three judges on human-annotated responses.
  • Metrics: Abstention precision is close to 1 for most models on most datasets, indicating that models rarely over-abstain.Abstention F1 balances precision and recall, and its model ranking mostly agrees with recall ranking.
  • Cross-dataset behavior: The correlation between response correctness and abstention recall varies substantially across datasets.This indicates that correctness and abstention recall do not show a uniform relationship across benchmark tasks.
  • Model comparison: Reasoning models rank among the strongest for average correctness but DeepSeek R1 Distill and s1.1 are near the weakest for abstention performance.Table 4 reports both average correctness and average abstention recall across datasets.
  • Scale: Increasing Llama 3.1 scale has limited effect across abstention F1, precision, recall, and response accuracy.The supplementary figure extends the scale analysis beyond abstention recall.

D.3 Effect of post-training

Post-training generally improves abstention for Llama models, although gains are uneven across scenarios and some training or task combinations degrade performance.

  • Tülu post-training: Tülu post-training generally improves Llama 3.1 8B abstention performance except on samples with underspecified contexts.Most improvements occur during supervised fine-tuning and direct preference optimization, while PPO RLVF degrades abstention performance.
  • Training stages: PPO RLVF degrades abstention performance after earlier gains from SFT and DPO.The reported pattern places the principal improvements in earlier post-training stages.
  • Scenario-specific effects: Tülu’s supplementary results show degraded precision for questions about stale data.The corresponding 70B analysis reports consistent results at larger scale.
  • Metrics: The supplementary comparisons report abstention F1, precision, recall, and response accuracy for post-training and instruction-tuning variants.These figures compare checkpoints or tuned models against Llama 3.1 base models.
  • Instruction tuning: Llama instruction tuning generally improves abstention recall at 8B, while underspecified contexts remain more difficult than unknown-answer or underspecified-intent cases.At 70B, underspecified-context and subjective questions remain challenging.

D.4 Effect of reasoning fine-tuning

The reasoning analyses examine fine-tuning, reasoning budgets, decoding, and judge inputs. Reasoning can improve answer accuracy in selected settings, but it often harms abstention recall and introduces control and evaluation trade-offs.

  • Decoding: Greedy decoding for s1.1 significantly improves accuracy on reasoning datasets.The reported inference setting uses temperature = 0.
  • Evaluation scope: Reasoning comparisons evaluate abstention recall, precision, and F1 across all datasets and reasoning datasets only.Supplementary figures pair reasoning models with corresponding instruction-tuned models.
  • Reasoning fine-tuning: Reasoning fine-tuning significantly harms abstention recall across all reasoning datasets.Accuracy gains are mixed: GPQA-Abstain improves significantly, while other datasets show minor gains or degradations.
  • Reasoning budget: Increasing reasoning budgets generally improves accuracy on GPQA-Abstain and MMLU-Math-Abstain, while models often use fewer tokens than the configured maximum.DeepSeek R1 Distill exits reasoning early on GSM8K-Abstain and UMWP, often below 512 tokens.
  • Generation control: DeepSeek R1 Distill can ignore termination controls and continue generating its reasoning chain, especially on GPQA-Abstain and MMLU-Math.The observed average reasoning length differs substantially from the set maximum without the trigger token.
  • Judge inputs: Using reasoning chains as well as final answers raises abstention recall but lowers precision because chains contain more uncertainty expressions.The comparison concerns DeepSeek R1 and s1.1 under two abstention-judge inputs.

E Code and reproducibility

AbstentionBench is released with reproducible tooling, a faster evaluation subset, and examples spanning appropriate abstentions and failures across scenarios.

  • Reproducible evaluation: The released AbstentionBench codebase reproduces evaluations across the full suite of datasets and supports measuring model abstention capabilities.The library includes scripts for downloading and preprocessing the required datasets and interfaces for running experiments across environments.
  • Fast subset: 4× faster inference is enabled by a fast subset containing 100 questions per benchmark, running in under 5 minutes on one NVIDIA Tesla V100 GPU with Llama 3.1 8B.
  • Fast subset: Abstention recall on the fast subset is within 5% of full-benchmark performance for most scenarios, except saturated stale samples.
  • Examples: The paper presents appropriate abstention examples from Qwen 2.5 32B Instruct and abstention failures from OLMo 7B Instruct and s1.1 32B.The examples cover general scenarios and failures on reasoning datasets.

F.1 Appropriate abstention from Qwen 2.5 32B

The examples show Qwen 2.5 32B Instruct abstaining appropriately across unknown-answer, subjective, underspecified, false-premise, and unsupported-query scenarios. They contrast calibrated responses with answers that invent facts, assume missing context, or claim unsupported capabilities.

  • Underspecified Intent: For underspecified intent, Qwen asks for the missing team identity instead of answering the NCAA tournament question as though the referent were known.
  • Answer Unknown: For unknown-answer questions, Qwen explains when the context lacks an answer and avoids asserting an event or date that is not established.Examples include an absent February university event and an unannounced 2050 Olympics host.
  • Answer Unknown: For current or outdated-information queries, Qwen states when it lacks real-time information and directs users to authoritative sources for changing streaming or audience-score data.
  • Subjective: For subjective questions, Qwen identifies that rankings such as the most innovative inventor depend on evaluation criteria rather than presenting one definitive choice.
  • Failure contrasts: The examples also expose failures where models invent definitions, accept false premises, infer unsupported personal preferences, or claim to translate an unavailable audio file.The unsupported DCCF definition, Orwell attribution, movie recommendations, and fabricated audio translation illustrate these failure modes.

F.3 Failure to abstain on reasoning datasets from s1.1 32B

The reasoning examples show models producing definitive answers despite missing information or unclear prompts. In the ladder problem, the ladder length and motion rate are unspecified; in the biology question, the referents and pathway are unclear, yet the model selects an option.

  • Ladder problem: The ladder question omits the ladder length and any rate needed to determine the top’s downward speed.The reasoning chain explicitly recognizes that dy/dt requires dx/dt or another relation, while the prompt supplies neither.
  • Ladder problem: The model nevertheless concludes that the answer is 6*sqrt(2) B after considering unsupported ladder-length assumptions.
  • Biology question: The biology prompt is unclear about what entities met and where the chain is heading, leaving the intended cellular pathway underdetermined.The reasoning chain notes that the options describe different cellular movements without clarifying the referents.
  • Biology question: Despite that ambiguity, the model selects ribosome to proteasome and supplies a detailed degradation narrative not established by the prompt.
  • Bolts question: The bolts question provides too little context to identify the project or determine a total, but the model answers 1,000,000.The reasoning considers several unrelated structures before producing the unsupported number.
Loading 2506.09038v1…