AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions

TL;DR

AbstentionBench investigates how well modern LLMs know when not to answer by evaluating abstention across 20 diverse datasets. It shows that abstention remains an open challenge, with model scaling offering limited gains and reasoning-focused fine-tuning often degrading abstention. The work demonstrates that post-training can help abstention in some contexts but not in underspecified cases, and that a system prompt can improve abstention without solving the fundamental uncertainty reasoning problem. By providing a large, automated benchmarking framework and an LLM-based judge, the paper lays a foundation for developing more reliable LLMs that appropriately abstain under uncertainty, underspecification, or conflicting premises.

Abstract

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datasets, including questions with unknown answers, underspecification, false premises, subjective interpretations, and outdated information.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.

Figures (24)

• Figure 1: (a) AbstentionBench evaluates model performance on over $35k$ unanswerable questions drawn from diverse scenarios. (b) Faced with an unanswerable question, an abstention response is desired, yet models often respond incorrectly. (c) Reasoning interventions worsen abstention compared with instruction-tuned baselines.
• Figure 2: AbstentionBench evaluates frontier LLMs across 20 datasets spanning diverse scenarios.
• Figure 3: Bigger or more powerful closed-source models aren't always better at abstention. (a) Average performance for open and proprietary LLMs. (b) Increasing model scale in Llama does not improve abstention. (c) Abstention performance distribution for Qwen across scenarios.
• Figure 4: Higher accuracy doesn't lead to better abstention. Abstention recall and response correctness exhibit variable degree of correlation on different datasets.
• Figure 5: Post-training improves response accuracy and abstention recall, but not for underspecified context. (a) Change in abstention recall of Tülu checkpoints vs. Llama 3.1 Base 8B. (b) Change in response accuracy. See \ref{['sec:app-additional-results']} for precision and F1 score. (c) Contribution of each post-training stage to change in recall: RLVR degrades abstention.