REVIEW 40 cited by
AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions
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
AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions
read the original 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datasets, including questions with unknown answers, underspecification, false premises, subjective interpretations, and outdated information.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.
Forward citations
Cited by 40 Pith papers
-
What Benchmarks Don't Measure: The Case for Evaluating Abstention Competence in Autonomous Agents
Current benchmarks overlook abstention competence in agents due to compliance bias; a new three-gap taxonomy and metrics (Safety Rate, Usability Rate, Informed Refusal Rate) demonstrate tunable safety-usability tradeo...
-
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.
-
Operadic consistency: a label-free signal for compositional reasoning failures in LLMs
Operadic consistency is a new per-question signal that correlates strongly with accuracy (r 0.86-0.94) across four multi-hop QA datasets and improves selective prediction over CoT-SC baselines.
-
PhantomFill: When the Form Demands an Answer, Language Models Invent One
Required JSON fields make LLMs invent answers to unanswerable questions 100% of the time in ten of thirteen models, even when an 'insufficient evidence' escape exists.
-
Not All Proofs Are Equal: Evaluating LLM Proof Quality Beyond Correctness
LLM proofs for hard math problems show large differences in quality metrics like conciseness and cognitive simplicity that correctness-only tests miss, along with trade-offs between quality and correctness.
-
Not All Proofs Are Equal: Evaluating LLM Proof Quality Beyond Correctness
ProofRank benchmark shows substantial differences in LLM proof quality not captured by correctness, with trade-offs between quality metrics and accuracy.
-
Knowing but Not Correcting: Routine Task Requests Suppress Factual Correction in LLMs
LLMs suppress factual corrections in task contexts despite internal knowledge of errors, with two training-free interventions shown to increase correction rates substantially.
-
Learning When Not to Decide: A Framework for Overcoming Factual Presumptuousness in AI Adjudication
A new structured prompting method (SPEC) helps AI detect insufficient evidence in adjudication tasks and defer decisions appropriately, reaching 89% accuracy on a benchmark varying information completeness from Colora...
-
Don't Start What You Can't Finish: A Counterfactual Audit of Support-State Triage in LLM Agents
LLM agents overcommit on non-complete tasks at 41.7% unless given explicit support-state categories, which raise typed deferral accuracy to 91.7%.
-
Generate, Filter, Control, Replay: A Comprehensive Survey of Rollout Strategies for LLM Reinforcement Learning
This survey introduces the Generate-Filter-Control-Replay (GFCR) taxonomy to structure rollout pipelines for RL-based post-training of reasoning LLMs.
-
BAS: A Decision-Theoretic Approach to Evaluating Large Language Model Confidence
BAS aggregates utility from an answer-or-abstain model across risk thresholds and is uniquely maximized by truthful confidence estimates.
-
MMOOC: A Comprehensive Benchmark for Out-of-Context Evaluation in Multimodal Large Language Models
MMOOC, a 41K-pair benchmark, shows current multimodal LLMs struggle to both refuse truly out-of-context questions and correctly answer questions that remain answerable despite misleading or shifted context.
-
Polistemics: Evaluating LLMs as Information Mediators in Politics & Elections
LLMs reliably mediate party positions only when evidence is clear; under absent, vague, or contradictory evidence they overclaim, fall back on party priors, and soften political rhetoric.
-
Polistemics: Evaluating LLMs as Information Mediators in Politics & Elections
A benchmark with AI judges shows that chatbots score high overall on political-information mediation but break down — overclaiming and sanitizing — specifically when evidence is absent, vague, or contradictory.
-
ReliableTableQA:How Much Supervision Does Reliability Annotation Need?
Reliability annotation for tabular QA reaches near-perfect accuracy with about 200 schema-stratified SFT examples, and GRPO adds measurable value only when SFT is undertrained.
-
Enhancing LLM Metacognition via Cognitive Pairwise Training
CPT is introduced as a pairwise reasoning-trace comparison stage that improves the reasoning-metacognition trade-off over standard SFT+RL pipelines across model scales.
-
Quantifying and Mitigating Premature Closure in Frontier LLMs
Frontier LLMs exhibit premature closure by selecting answers at high rates on medical tasks where the correct choice was removed and on open-ended queries, with safety prompting reducing but not eliminating the behavior.
-
MedVIGIL: Evaluating Trustworthy Medical VLMs Under Broken Visual Evidence
MedVIGIL provides a 300-case evaluation suite with 2556 probes that measures silent failures in medical VLMs under broken evidence, showing the best model at 69.2 on the composite score versus a human radiologist at 83.3.
-
MedVIGIL: Evaluating Trustworthy Medical VLMs Under Broken Visual Evidence
MedVIGIL introduces a clinician-supervised benchmark showing medical VLMs frequently give fluent answers on broken visual evidence, with top models 14 points below human radiologists on the composite score.
-
Coding Agents Don't Know When to Act
Coding agents exhibit action bias by proposing undesirable changes on already-fixed issues 35-65% of the time, and explicit reproduction instructions only partially mitigate this while creating new abstention errors.
-
Automatically Evolving Prompt Guidelines for Task-Specific Optimization
AGOPS automatically evolves task-specific prompt guidelines from reference answers and reports recovering 15.5–81.7% of the performance lost to underspecified prompts.
-
Knowing but Not Correcting: Routine Task Requests Suppress Factual Correction in LLMs
Task context suppresses factual correction in LLMs at the response-selection stage even when the model has encoded the error, and two training-free interventions raise correction rates substantially.
-
Ask Before You Diagnose: Safe-Psych, a Sequential Evaluation Benchmark for LLMs in Psychiatry
A new benchmark of 1,048 psychiatric notes shows LLMs frequently diagnose too early under incomplete evidence, and safety prompting only trades premature diagnoses for excessive abstention.
-
Hallucinations Undermine Trust; Metacognition is a Way Forward
LLMs need metacognition to align expressed uncertainty with their actual knowledge boundaries, moving beyond knowledge expansion to reduce confident errors.
-
Unsupervised Confidence Calibration for Reasoning LLMs from a Single Generation
Unsupervised single-generation confidence calibration for reasoning LLMs via offline self-consistency proxy distillation outperforms baselines on math and QA tasks and improves selective prediction.
-
Causal Evidence that Language Models use Confidence to Drive Behavior
Language models deploy multidimensional internal confidence representations and threshold-based policies to control abstention behavior, with causal support from activation steering experiments.
-
Decoupling Reasoning and Confidence: Resurrecting Calibration in Reinforcement Learning from Verifiable Rewards
DCPO decouples reasoning optimization from calibration in RLVR to fix overconfidence in LLMs without losing accuracy.
-
Decoupling Reasoning and Confidence: Resurrecting Calibration in Reinforcement Learning from Verifiable Rewards
DCPO decouples reasoning and calibration in RLVR, matching GRPO accuracy while reducing overconfidence via a claimed fundamental gradient conflict.
-
UCPO: Uncertainty-Aware Policy Optimization
UCPO uses decoupled advantage normalization and a dynamic uncertainty reward to train LLMs to express uncertainty, improving answer reliability (PAQ) but sometimes lowering F1.
-
Scaling with Confidence: Calibrating Confidence of LLMs for Adaptive Test Time Scaling
C3RL is a new RL algorithm combining correctness, calibration, and reference accuracy rewards to improve LLM confidence calibration, enabling CAS to outperform majority voting with up to 12.33x lower inference cost.
-
Legal Reasoning Is Not Lawyering: Rethinking Legal Benchmarks for Pro Se Access to Justice
Legal AI benchmarks must evaluate robustness to pro se litigant inputs rather than expert-preprocessed ones to support access-to-justice claims.
-
Bridging the Detection-to-Abstention Gap in Reasoning Models under Insufficient Information
JTS trains reasoning models via supervised warm-up and missing-premise RL to make an explicit answerability commitment that triggers early termination on unanswerable inputs, raising Abstention@Detection near saturation.
-
When No Benchmark Exists: Validating Comparative LLM Safety Scoring Without Ground-Truth Labels
A formalization of benchmarkless LLM safety scoring validated via an instrumental-validity chain of contrast separation, target variance dominance, and rerun stability, demonstrated on Norwegian scenarios.
-
Beyond Benchmarks: MathArena as an Evaluation Platform for Mathematics with LLMs
MathArena is a maintained platform evaluating LLMs across olympiad problems, proofs, research questions, and formal proofs, with GPT-5.5 reaching 98% on 2026 USAMO and 74% on research-level tasks.
-
Beyond Benchmarks: MathArena as an Evaluation Platform for Mathematics with LLMs
MathArena is broadened into a maintained platform with new benchmarks for proofs, research questions, and formal verification, where GPT-5.5 scores 98% on 2026 USAMO and 74% on research-level tasks.
-
Stateful Evidence-Driven Retrieval-Augmented Generation with Iterative Reasoning
A stateful iterative RAG system converts retrieved documents into scored reasoning units, maintains supportive and non-supportive evidence, and performs deficiency-driven query refinement to achieve more robust QA per...
-
Position: The Hidden Costs and Measurement Gaps of Reinforcement Learning with Verifiable Rewards
The paper identifies confounds in RLVR evaluations that inflate apparent gains and proposes a minimum standard for budget-matched, contamination-aware assessment with calibration tracking.
-
ConfidenceBench: Evaluating Confidence Calibration in Large Language Models
Frontier LLMs' verbalized confidence is often miscalibrated: the most accurate model is not the best-calibrated, and several models score worse than a calibrated random baseline.
-
Prompt Governance? On Governing Technologies Governed by Natural Language
Literature on system prompts for AI shows fragmented and contradictory claims that complicate policy efforts to use them as reliable governance mechanisms.
-
Treatment, evidence, imitation, and chat
LLMs cannot solve the medical treatment problem through imitation alone because it requires evidence from experiments or observations, posing ethical challenges for training such systems.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.