Rewarding Efficient Reasoning Improves Abstention on Underspecified Tasks in Reasoning Models
Summary
This paper introduces a novel GRPO reward to improve abstention in large reasoning models on underspecified tasks, enhancing efficiency and human-like reasoning while maintaining performance.
arXiv:2609.20846v1 Announce Type: new
Abstract: While modern large reasoning models (LRMs) excel at providing correct answers in many tasks, we provide additional evidence for the observation that they often struggle with a critical capability: knowing when to abstain from answering. We analyze this gap by comparing LRM behavior to results from a human study, revealing that human reasoning effort on unanswerable tasks is upper-bounded by answerable tasks, whereas LRMs waste computational resources by generating longer Chains of Thought (CoTs) on unanswerable than on answerable prompts. To overcome this inefficiency, we take inspiration from a resource-rational perspective on human cognition and introduce a novel GRPO reward that encourages efficient reasoning about whether the task contains all the information needed to solve it. Fine-tuning several 4B LRMs with this reward leads to human-like abstention performance gains (+12.8% on average) while retaining answering capabilities and boosting the models' efficiency (44% shorter CoTs on average).
View Cached Full Text
Cached at: 09/21/26, 09:01 AM
# Rewarding Efficient Reasoning Improves Abstention on Underspecified Tasks in Reasoning Models Source: [https://arxiv.org/html/2609.20846](https://arxiv.org/html/2609.20846) Polina Tsvilodub1,⋆\\star,Max Höth2,3,Michael Franke1, Björn Deiseroth2,3,†\\dagger,Carina Kauf⋆,†\\star,\\dagger 1University of Tübingen,2Aleph Alpha Research,3Lab1141 Correspondence:[polina\.tsvilodub@uni\-tuebingen\.de](mailto:[email protected]) ###### Abstract While modern large reasoning models \(LRMs\) excel at providing correct answers in many tasks, we provide additional evidence for the observation that they often struggle with a critical capability: knowing when to abstain from answering\. We analyze this gap by comparing LRM behavior to results from a human study, revealing that human reasoning effort on unanswerable tasks is upper\-bounded by answerable tasks, whereas LRMs waste computational resources by generating longer Chains of Thought \(CoTs\) on unanswerable than on answerable prompts\. To overcome this inefficiency, we take inspiration from a resource\-rational perspective on human cognition and introduce a novel GRPO reward that encourages efficient reasoning about whether the task contains all the information needed to solve it\. Fine\-tuning several 4B LRMs with this reward leads to human\-like abstention performance gains \(\+12\.8% on average\) while retaining answering capabilities and boosting the models’ efficiency \(44% shorter CoTs on average\)\. Rewarding Efficient Reasoning Improves Abstention on Underspecified Tasks in Reasoning Models Polina Tsvilodub1,⋆\\star, Max Höth2,3, Michael Franke1,Björn Deiseroth2,3,†\\dagger,Carina Kauf⋆,†\\star,\\dagger1University of Tübingen,2Aleph Alpha Research,3Lab1141Correspondence:[polina\.tsvilodub@uni\-tuebingen\.de](mailto:[email protected]) ††footnotetext:⋆\\starWork done while working at Aleph Alpha Research\.††footnotetext:†\\daggerJoint senior authorship\.Figure 1:Overview of our approach and contributions\.\(A\): Example of anabstentionversion and ananswerableversion of a question from QuestBench\(Liet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib2)\)that was also used for the human study\.\(B\): Overview of the SURE objective proposed in this work\. The example CoT chunks are stylized\.\(C\): Our SURE GRPO objective makes LRMs’ abstention more human\-like\. Left: SURE objective improvesabstentionperformance of an off\-the\-shelf reasoning model \(Phi\-4\-mini\-reasoning\), while retainingansweringperformance\. Right: the SURE objective reduces redundant CoT tokens, allocating reasoning resources on abstention in a more human\-like way\.## 1Introduction Large Language Models \(LLMs\) have achieved impressive performance across various tasks and domains\(Brownet al\.,[2020](https://arxiv.org/html/2609.20846#bib.bib17); Bubecket al\.,[2023](https://arxiv.org/html/2609.20846#bib.bib18)\)\. Recently, fine\-tuning LLMs specifically for complex reasoning tasks by incentivizing models to produce long chains of thought \(CoT\) has become common\(Weiet al\.,[2022](https://arxiv.org/html/2609.20846#bib.bib34); Shaoet al\.,[2024](https://arxiv.org/html/2609.20846#bib.bib14); Muennighoffet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib19)\)\. This approach has given rise to so\-called Large Reasoning Models \(LRMs\) that perform particularly well on tasks such as mathematical reasoning or coding\(Guoet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib5)\)\. However, while LRMs excel on answerable tasks, their capability to accurately identifywhen not to answer, i\.e\., when toabstain, remains subpar, even though this capability is critical for user\-facing deployment\(Kirichenkoet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib1)\)\. Abstention is an umbrella term for a range of behaviors where models refuse to directly answer a query, like outputting “I don’t know”, hedging, or asking for clarification\(Kirichenkoet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib1); Wenet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib36)\)\. Depending on context, abstention is expected, e\.g\., \(i\) when the prompt is underspecified, \(ii\) the answer is generally unknown or \(iii\) the prompt should not be answered for safety reasons\. Crucially, because real\-world user inputs are frequently vague or linguistically underspecified\(Kuhnet al\.,[2023](https://arxiv.org/html/2609.20846#bib.bib48); Zhanget al\.,[2024](https://arxiv.org/html/2609.20846#bib.bib21)\), training effective LRMs requires balancing caution with helpfulness: models should not refuse every underspecified request, but abtain only in critical cases, while maintaining high accuracy on tasks where an answer is expected\(Varshneyet al\.,[2024](https://arxiv.org/html/2609.20846#bib.bib38)\)\. In contrast to other related work in this space \(c\.f\. Section[2](https://arxiv.org/html/2609.20846#S2)\), here, we take inspiration from a*resource\-rational*perspective on human cognition\(Lieder and Griffiths,[2020](https://arxiv.org/html/2609.20846#bib.bib8)\)and results from our own, novel experiment with human participantsto evaluate and improve abstention efficiency and performance of LRMs\. We show that humans identify with high accuracy both when to answer and when to abstain \(see Figure[1](https://arxiv.org/html/2609.20846#S0.F1)\(C, left\) and Section[4](https://arxiv.org/html/2609.20846#S4)\), and that their reasoning effort on unanswerable tasks is upper\-bounded by answerable tasks \(Figure[1](https://arxiv.org/html/2609.20846#S0.F1)\(C, right\)\)\. By contrast, LRMs often produce excessively long CoTs \(i\.e\., “overthink”\) specifically in cases when they should abstain \(Figure[1](https://arxiv.org/html/2609.20846#S0.F1)\(C, right\)\)\. Our contributions are as follows \(visualized in Figure[1](https://arxiv.org/html/2609.20846#S0.F1)\): - •We conduct a human study, showing thathumans accurately identify when a question is unanswerable, and do so with the same reasoning resources as for answering\. - •We compare reasoning models \(4B–32B parameters, six model families\) to human results and find a clear misalignment:LRMs abstain far worse and less efficiently than humans\. - •We propose theSUfficiency\-awareReasoningEfficiency\(SURE\) rewardfor Group Relative Policy Optimization \(GRPO\): it combines an outcome reward with a process reward that penalizes reasoning beyond the point at which the CoT determines whether task\-crucial information is missing\.SURE\-fine\-tuned LRMs showed improved and more efficient abstention performance, while keeping robust answering capabilities\. ## 2Related work #### Abstention in LLMs\. Abstention capabilities of LLMs have received increasing attention\(Kirichenkoet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib1); Wenet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib36)\), and have been evaluated, e\.g\., on ambiguous or underspecified tasks\(e\.g\., Slobodkinet al\.,[2023](https://arxiv.org/html/2609.20846#bib.bib40); Sunet al\.,[2024](https://arxiv.org/html/2609.20846#bib.bib42); Zhanget al\.,[2024](https://arxiv.org/html/2609.20846#bib.bib21)\), on tasks with unknown answers\(Amayuelaset al\.,[2024](https://arxiv.org/html/2609.20846#bib.bib27)\), or as a possible mitigation of LLM hallucinations\(Tonmoyet al\.,[2024](https://arxiv.org/html/2609.20846#bib.bib53)\)\. Other work has focused on clarification question asking in LLMs\(Kunduet al\.,[2020](https://arxiv.org/html/2609.20846#bib.bib60); Andukuriet al\.,[2024](https://arxiv.org/html/2609.20846#bib.bib28); Testoni and Fernández,[2024](https://arxiv.org/html/2609.20846#bib.bib61)\), also when the task is underspecified\(Liet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib2); Lachenmaieret al\.,[2026](https://arxiv.org/html/2609.20846#bib.bib25); Wanget al\.,[2026](https://arxiv.org/html/2609.20846#bib.bib31)\)\. A related line of work focuses on evaluating how well LLMs express uncertainty\(Kadavathet al\.,[2022](https://arxiv.org/html/2609.20846#bib.bib44); Tianet al\.,[2023](https://arxiv.org/html/2609.20846#bib.bib26)\)\. Several studies have proposed approaches for improving LLMs’ or LRMs’ abstention capabilities through prompting\(Denget al\.,[2024](https://arxiv.org/html/2609.20846#bib.bib46)\)or fine\-tuning to produce correct answers\(Chenet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib45); Zhaiet al\.,[2026](https://arxiv.org/html/2609.20846#bib.bib22)\), while less work has considered process rewards\(Lightmanet al\.,[2024](https://arxiv.org/html/2609.20846#bib.bib37)\)or the efficiency of the reasoning process on abstention tasks\(but see Guet al\.,[2026](https://arxiv.org/html/2609.20846#bib.bib23)\)\. #### Efficiency in LRMs\. Efforts to improve CoT efficiency have applied length penalties during RL fine\-tuning of LRMs on answerable tasks\(Teamet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib32)\), often aiming to allocate longer CoTs to harder than to simpler prompts\(Linget al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib33); Xianget al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib35)\)\. Early\-exiting approaches force efficient termination of CoTs during inference\(Yanget al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib29); Wanget al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib30)\)\. Most abstention work focuses only on LLMs, while work on answerable tasks has also compared LLMs to human reasoning\(e\.g\., Eisapeet al\.,[2024](https://arxiv.org/html/2609.20846#bib.bib52); Liuet al\.,[2024](https://arxiv.org/html/2609.20846#bib.bib51)\)\. We take inspiration fromde Vardaet al\.\([2025](https://arxiv.org/html/2609.20846#bib.bib47)\)who show that LRMs’ CoTs align with human reaction times \(RTs\) and capture reasoning demands on various answerable tasks, and evaluate human performance also on unanswerable tasks to ground the assessment of LRMs on abstention\. #### Resource Rationality in Humans\. Work within the resource rationality framework has shown that humans flexibly allocate their reasoning resources, often measured through time allocated for solving a task\(Lieder and Griffiths,[2020](https://arxiv.org/html/2609.20846#bib.bib8)\), and higher reasoning time often leads to more accurate task performance\(Wickelgren,[1977](https://arxiv.org/html/2609.20846#bib.bib59)\)\. Yet while previous work has investigated factors influencing abstention\(Undorfet al\.,[2021](https://arxiv.org/html/2609.20846#bib.bib54); Lawet al\.,[2022](https://arxiv.org/html/2609.20846#bib.bib55)\)and clarification question production\(Clark and Wilkes\-Gibbs,[1986](https://arxiv.org/html/2609.20846#bib.bib58); Purveret al\.,[2001](https://arxiv.org/html/2609.20846#bib.bib57); Aliet al\.,[2026](https://arxiv.org/html/2609.20846#bib.bib56); Tsvilodubet al\.,[2026](https://arxiv.org/html/2609.20846#bib.bib3)\)in humans, the resource allocation in reasoning about abstention remains less clear\. ## 3Experiment Design & Dataset Following definitions byKirichenkoet al\.\([2025](https://arxiv.org/html/2609.20846#bib.bib1)\); Wenet al\.\([2025](https://arxiv.org/html/2609.20846#bib.bib36)\), we investigate whether LRMs refuse to answer in any form \(e\.g\., hedging, outputting “I don’t know”, or asking for clarification\) when given queries for which abstention is expected\. An LLM judge provides binary annotations of the outputs \(see Section[5](https://arxiv.org/html/2609.20846#S5)for details and Appendix[A\.1](https://arxiv.org/html/2609.20846#A1.SS1)for the prompt\)\. Humans are evaluated analogously, via a binary forced\-choice task asking whether a question is answerable \(Section[4](https://arxiv.org/html/2609.20846#S4)\)\. Our experiments vary thequestion type\(answerable vs\. unanswerable\) by using evaluation datasets which contain both question types \(closely matched in difficulty\) for evaluating and subsequent fine\-tuning \(Section[6](https://arxiv.org/html/2609.20846#S6)\) LRMs: QuestBench\(Liet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib2)\)and AbstentionBench\(Kirichenkoet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib1)\)\. An example from AbstentionBench is shown in Figure[1](https://arxiv.org/html/2609.20846#S0.F1)\(A\)\. For QuestBench, we use the GSM\-Q subset as unanswerable questions which consist of grade school level math tasks from the GSM8K dataset \(derived fromLiet al\.,[2024](https://arxiv.org/html/2609.20846#bib.bib6), which is used for answerable questions\)\.Liet al\.\([2025](https://arxiv.org/html/2609.20846#bib.bib2)\)constructed GSM\-Q by removing a single variable in each question, resulting in tasks where crucial information is missing\. The questions are paired\. The questions also vary with respect to the number of steps that are needed to solve them \(i\.e\., in their difficulty\)\. AbstentionBench\(Kirichenkoet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib1)\)consists of a combination of 20 datasets covering different tasks with both answerable and unanswerable prompts\. The tasks cover different reasons for abstention \(underspecified context, but also underspecified intent, stale data, false premise, or where the answer is unknown or subjective\)\. We exclude questions with more than 2048 tokens\. We use 500 test samples from each benchmark across all reported evaluations\. The QuestBench split of the unanswerable test set evaluates abstention on underspecified prompts; the AbstentionBench split evaluates abstention on diverse tasks, approximately balanced across the benchmark’s distribution of abstention reasons\. ## 4Humans Answer and Abstain Accurately & Efficiently To ground LRM evaluations, we draw on insights about human behavior on the same benchmarks\. If human task solving is conceptualized as a search over a problem space\(Simon and Newell,[1971](https://arxiv.org/html/2609.20846#bib.bib7)\), one hypothesis is that the search will be terminated as soon as a gap in the problem representation \(i\.e\., missing information\) is encountered, predicting that the resources for abstention are upper\-bounded by the respective answerable tasks\. Here, we empirically compare human performance and reasoning effort on unanswerable vs\. answerable tasks through an exploratory web\-based experiment, investigating the following questions:111The materials can be viewed at:[https://github\.com/polina\-tsvilodub/reasoning\-under\-missing\-info](https://github.com/polina-tsvilodub/reasoning-under-missing-info)\.\(1\) Do humans accurately identify whether a question is \(un\)answerable? \(2\) Are human completion times \(i\.e\., “reasoning effort”\) on unanswerable tasks upper\-bounded by answerable tasks? \(3\) When participants are additionally incentivized to accurately complete certain trials, do their performance and reasoning effort change? #### Materials, Procedure & Participants\. We devised a2×2×32\\times 2\\times 3design with factorsquestion type\(answerable vs\. unanswerable\),question domain\(math tasks from QuestBench vs\. common sense questions from AbstentionBench\), andimportanceof solving the task \(high vs\. default vs\. low; operationalized through different numbers of points for completing a given trial correctly, and bonus payments proportional to achieved points above a threshold\)\. The default condition only contained task instructions\. The test items were selected by randomly sampling six items per domain and question type, resulting in 24 items\. The items were additionally filtered by the authors for naturalness\. Each item was used in the three importance conditions\. Full details about the materials are reported in Appendix[B](https://arxiv.org/html/2609.20846#A2)\. Participants \(N=126N=126, recruited via Prolific\) were self\-reported native English speakers with approval rates over 95% and at least five prior studies\. Because participants might hesitate to abstain in an experimental setting, each participant first viewed examples of each question type \(see Appendix[B](https://arxiv.org/html/2609.20846#A2)\)\. Then, they completed six main trials \(one per question type×\\timesimportance condition, with three trials per domain\), and one attention check\. On each trial, participants read a question, embedded in a randomly sampled importance prompt condition\. They first only saw a forced choice \(FC\) task where they indicated whether the question is “Answerable” or “Not definitively answerable”\. Once they answered the FC task, two text boxes appeared, one for an explanation of the solution steps, and the other either for the final answer to the answerable questions, or for an explanation of what information is missing in the unanswerable questions\. The attention check trials were visually identical, and asked participants to provide specific answers\. Participants took 17 minutes on average, and were reimbursed £1\.20 with up to £0\.20 bonus\. Full experiment details are reported in Appendix[B](https://arxiv.org/html/2609.20846#A2)\.  Figure 2:Left: FC accuracy by question type across importance conditions\.Right: Typing accuracy \(correctness of answers or identified missing information\) across importance conditions\. Error bars are bootstrapped 95\-% CIs\.\. #### Results\. We analyze the accuracy of the responses provided to the FC question, the accuracy of the typed answers \(pooled across the text fields\), and reaction time \(RT\) per trial\. After excluding participants who failed any part of the attention check, we analyze data from 83 participants\. We analyze all results with Bayesian mixed\-effect regression models, always including maximal converging random effects structure\. Posterior means and 95% credible intervals are reported\. #### 1\. Humans identify when to answer or abstain well above chance, but more accurately when to answer\. The forced choice accuracy for answerable and unanswerable questions across importance conditions and domains is shown in Figure[2](https://arxiv.org/html/2609.20846#S4.F2)\(left\)\. We analyze the forced\-choice answer accuracy using a logistic regression model\.222Model in R syntax:accuracy∼\\simdomain \* prompt \* question\_type \+ \(1 \+ domain \* prompt \* question\_type \| subject\) \+ \(1 \| itemId\)Humans were credibly more accurate at identifying answerable questions \(β=5\.86\[0\.21,14\.02\]\\beta=5\.86\[0\.21,14\.02\]across domains\), driven by credible differences on math questions \(β=16\.72\[6\.78,34\.0\]\\beta=16\.72\[6\.78,34\.0\]\), but not in the common sense questions \(β=6\.71\[−14\.48,29\.26\]\\beta=6\.71\[\-14\.48,29\.26\]\)\. Still, humans identified when to abstain credibly above chance \(posterior probability of effect 96%\)\. #### 2\. Humans reasoning effort on unanswerable tasks is upper\-bounded by answerable tasks\. The total response time \(RT\) taken by the participants on each answerable and unanswerable trial is shown in Figure[1](https://arxiv.org/html/2609.20846#S0.F1)\(C\) \(right\)\. Separate analyses of FC and typing times are reported in Appendix[B](https://arxiv.org/html/2609.20846#A2)\. All RTs are log\-transformed\. We use a Bayesian linear mixed effect regression model\.333Model in R syntax:log\(RT\)∼\\simdomain \* importance \* question\_type \+ \(1 \+ question\_type \+ domain \| subject\) \+ \(1 \| itemId\)\.Participants’ RTs didn’t differ credibly between the answerable and abstention tasks across domains and conditions \(β=0\.03\[−0\.14,0\.19\\beta=0\.03\[\-0\.14,0\.19\)\. However, humans did have a marginally credibly higher RT for answerable than abstention math tasks \(β=0\.13\[−0\.00,0\.27\]\\beta=0\.13\[\-0\.00,0\.27\]\), but not common sense tasks \(β=0\.03\[−0\.64,0\.72\]\\beta=0\.03\[\-0\.64,0\.72\]\)\. In general, RTs were higher on math than the common sense questions \(across question types:β=0\.54\[0\.12,0\.98\]\\beta=0\.54\[0\.12,0\.98\]\)\. #### 3\. Humans are more accurate and reason slower when incentivized accordingly\. The accuracy of the typed answers by importance condition is shown in Figure[2](https://arxiv.org/html/2609.20846#S4.F2)\(right\)\. Humans are marginally, but credibly more accurate in the high\-importance than the low\-importance condition across domains \(posterior probability 96\.45%\), but visual inspection suggests that the trend is driven by answerable questions \(Figure[2](https://arxiv.org/html/2609.20846#S4.F2), right\)\. Additionally, humans took longer in the high than default prompt condition \(β=0\.12\[0\.01,0\.23\]\\beta=0\.12\[0\.01,0\.23\]\), an effect driven primarily by differences in the common sense domain \(high vs\. default:β=0\.18\[0\.01,0\.35\]\\beta=0\.18\[0\.01,0\.35\]\)\. In sum, these results suggest that humans accurately identify when to abstain, while still answering accurately, and do so with roughly the same reasoning resources spent on answering and abstention\. We consider these patterns as a human baseline for comparison with off\-the\-shelf LRMs\. ## 5LRMs’ Abstention Performance Is Worse Than in Instruct Models We first evaluate reasoning models from six different families on our evaluation set \(Section[3](https://arxiv.org/html/2609.20846#S3)\)\. We evaluate all models in a free generation setting, with maximally 10000 new tokens\. The default generation configuration of each model was used\. We use an LLM as a judge \(across experiments:Qwen3\-30B\-A3B\-Instruct\-2507,Team,[2025](https://arxiv.org/html/2609.20846#bib.bib9)\) to evaluate whether an answer is an abstention, and to evaluate the correctness of outputs for answerable questions\. The prompts are in Appendix[A\.1](https://arxiv.org/html/2609.20846#A1.SS1),[A\.2](https://arxiv.org/html/2609.20846#A1.SS2)\. To evaluate the initial LRM performance in the strongest baseline setting, we designed a prompt stating that the task might be missing information, in which case the model should abstain and stop reasoning as soon possible \(the full prompt and details are in Appendix[A\.3](https://arxiv.org/html/2609.20846#A1.SS3)\)\. We evaluate a range of models that have both an instruction fine\-tuned and a reasoning variant available:Qwen3\-4B\-2507,Qwen3\-32B\(Team,[2025](https://arxiv.org/html/2609.20846#bib.bib9)\),Olmo\-3\.1\-32B\(Olmoet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib10)\),Falcon\-H1\-7B\(Teamet al\.,[2026](https://arxiv.org/html/2609.20846#bib.bib11)\),Phi\-4\-mini\(Xuet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib12)\)DeepSeek\-R1\-Distill\-Llama\-8B\(DeepSeek\-AI,[2025](https://arxiv.org/html/2609.20846#bib.bib62)\)\.444For DeepSeek, theLlama\-3\.1\-8B\-Instructwas used as the instruction\-tuned variant\.For subsequent fine\-tuning, for comparability across architectures, we focus on models with approximately 4B parameters:Qwen3\-4B\-Thinking\-2507\(Team,[2025](https://arxiv.org/html/2609.20846#bib.bib9)\),Phi4\-mini\-reasoning\(Xuet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib12)\)andNVIDIA\-Nemotron\-3\-nano\-4B\-BF16\(used in the enabled thinking mode,Blakemanet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib13)\)\. Figure 3:Reasoning models tend to perform worse on abstention tasks, and inefficiently allocate the compute budget by producing verbose CoTs when they should abstain\. Accuracy \(upper row\) and length of generations \(in tokens, bottom row\) of reasoning models from six different families on answerable vs\. unanswerable questions \(x\-axis\), averaged across two benchmarks\. Error bars show 95% bootstrapped CIs\.#### LRMs Perform Worse than Instruction\-Tuned Models on Abstention in Terms of Performance and Efficiency\. We first compare the performance of seven different reasoning models on unanswerable questions vs\. answerable questions\. Results across the benchmarks are shown in Figure[3](https://arxiv.org/html/2609.20846#S5.F3)\. Accuracy on answerable questions measures answer correctness, while accuracy on unanswerable samples measures whether the model correctly abstains rather than guesses\. We analyze the results with a Bayesian logistic regression model, regressing the accuracy against the question type, model, and their interaction\.555Model in R syntax:accuracy∼\\simquestion\_type \* model\_name\. Note that we include human results and humans as a “model” to enable comparison\.For each single LRM as well as across models, the abstention performance was credibly lower than answering performance \(β=0\.96\[0\.87,1\.06\]\\beta=0\.96\[0\.87,1\.06\]\) \(Figure[3](https://arxiv.org/html/2609.20846#S5.F3), upper row\)\. All LRMs abstain credibly worse than humans \(β=1\.03\[0\.73,1\.35\]\\beta=1\.03\[0\.73,1\.35\]\)\. For three of the models, the gap between the abstention and answering performance was larger than for humans with a posterior probability over 85%\. Additionally, the models’ CoTs on unanswerable questions are at least as long as those for the answerable condition \(Figure[3](https://arxiv.org/html/2609.20846#S5.F3), bottom row\)\. They are even longer than the answerable question CoTs with a posterior probability of 85%\.666Underlying linear regression model:avg tokens∼\\simquestion\_type\.This suggests that the models might tend towards more inefficient CoTs or loops when faced with unanswerable questions which might be missing information samples\. To investigate whether these abstention dynamics are due to the reasoning fine\-tuning, we compare the reasoning models to their respective instruction fine\-tuned versions\. The accuracies and CoT lengths on answerable vs\. unanswerable questions across benchmarks are shown in Figure[6](https://arxiv.org/html/2609.20846#A2.F6)in the Appendix\. The instruct models performed credibly better on abstention than answering tasks, driven by four of the six model pairs\.777Logistic regression model in R syntax:accuracy∼\\simquestion\_type \* model\_type \+ \(1 \| model\_family\)\.The gap between the answering and abstention performance was credibly bigger in the reasoning than instruction models \(β=1\.67\[1\.54,1\.80\]\\beta=1\.67\[1\.54,1\.80\]\)\. The lengths of the CoTs in the instruct models did not credibly differ between answering and abstention \(β=−24\.49\[−759\.29,713\.78\]\\beta=\-24\.49\[\-759\.29,713\.78\]\)\.888Underlying linear regression model:avg tokens∼\\simquestion\_type \* model\_type\.Therefore, the gap in performance and the CoT allocation is a consequence of the reasoning fine\-tuning of the models\. #### LRMs Identify Missing Information in Abstention Tasks Early on\. As hypothesized in Section[4](https://arxiv.org/html/2609.20846#S4), a rational allocation of reasoning resources might require stopping the reasoning once missing information is identified\. We explore to which extent LRMs’ CoTs reflect such a strategy by annotating whether the CoT sentences contain reasoning about missing information for the task \(see Section[6\.1](https://arxiv.org/html/2609.20846#S6.SS1)for details\) and find a stricking discrepancy\. Figure[7](https://arxiv.org/html/2609.20846#A3.F7)in the Appendix shows that only up to∼25%\\sim 25\\%of the CoT on abstention tasks is required until missing information is identified, suggesting that a large portion of the CoT is redundant and might lead to the observed misalignment with human behavior\. We use these results as a departure point for developing a fine\-tuning approach for improving efficiency and abstention performance of LRMs, described next\. ## 6Fine\-tuning LRMs to Reason about Missing Information Leads to Efficient Human\-like Accurate Abstention The results from Section[5](https://arxiv.org/html/2609.20846#S5.SS0.SSS0.Px1)suggest that for abstention tasks, a large portion of the CoT which might contain, e\.g\., “self\-verification” like repeating solution attempts\(Muennighoffet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib19)\), might worsen performance\. Therefore, we develop an objective for reinforcement learning \(RL\) fine\-tuning with GRPO specifically aiming to reduce the redundancy in the CoT \(c\.f\. Figure[7](https://arxiv.org/html/2609.20846#A3.F7)\), in order to improve the efficiency and accuracy when abstention is needed while maintaining response accuracy and propensity on sufficiently specified tasks\. We construct a fine\-tuning dataset \(disjoint from the evaluation samples\) from AbstentionBench and QuestBench GSM\-Q, with about 60% unanswerable and 40% answerable samples\.999We use GSM8K\(Cobbeet al\.,[2021](https://arxiv.org/html/2609.20846#bib.bib4)\)for the answerable counterpart to QuestBench GSM\-Q\. ### 6\.1SURE Fine\-Tuning of Reasoning Models #### GRPO\. Group Relative Policy Optimization \(GRPO\)Shaoet al\.\([2024](https://arxiv.org/html/2609.20846#bib.bib14)\)aligns language models without a value network by estimating advantagesAi=\(ri−μ\)/σA\_\{i\}=\(r\_\{i\}\-\\mu\)/\\sigmafrom a group ofGGoutputsoio\_\{i\}sampled for queryqq, whereμ\\muandσ\\sigmaare the group’s reward mean and standard deviation\. The policyπθ\\pi\_\{\\theta\}maximizes the objective𝒥\(θ\)\\mathcal\{J\}\(\\theta\): 𝒥\(θ\)\\displaystyle\\mathcal\{J\}\(\\theta\)=𝔼\[1G∑i=1G\(min\(ρiAi,\\displaystyle=\\mathbb\{E\}\\Bigg\[\\frac\{1\}\{G\}\\sum\_\{i=1\}^\{G\}\\Big\(\\min\\big\(\\rho\_\{i\}A\_\{i\},\(1\)clip\(ρi,1−ϵ,1\+ϵ\)Ai\)−β𝔻KL\(πθ∥πref\)\)\],\\displaystyle\\text\{clip\}\(\\rho\_\{i\},1\-\\epsilon,1\+\\epsilon\)A\_\{i\}\\big\)\-\\beta\\mathbb\{D\}\_\{\\text\{KL\}\}\(\\pi\_\{\\theta\}\\\|\\pi\_\{\\text\{ref\}\}\)\\Big\)\\Bigg\],whereρi=πθ\(oi\|q\)/πθold\(oi\|q\)\\rho\_\{i\}=\\pi\_\{\\theta\}\(o\_\{i\}\|q\)/\\pi\_\{\\theta\_\{\\text\{old\}\}\}\(o\_\{i\}\|q\)is the policy ratio,ϵ\\epsilonis the clip margin, andβ\\betascales the token\-level KL divergence𝔻KL\\mathbb\{D\}\_\{\\text\{KL\}\}from the reference modelπref\\pi\_\{\\text\{ref\}\}\. Typically, the rewardrir\_\{i\}relies solely on a final accuracy indicator,ri=acc\(oi\)r\_\{i\}=\\text\{acc\}\(o\_\{i\}\)\(accuracy\-onlyreward\)\. #### Baseline reward\. To explicitly suppress verbosity, we modify the reward to incorporate a standard length penalty:ri=acc\(oi\)−λ\|oi\|r\_\{i\}=\\text\{acc\}\(o\_\{i\}\)\-\\lambda\|o\_\{i\}\|, where\|oi\|\|o\_\{i\}\|is the output length andλ\\lambdacontrols the penalty magnitude \(accuracy \+ length penaltyreward\)\. #### SURE reward\. To explicitly incentivize the model to halt its reasoning once it detects missing information, we formulate a compositeSUfficiency\-awareReasoningEfficiency reward\. An intuitive explanation of the reward is presented in Figure[1](https://arxiv.org/html/2609.20846#S0.F1)\(B\)\. For a given sampled outputoio\_\{i\}, the total rewardrir\_\{i\}combines an outcome accuracy score with a process\-focused efficiency terme\(oi\)e\(o\_\{i\}\): ri=0\.5\+α⋅acc\(oi\)\+αeff⋅e\(oi\)r\_\{i\}=0\.5\+\\alpha\\cdot\\text\{acc\}\(o\_\{i\}\)\+\\alpha\_\{\\text\{eff\}\}\\cdot e\(o\_\{i\}\)\(2\)whereacc\(oi\)\\text\{acc\}\(o\_\{i\}\)measures the binary task accuracy, evaluating both answerable and abstention cases\. The terme\(oi\)e\(o\_\{i\}\)captures the proportion of the reasoning trace that is redundant: e\(oi\)=n−kne\(o\_\{i\}\)=\\frac\{n\-k\}\{n\}\(3\)Here,nnis the total number of sentences in the reasoning trace ofoio\_\{i\}, andkkis the index of the first sentence where missing task\-relevant information is identified\. The scalar weightsα\\alphaandαeff\\alpha\_\{\\text\{eff\}\}balance the outcome and process signals and are set to 0\.5 across experiments\. To localizekkwithin a reasoning trace, we generate a complete model rolloutoio\_\{i\}, segment this rollout into discrete chunks based on punctuation boundaries, and apply an LLM as a judge \(Qwen3\-30B\-A3B\-Instruct\-2507\) to assess each chunk individually\. For each chunk, the judge returns a binary annotation whether it mentions that the task is missing information\. The specific judge prompt and further implementation details are provided in Appendix[A\.4](https://arxiv.org/html/2609.20846#A1.SS4)\. We note that the reward in Eq\.[2](https://arxiv.org/html/2609.20846#S6.E2)can be optimized through two strategies: either decreasingnnwhilekkremains constant, or increasingkkwhilennremains constant\. We explore empirically what happens to the models when using this objective within GRPO training\. Figure 4:SURE GRPO training results: Performance of models averaged across three families trained with our SURE GRPO objective against two strong baseline objectives \(“Accuracy \+ Length penalty” and “Accuracy\-only”\), relative to the performance of the initial LRMs \(“Base”\)\. 95% bootstrapped CIs are shown\.Left: recall on abstention tasks; increase indicates that models guessed answers less\.Right: accuracy of provided answers on the answerable questions; difference to the base models indicates deterioration of the models’ capabilities\. Only SURE objective combines stable performance and efficiency gains across models \(see Table[1](https://arxiv.org/html/2609.20846#S6.T1)\)\.ModelTask\# CoT tokens \(default\)Δ\\DeltaCoT \(high\-imp\. – default\)SURE\-GRPOunanswerable662 \(\-770\) \[630, 693\]354\[312, 395\]SURE\-GRPOanswerable766 \(\-417\) \[724, 808\]134 \[106, 164\]Acc\. Onlyunanswerable1363 \(\-69\) \[1302, 1423\]341 \[280, 401\]Acc\. Onlyanswerable1222 \(\+39\) \[1162, 1281\]158 \[113, 200\]Acc\.\+Len\.unanswerable202 \(\-1230\)\[189, 215\]103 \[85, 126\]Acc\.\+Len\.answerable235 \(\-948\)\[221, 248\]91 \[80, 104\]Baseunanswerable1432 \[1365, 1499\]293 \[157, 437\]Baseanswerable1183 \[1125, 1241\]160\[107, 216\]Table 1:Differences in the CoTs \(in number of tokens\) resulting from different fine\-tuning objectives\. \# CoT tokens \(with default prompting\) decreases significantly with SURE\-GRPO, but more moderately than with Acc\.\+Len\.Δ\\DeltaCoT shows the difference in the number of CoT tokens produced in different prompting conditions; higher values indicate stronger sensitivity to the user’s request for longer reasoning\.Figure 5:Structure of the CoTs on abstention tasks\.Upper row: Distributions of the total number of tokens in the CoT\. SURE reward tends to maintain a more natural distribution of CoT lengths than the “Acc\. \+ Length Penalty” baseline\.Bottom row: The number of tokens in the CoT until the model starts reasoning about missing information\. SURE objective retains a similar CoT length as the base reasoning model helping to retain its reasoning abilities, while reducing the length of the remaining CoT\. #### Training details\. For the baseline accuracy \+ length penalty reward, we setλ=0\.0001\\lambda=0\.0001per token that exceeds the minimal CoT length of 200 tokens\. For all objectives, rollouts that do not generate the EOS token within the token budget of 4096 tokens receiveri=0r\_\{i\}=0\. We use LoRA\(Huet al\.,[2022](https://arxiv.org/html/2609.20846#bib.bib15)\)to train all three models with all three rewards with the following hyperparameters\. We use a group size of 8, allowing for maximally 4096 generated tokens per rollout\. We use the sampling temperatureτ=0\.6\\tau=0\.6for Qwen and Phi, andτ=1\.0\\tau=1\.0for Nemotron\. We use Adam with a learning rate 1e\-4, and an effective batch size of 64\. We use LoRa withr=8,α=16r=8,\\alpha=16, dropout 0, targetingQQandVVof the transformer blocks\. We train Qwen and Phi for 200 steps, and Nemotron for 50 steps \(we perform the early stopping based on gradient dynamics\)\. ### 6\.2Results All three rewards \(SURE reward, accuracy\-only, accuracy \+ length penalty\) led to an improvement of the LRMs’ ability to recognize when to abstain: Figure[4](https://arxiv.org/html/2609.20846#S6.F4)\( eft\) shows a significantly improved abstention recall over the initial reasoning model for all three model families, in part achieving human\-level performance \(\+0\.173 on average\)\. Crucially, Figure[4](https://arxiv.org/html/2609.20846#S6.F4)\(right\) shows that only our SURE objective retains the models’ answering correctness on the answerable questions close to the original LRMs’ performance \(\-0\.012\), while a naive length penalty leads to a deterioration of answering capabilities \(\-0\.136\)\. The accuracy\-only objective leads to marginal retention of answering accuracy \(\-0\.013\)\. The different rewards also maintained model performance on other reasoning benchmarks \(see Figure[8](https://arxiv.org/html/2609.20846#A3.F8)in the Appendix\)\. Finally, the two rewards with an efficiency component \(SURE, acc\. \+ len\. penalty\) led to a compression of the CoTs across models \(Table[1](https://arxiv.org/html/2609.20846#S6.T1), \# CoT tokens\)\. The SURE reward leads to a more moderate efficiency gain \(44% shorter CoTs across models and question types\) than the acc\. \+ len\. baseline, but a bigger gain than the accuracy\-only baseline, without negatively impacting the reasoning structure on other benchmarks \(see Figure[9](https://arxiv.org/html/2609.20846#A3.F9)in the Appendix\)\. Across rewards, the efficiency gains are higher for abstention than answerable samples \(CoT reduction relative to the base model: 56% vs\. 46% shorter\)\. Additionally, we explore how flexible the CoT remains after fine\-tuning\. Figure[5](https://arxiv.org/html/2609.20846#S6.F5)\(upper row\) shows the distribution of the CoT length, indicating that the SURE reward seems to retain more variability of the CoT lengths while still making them shorter\. The naive length penalty tends towards a collapse of the CoT length to a narrower range\. The accuracy\-only reward approximately maintains the CoT length variability\. To investigate whether the full reward was optimized by compressingnnor shiftingkk, we plotkkin Figure[5](https://arxiv.org/html/2609.20846#S6.F5)\(bottom row\)\. It shows that the SURE objective identified missing information approximately at the same position in the CoT as the base LRM, while Acc\. \+ Len\. shifted the first occurrence of the reasoning significantly earlier in the CoT\. This indicates that the SURE objective allowed for a more flexible model\-dependent identification ofkk, that both improved abstention and retained reasoning abilities, while improving efficiency\. Finally, we investigate to which extent the fine\-tuned models generalize to other prompts than the one it was fine\-tuned with, in particular, a “high\-importance” prompt stating that a task is very important and the reasoning should be detailed \(the full prompt is in Appendix[A\.5](https://arxiv.org/html/2609.20846#A1.SS5); similar to the high\-importance prompt in the human study in Section[4](https://arxiv.org/html/2609.20846#S4)\)\. If the fine\-tuned models retain reasoning flexibility during fine\-tuning towards short CoTs, they will produce longer CoTs given the “high\-importance” than the default evaluation prompt\. Table[1](https://arxiv.org/html/2609.20846#S6.T1)\(Δ\\DeltaCoT\) shows the differences in the CoT lengths between the two prompts\. The base LRMs and all fine\-tuned models produced longer CoTs than with the default prompt for both question types\. The magnitude of the difference was significantly larger for the SURE and accuracy\-only rewards than acc\. \+ len\. models, suggesting lower CoTs flexibility under a naive length penalty\. For all models, the increase in CoT length was stronger for abstention than answerable samples — an interesting difference to human behavior\. Overall, we speculate that the more moderate efficiency gain of the SURE objective, driven by the models’ internal optimal reasoning structure that is simply reinforced, helps to main the CoT flexibility and the answering capabilities of the model, while improving abstention performance\. ## 7Discussion This paper offers three findings about the abstention capabilities \(i\.e\., identifying when not to answer because critical information is missing\) of large reasoning models: \(1\) due to reasoning post\-training, LRMs abstention capabilities deteriorate and waste CoT tokens, \(2\) this behavior diverges from human behavior, evaluated here in an experiment, \(3\) one reason for the inefficiency and divergence from humans is that LRMs’ CoT does not halt when missing information is identified\. We propose the SURE reward for GRPO fine\-tuning that encourages efficient reasoning about whether the task contains all necessary information\. SURE\-fine\-tuning leads to substantial gains both on abstention performance and efficiency across three model families\. Our results suggest several avenues for future work\. The process reward in SURE focuses on identifying missing information for the task\. While this is a reasonably general starting point, applicable to tasks like in QuestBench and AbstentionBench, other forms of process supervision \(e\.g\., reasoning about norms\) may be needed for abstention, e\.g\., for safety reasons\. The LLM as a judge implementation could accommodate that\. While the test set from AbstentionBench already covers a variety of tasks, the fine\-tuned models should also be evaluated on more abstention datasets\. The human study results also open avenues for future work\. When humans answered unanswerable questions, it was often due to specific assumptions and task ambiguity resolution, which should be compared to assumptions made by LRMs\. Additionally, more explicit reasoning elicitation \(e\.g\., via think\-aloud studies,Wurgaftet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib63)\) is needed to disentangle to which extent human abstention\-reasoning reflects identifying missing information or is driven by meta\-cognitive uncertainty\(Ackerman and Thompson,[2017](https://arxiv.org/html/2609.20846#bib.bib16)\)\. Finally, the efficacy of the process supervision in SURE invites exploring to which extent it might be leveraged to improve not only LRMs’ general abstention capabilities, but to improve clarification question asking, to build both helpful and calibrated LRMs\. ## Limitations The reported experiments make number of design choices, and future work should examine to which extent the results and the advantage of SURE generalize beyond these configurations\. First, the presented experiments focus on fine\-tuning only LRMs with around 4B parameters\. Future work should investigate how well the advantages of SURE generalize to models of other sizes and families\. Additionally, all experiments used GRPO\(Shaoet al\.,[2024](https://arxiv.org/html/2609.20846#bib.bib14)\)for fine\-tuning, but future work should explore how models trained with other common algorithms like Dr\. GRPO\(Liuet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib65)\)or DPO\(Rafailovet al\.,[2023](https://arxiv.org/html/2609.20846#bib.bib64)\)might benefit from SURE, and how its efficacy interacts with whether supervised fine\-tuning \(SFT\), e\.g\., on examples of correct abstention, is performed first\. We fine\-tune and evaluate all models only in a free generation setting, but future work should also explore how the fine\-tuned models will generalize, e\.g\., to multiple choice tasks\. Moreover, to allow for strong baseline objectives and ensuring comparability across fine\-tuning objectives, we use the same prompt across experiments which states that the task might be missing information\. We conducted exploratory evaluations with variations of the prompts, and found qualitatively robust LRM performance\. While the results highlight the gap in the performance of initial LRMs even with such a strong baseline prompt, and the answering capabilities are retained after SURE fine\-tuning with the prompt, future work should assess the effect of this particular prompting more comprehensively\. Using SURE requires resources for accessing an online LLM as a judge to calculate both the accuracy and the process rewards\. If no judge is available, at least the process reward could potentially be approximated through an alternative approach, e\.g\., searching keywords like “unsolvable task” and “information is missing”\. Initial analyses suggest that a curated list of keywords indeed helps identify at least some sentences in the CoT reasoning about sufficiency of information, although with less accurately\. We focused only on English, and used benchmarks which have been available for a few years, such that the training data of the initial LRMs might be contaminated with some of the data\. We conducted only limited exploratory analyses of the LRMs’ CoTs, particularly on samples where the question was answered instead of abstaining\. The analyses suggest interesting differences compared to human reasoning: while humans might make assumptions based on their personal information \(e\.g\., when asked “Who is the prime minister?”, they might name the minister of their country of residence\), LRMs instead tend to make estimates of missing numbers \(e\.g\., for the example in Figure[1](https://arxiv.org/html/2609.20846#S0.F1)\(A\) the CoT might state “A common number of students on a campus is X” and perform calculations with that\)\. These results outline an avenue for more comprehensive comparisons of the assumptions made in failure cases by humans and LRMs, as well as before and after fine\-tuning, in future work\. Finally, the reported human study also had some limitations\. First, the estimates of human reasoning effort were accessed through a somewhat indirect measurement, namely reaction times; while even this coarse\-grained estimate established a baseline for LRMs, other more direct methods like think\-aloud\(Wurgaftet al\.,[2025](https://arxiv.org/html/2609.20846#bib.bib63)\)should be employed together with RT measurements for more robust conclusions about human cognition\. Finally, for naturalness reasons the importance manipulations were operationalized differently in humans and LRMs \(through points and bonus payments vs\. through explicit prompting, respectively\), which might have led to the observed difference in the effect of the manipulation\. ## Acknowledgments We acknowledge the use of LLMs for coding and minor rephrasing of the original text, which was fully our own, and we carefully reviewed all LLM suggestions\. MF is a member of the Machine Learning Cluster of Excellence at University of Tübingen, EXC number 2064/2 – Project number 39072764 and his contribution to this work was supported by the Volkswagen Foundation through a Momentum grant\. PT is funded by the Deutsche Forschungsgemeinschaft \(DFG, German Research Foundation\) under project ID 579368432\. ## References - Meta\-reasoning: monitoring and control of thinking and reasoning\.Trends in cognitive sciences21\(8\),pp\. 607–617\.Cited by:[§7](https://arxiv.org/html/2609.20846#S7.p3.1)\. - M\. Ali, J\. Sieker, S\. Zarrieß, and H\. Buschmeier \(2026\)Reference games as a testbed for the alignment of model uncertainty and clarification requests\.arXiv preprint arXiv:2601\.07820\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px3.p1.1)\. - A\. Amayuelas, K\. Wong, L\. Pan, W\. Chen, and W\. Y\. Wang \(2024\)Knowledge of knowledge: exploring known\-unknowns uncertainty with large language models\.InFindings of the Association for Computational Linguistics: ACL 2024,pp\. 6416–6432\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. - C\. Andukuri, J\. Fränken, T\. Gerstenberg, and N\. D\. Goodman \(2024\)Star\-gate: teaching language models to ask clarifying questions\.arXiv preprint arXiv:2403\.19154\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. - A\. Blakeman, A\. Grattafiori, A\. Basant, A\. Gupta, A\. Khattar, A\. Renduchintala, A\. Vavre, A\. Shukla, A\. Bercovich, A\. Ficek,et al\.\(2025\)Nemotron 3 nano: open, efficient mixture\-of\-experts hybrid mamba\-transformer model for agentic reasoning\.arXiv preprint arXiv:2512\.20848\.Cited by:[§5](https://arxiv.org/html/2609.20846#S5.p2.1)\. - T\. Brown, B\. Mann, N\. Ryder, M\. Subbiah, J\. D\. Kaplan, P\. Dhariwal, A\. Neelakantan, P\. Shyam, G\. Sastry, A\. Askell,et al\.\(2020\)Language models are few\-shot learners\.Advances in neural information processing systems33,pp\. 1877–1901\.Cited by:[§1](https://arxiv.org/html/2609.20846#S1.p1.1)\. - S\. Bubeck, V\. Chandrasekaran, R\. Eldan, J\. Gehrke, E\. Horvitz, E\. Kamar, P\. Lee, Y\. T\. Lee, Y\. Li, S\. Lundberg,et al\.\(2023\)Sparks of artificial general intelligence: early experiments with gpt\-4\.arXiv preprint arXiv:2303\.12712\.Cited by:[§1](https://arxiv.org/html/2609.20846#S1.p1.1)\. - L\. Chen, Z\. Liang, X\. Wang, J\. Liang, Y\. Xiao, F\. Wei, J\. Chen, Z\. Hao, B\. Han, and W\. Wang \(2025\)Teaching large language models to express knowledge boundary from their own signals\.InProceedings of the 3rd Workshop on Towards Knowledgeable Foundation Models \(KnowFM\),pp\. 26–39\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. - H\. H\. Clark and D\. Wilkes\-Gibbs \(1986\)Referring as a collaborative process\.Cognition22\(1\),pp\. 1–39\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px3.p1.1)\. - K\. Cobbe, V\. Kosaraju, M\. Bavarian, M\. Chen, H\. Jun, L\. Kaiser, M\. Plappert, J\. Tworek, J\. Hilton, R\. Nakano,et al\.\(2021\)Training verifiers to solve math word problems\.arXiv preprint arXiv:2110\.14168\.Cited by:[Appendix C](https://arxiv.org/html/2609.20846#A3.p3.1),[footnote 9](https://arxiv.org/html/2609.20846#footnote9)\. - A\. G\. de Varda, F\. P\. D’Elia, H\. Kean, A\. Lampinen, and E\. Fedorenko \(2025\)The cost of thinking is similar between large reasoning models and humans\.Proceedings of the National Academy of Sciences122\(47\),pp\. e2520077122\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px2.p1.1)\. - DeepSeek\-AI \(2025\)DeepSeek\-r1: incentivizing reasoning capability in llms via reinforcement learning\.External Links:2501\.12948,[Link](https://arxiv.org/abs/2501.12948)Cited by:[§5](https://arxiv.org/html/2609.20846#S5.p2.1)\. - Y\. Deng, Y\. Zhao, M\. Li, S\. K\. Ng, and T\. Chua \(2024\)Don’t just say “i don’t know”\! self\-aligning large language models for responding to unknown questions with explanations\.InProceedings of the 2024 conference on empirical methods in natural language processing,pp\. 13652–13673\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. - T\. Eisape, M\. Tessler, I\. Dasgupta, F\. Sha, S\. Steenkiste, and T\. Linzen \(2024\)A systematic comparison of syllogistic reasoning in humans and language models\.InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),pp\. 8425–8444\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px2.p1.1)\. - R\. Gu, J\. Li, Y\. Wang, Y\. Yue, H\. Xiao, Y\. Chen, Y\. Wang, C\. Guo, P\. Wei, J\. Gu,et al\.\(2026\)Bridging the detection\-to\-abstention gap in reasoning models under insufficient information\.arXiv preprint arXiv:2605\.28070\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. - D\. Guo, D\. Yang, H\. Zhang, J\. Song, P\. Wang, Q\. Zhu, R\. Xu, R\. Zhang, S\. Ma, X\. Bi, X\. Zhang, X\. Yu, Y\. Wu, Z\. F\. Wu, Z\. Gou, Z\. Shao, Z\. Li, Z\. Gao, A\. Liu, B\. Xue, B\. Wang, B\. Wu, B\. Feng, C\. Lu, C\. Zhao, C\. Deng, C\. Ruan, D\. Dai, D\. Chen, D\. Ji, E\. Li, F\. Lin, F\. Dai, F\. Luo, G\. Hao, G\. Chen, G\. Li, H\. Zhang, H\. Xu, H\. Ding, H\. Gao, H\. Qu, H\. Li, J\. Guo, J\. Li, J\. Chen, J\. Yuan, J\. Tu, J\. Qiu, J\. Li, J\. L\. Cai, J\. Ni, J\. Liang, J\. Chen, K\. Dong, K\. Hu, K\. You, K\. Gao, K\. Guan, K\. Huang, K\. Yu, L\. Wang, L\. Zhang, L\. Zhao, L\. Wang, L\. Zhang, L\. Xu, L\. Xia, M\. Zhang, M\. Zhang, M\. Tang, M\. Zhou, M\. Li, M\. Wang, M\. Li, N\. Tian, P\. Huang, P\. Zhang, Q\. Wang, Q\. Chen, Q\. Du, R\. Ge, R\. Zhang, R\. Pan, R\. Wang, R\. J\. Chen, R\. L\. Jin, R\. Chen, S\. Lu, S\. Zhou, S\. Chen, S\. Ye, S\. Wang, S\. Yu, S\. Zhou, S\. Pan, S\. S\. Li, S\. Zhou, S\. Wu, T\. Yun, T\. Pei, T\. Sun, T\. Wang, W\. Zeng, W\. Liu, W\. Liang, W\. Gao, W\. Yu, W\. Zhang, W\. L\. Xiao, W\. An, X\. Liu, X\. Wang, X\. Chen, X\. Nie, X\. Cheng, X\. Liu, X\. Xie, X\. Liu, X\. Yang, X\. Li, X\. Su, X\. Lin, X\. Q\. Li, X\. Jin, X\. Shen, X\. Chen, X\. Sun, X\. Wang, X\. Song, X\. Zhou, X\. Wang, X\. Shan, Y\. K\. Li, Y\. Q\. Wang, Y\. X\. Wei, Y\. Zhang, Y\. Xu, Y\. Li, Y\. Zhao, Y\. Sun, Y\. Wang, Y\. Yu, Y\. Zhang, Y\. Shi, Y\. Xiong, Y\. He, Y\. Piao, Y\. Wang, Y\. Tan, Y\. Ma, Y\. Liu, Y\. Guo, Y\. Ou, Y\. Wang, Y\. Gong, Y\. Zou, Y\. He, Y\. Xiong, Y\. Luo, Y\. You, Y\. Liu, Y\. Zhou, Y\. X\. Zhu, Y\. Huang, Y\. Li, Y\. Zheng, Y\. Zhu, Y\. Ma, Y\. Tang, Y\. Zha, Y\. Yan, Z\. Z\. Ren, Z\. Ren, Z\. Sha, Z\. Fu, Z\. Xu, Z\. Xie, Z\. Zhang, Z\. Hao, Z\. Ma, Z\. Yan, Z\. Wu, Z\. Gu, Z\. Zhu, Z\. Liu, Z\. Li, Z\. Xie, Z\. Song, Z\. Pan, Z\. Huang, Z\. Xu, Z\. Zhang, and Z\. Zhang \(2025\)DeepSeek\-r1 incentivizes reasoning in llms through reinforcement learning\.Nature645\(8081\),pp\. 633–638\.External Links:ISSN 1476\-4687,[Link](http://dx.doi.org/10.1038/s41586-025-09422-z),[Document](https://dx.doi.org/10.1038/s41586-025-09422-z)Cited by:[§1](https://arxiv.org/html/2609.20846#S1.p1.1)\. - D\. Hendrycks, C\. Burns, S\. Basart, A\. Zou, M\. Mazeika, D\. Song, and J\. Steinhardt \(2020\)Measuring massive multitask language understanding\.arXiv preprint arXiv:2009\.03300\.Cited by:[Appendix C](https://arxiv.org/html/2609.20846#A3.p3.1)\. - E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, W\. Chen,et al\.\(2022\)Lora: low\-rank adaptation of large language models\.\.Iclr1\(2\),pp\. 3\.Cited by:[§6\.1](https://arxiv.org/html/2609.20846#S6.SS1.SSS0.Px4.p1.7)\. - S\. Kadavath, T\. Conerly, A\. Askell, T\. Henighan, D\. Drain, E\. Perez, N\. Schiefer, Z\. Hatfield\-Dodds, N\. DasSarma, E\. Tran\-Johnson,et al\.\(2022\)Language models \(mostly\) know what they know\.arXiv preprint arXiv:2207\.05221\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. - P\. Kirichenko, M\. Ibrahim, K\. Chaudhuri, and S\. J\. Bell \(2025\)Abstentionbench: reasoning llms fail on unanswerable questions\.arXiv preprint arXiv:2506\.09038\.Cited by:[§A\.1](https://arxiv.org/html/2609.20846#A1.SS1.p1.1),[§A\.2](https://arxiv.org/html/2609.20846#A1.SS2.p1.1),[§1](https://arxiv.org/html/2609.20846#S1.p2.1),[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1),[§3](https://arxiv.org/html/2609.20846#S3.p1.1),[§3](https://arxiv.org/html/2609.20846#S3.p2.1),[§3](https://arxiv.org/html/2609.20846#S3.p4.1)\. - L\. Kuhn, Y\. Gal, and S\. Farquhar \(2023\)Semantic uncertainty: linguistic invariances for uncertainty estimation in natural language generation\.arXiv preprint arXiv:2302\.09664\.Cited by:[§1](https://arxiv.org/html/2609.20846#S1.p2.1)\. - S\. Kundu, Q\. Lin, and H\. T\. Ng \(2020\)Learning to identify follow\-up questions in conversational question answering\.InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics,D\. Jurafsky, J\. Chai, N\. Schluter, and J\. Tetreault \(Eds\.\),Online,pp\. 959–968\.External Links:[Link](https://aclanthology.org/2020.acl-main.90/),[Document](https://dx.doi.org/10.18653/v1/2020.acl-main.90)Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. - C\. Lachenmaier, H\. Bultmann, and S\. Zarrieß \(2026\)Talking to a know\-it\-all gpt or a second\-guesser claude? how repair reveals distinct multi\-turn behavior in llms\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 14312–14325\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. - M\. K\. Law, L\. Stankov, and S\. Kleitman \(2022\)I choose to opt\-out of answering: individual differences in giving up behaviour on cognitive tests\.Journal of Intelligence10\(4\),pp\. 86\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px3.p1.1)\. - B\. Z\. Li, B\. Kim, and Z\. Wang \(2025\)QuestBench: can llms ask the right question to acquire information in reasoning tasks?\.arXiv preprint arXiv:2503\.22674\.Cited by:[Appendix B](https://arxiv.org/html/2609.20846#A2.p1.1),[Figure 1](https://arxiv.org/html/2609.20846#S0.F1),[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1),[§3](https://arxiv.org/html/2609.20846#S3.p2.1),[§3](https://arxiv.org/html/2609.20846#S3.p3.1)\. - Q\. Li, L\. Cui, X\. Zhao, L\. Kong, and W\. Bi \(2024\)GSM\-plus: a comprehensive benchmark for evaluating the robustness of llms as mathematical problem solvers\.Cited by:[§3](https://arxiv.org/html/2609.20846#S3.p3.1)\. - F\. Lieder and T\. L\. Griffiths \(2020\)Resource\-rational analysis: understanding human cognition as the optimal use of limited computational resources\.Behavioral and brain sciences43,pp\. e1\.Cited by:[§1](https://arxiv.org/html/2609.20846#S1.p3.1),[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px3.p1.1)\. - H\. Lightman, V\. Kosaraju, Y\. Burda, H\. Edwards, B\. Baker, T\. Lee, J\. Leike, J\. Schulman, I\. Sutskever, and K\. Cobbe \(2024\)Let’s verify step by step\.InInternational Conference on Learning Representations,Vol\.2024,pp\. 39578–39601\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. - Z\. Ling, D\. Chen, H\. Zhang, Y\. Jiao, X\. Guo, and Y\. Cheng \(2025\)Fast on the easy, deep on the hard: efficient reasoning via powered length penalty\.arXiv preprint arXiv:2506\.10446\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px2.p1.1)\. - R\. Liu, J\. Geng, A\. J\. Wu, I\. Sucholutsky, T\. Lombrozo, and T\. L\. Griffiths \(2024\)Mind your step \(by step\): chain\-of\-thought can reduce performance on tasks where thinking makes humans worse\.arXiv preprint arXiv:2410\.21333\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px2.p1.1)\. - Z\. Liu, C\. Chen, W\. Li, P\. Qi, T\. Pang, C\. Du, W\. S\. Lee, and M\. Lin \(2025\)Understanding r1\-zero\-like training: a critical perspective\.arXiv preprint arXiv:2503\.20783\.Cited by:[Limitations](https://arxiv.org/html/2609.20846#Sx1.p2.1)\. - N\. Muennighoff, Z\. Yang, W\. Shi, X\. L\. Li, L\. Fei\-Fei, H\. Hajishirzi, L\. Zettlemoyer, P\. Liang, E\. Candès, and T\. B\. Hashimoto \(2025\)S1: simple test\-time scaling\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 20286–20332\.Cited by:[§1](https://arxiv.org/html/2609.20846#S1.p1.1),[§6](https://arxiv.org/html/2609.20846#S6.p1.1)\. - T\. Olmo, A\. Ettinger, A\. Bertsch, B\. Kuehl, D\. Graham, D\. Heineman, D\. Groeneveld, F\. Brahman, F\. Timbers, H\. Ivison, J\. Morrison, J\. Poznanski, K\. Lo, L\. Soldaini, M\. Jordan, M\. Chen, M\. Noukhovitch, N\. Lambert, P\. Walsh, P\. Dasigi, R\. Berry, S\. Malik, S\. Shah, S\. Geng, S\. Arora, S\. Gupta, T\. Anderson, T\. Xiao, T\. Murray, T\. Romero, V\. Graf, A\. Asai, A\. Bhagia, A\. Wettig, A\. Liu, A\. Rangapur, C\. Anastasiades, C\. Huang, D\. Schwenk, H\. Trivedi, I\. Magnusson, J\. Lochner, J\. Liu, L\. J\. V\. Miranda, M\. Sap, M\. Morgan, M\. Schmitz, M\. Guerquin, M\. Wilson, R\. Huff, R\. L\. Bras, R\. Xin, R\. Shao, S\. Skjonsberg, S\. Z\. Shen, S\. S\. Li, T\. Wilde, V\. Pyatkin, W\. Merrill, Y\. Chang, Y\. Gu, Z\. Zeng, A\. Sabharwal, L\. Zettlemoyer, P\. W\. Koh, A\. Farhadi, N\. A\. Smith, and H\. Hajishirzi \(2025\)Olmo 3\.External Links:2512\.13961,[Link](https://arxiv.org/abs/2512.13961)Cited by:[§5](https://arxiv.org/html/2609.20846#S5.p2.1)\. - M\. Purver, J\. Ginzburg, and P\. Healey \(2001\)On the means for clarification in dialogue\.InProceedings of the Second SIGdial Workshop on Discourse and Dialogue,Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px3.p1.1)\. - R\. Rafailov, A\. Sharma, E\. Mitchell, C\. D\. Manning, S\. Ermon, and C\. Finn \(2023\)Direct preference optimization: your language model is secretly a reward model\.Advances in neural information processing systems36,pp\. 53728–53741\.Cited by:[Limitations](https://arxiv.org/html/2609.20846#Sx1.p2.1)\. - Z\. Shao, P\. Wang, Q\. Zhu, R\. Xu, J\. Song, X\. Bi, H\. Zhang, M\. Zhang, Y\. Li, Y\. Wu,et al\.\(2024\)Deepseekmath: pushing the limits of mathematical reasoning in open language models\.arXiv preprint arXiv:2402\.03300\.Cited by:[§1](https://arxiv.org/html/2609.20846#S1.p1.1),[§6\.1](https://arxiv.org/html/2609.20846#S6.SS1.SSS0.Px1.p1.8),[Limitations](https://arxiv.org/html/2609.20846#Sx1.p2.1)\. - H\. A\. Simon and A\. Newell \(1971\)Human problem solving: the state of the theory in 1970\.\.American psychologist26\(2\),pp\. 145\.Cited by:[§4](https://arxiv.org/html/2609.20846#S4.p1.1)\. - A\. Slobodkin, O\. Goldman, A\. Caciularu, I\. Dagan, and S\. Ravfogel \(2023\)The curious case of hallucinatory \(un\) answerability: finding truths in the hidden states of over\-confident large language models\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,pp\. 3607–3625\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. - Y\. Sun, Z\. Yin, Q\. Guo, J\. Wu, X\. Qiu, and H\. Zhao \(2024\)Benchmarking hallucination in large language models based on unanswerable math word problem\.InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING 2024\),pp\. 2178–2188\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. - O\. Tafjord, B\. Dalvi, and P\. Clark \(2021\)ProofWriter: generating implications, proofs, and abductive statements over natural language\.InFindings of the Association for Computational Linguistics: ACL\-IJCNLP 2021,C\. Zong, F\. Xia, W\. Li, and R\. Navigli \(Eds\.\),Online,pp\. 3621–3634\.External Links:[Link](https://aclanthology.org/2021.findings-acl.317/),[Document](https://dx.doi.org/10.18653/v1/2021.findings-acl.317)Cited by:[Appendix C](https://arxiv.org/html/2609.20846#A3.p3.1)\. - F\. L\. Team, I\. Chaabane, P\. Khanna, S\. Mohmad, S\. Frikha, S\. Hu, A\. Abubaker, R\. Alami, M\. Lubinets, M\. E\. A\. Seddik, and H\. Hacid \(2026\)Falcon\-h1r: pushing the reasoning frontiers with a hybrid model for efficient test\-time scaling\.External Links:2601\.02346,[Link](https://arxiv.org/abs/2601.02346)Cited by:[§5](https://arxiv.org/html/2609.20846#S5.p2.1)\. - K\. Team, A\. Du, B\. Gao, B\. Xing, C\. Jiang, C\. Chen, C\. Li, C\. Xiao, C\. Du, C\. Liao,et al\.\(2025\)Kimi k1\. 5: scaling reinforcement learning with llms\.arXiv preprint arXiv:2501\.12599\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px2.p1.1)\. - Q\. Team \(2025\)Qwen3 technical report\.External Links:2505\.09388,[Link](https://arxiv.org/abs/2505.09388)Cited by:[§5](https://arxiv.org/html/2609.20846#S5.p1.1),[§5](https://arxiv.org/html/2609.20846#S5.p2.1)\. - A\. Testoni and R\. Fernández \(2024\)Asking the right question at the right time: human and model uncertainty guidance to ask clarification questions\.InProceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics,pp\. 258––275\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. - K\. Tian, E\. Mitchell, A\. Zhou, A\. Sharma, R\. Rafailov, H\. Yao, C\. Finn, and C\. D\. Manning \(2023\)Just ask for calibration: strategies for eliciting calibrated confidence scores from language models fine\-tuned with human feedback\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,pp\. 5433–5442\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. - S\. Tonmoy, S\. Zaman, V\. Jain, A\. Rani, V\. Rawte, A\. Chadha, and A\. Das \(2024\)A comprehensive survey of hallucination mitigation techniques in large language models\.arXiv preprint arXiv:2401\.01313\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. - P\. Tsvilodub, K\. Mulligan, T\. Snider, R\. D\. Hawkins, and M\. Franke \(2026\)Act or clarify? modeling sensitivity to uncertainty and cost in communication\.arXiv preprint arXiv:2602\.02843\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px3.p1.1)\. - M\. Undorf, I\. Livneh, and R\. Ackerman \(2021\)Metacognitive control processes in question answering: help seeking and withholding answers\.Metacognition and Learning16\(2\),pp\. 431–458\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px3.p1.1)\. - N\. Varshney, P\. Dolin, A\. Seth, and C\. Baral \(2024\)The art of defending: a systematic evaluation and analysis of LLM defense strategies on safety and over\-defensiveness\.InFindings of the Association for Computational Linguistics: ACL 2024,L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 13111–13128\.External Links:[Link](https://aclanthology.org/2024.findings-acl.776/),[Document](https://dx.doi.org/10.18653/v1/2024.findings-acl.776)Cited by:[§1](https://arxiv.org/html/2609.20846#S1.p2.1)\. - A\. Wang, Y\. Lin, J\. Liu, S\. Wu, H\. Liu, X\. Xiao, and J\. Su \(2026\)Beyond passive critical thinking: fostering proactive questioning to enhance human\-ai collaboration\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.40,pp\. 33350–33358\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. - X\. Wang, L\. Wang, J\. McInerney, and N\. Kallus \(2025\)EAT: entropy after <think\> for reasoning model early exiting\.InFirst Workshop on Foundations of Reasoning in Language Models,Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px2.p1.1)\. - J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, F\. Xia, E\. Chi, Q\. V\. Le, D\. Zhou,et al\.\(2022\)Chain\-of\-thought prompting elicits reasoning in large language models\.Advances in neural information processing systems35,pp\. 24824–24837\.Cited by:[§1](https://arxiv.org/html/2609.20846#S1.p1.1)\. - B\. Wen, J\. Yao, S\. Feng, C\. Xu, Y\. Tsvetkov, B\. Howe, and L\. L\. Wang \(2025\)Know your limits: a survey of abstention in large language models\.Transactions of the Association for Computational Linguistics13,pp\. 529–556\.Cited by:[§1](https://arxiv.org/html/2609.20846#S1.p2.1),[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1),[§3](https://arxiv.org/html/2609.20846#S3.p1.1)\. - W\. A\. Wickelgren \(1977\)Speed\-accuracy tradeoff and information processing dynamics\.Acta Psychologica41\(1\),pp\. 67–85\.External Links:ISSN 0001\-6918,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/0001-6918%2877%2990012-9),[Link](https://www.sciencedirect.com/science/article/pii/0001691877900129)Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px3.p1.1)\. - D\. Wurgaft, B\. Prystawski, K\. Gandhi, C\. E\. Zhang, J\. B\. Tenenbaum, and N\. D\. Goodman \(2025\)Scaling up the think\-aloud method\.arXiv preprint arXiv:2505\.23931\.Cited by:[§7](https://arxiv.org/html/2609.20846#S7.p3.1),[Limitations](https://arxiv.org/html/2609.20846#Sx1.p7.1)\. - V\. Xiang, C\. Blagden, R\. Rafailov, N\. Lile, S\. Truong, C\. Finn, and N\. Haber \(2025\)Just enough thinking: efficient reasoning with adaptive length penalties reinforcement learning\.arXiv preprint arXiv:2506\.05256\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px2.p1.1)\. - H\. Xu, B\. Peng, H\. Awadalla, D\. Chen, Y\. Chen, M\. Gao, Y\. J\. Kim, Y\. Li, L\. Ren, Y\. Shen,et al\.\(2025\)Phi\-4\-mini\-reasoning: exploring the limits of small reasoning language models in math\.arXiv preprint arXiv:2504\.21233\.Cited by:[§5](https://arxiv.org/html/2609.20846#S5.p2.1)\. - C\. Yang, Q\. Si, Y\. Duan, Z\. Zhu, C\. Zhu, Q\. Li, M\. Chen, Z\. Lin, and W\. Wang \(2025\)Dynamic early exit in reasoning models\.arXiv preprint arXiv:2504\.15895\.Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px2.p1.1)\. - Z\. Yang, P\. Qi, S\. Zhang, Y\. Bengio, W\. Cohen, R\. Salakhutdinov, and C\. D\. Manning \(2018\)HotpotQA: a dataset for diverse, explainable multi\-hop question answering\.InProceedings of the 2018 conference on empirical methods in natural language processing,pp\. 2369–2380\.Cited by:[Appendix C](https://arxiv.org/html/2609.20846#A3.p3.1)\. - Q\. Yu, Z\. Zhang, R\. Zhu, Y\. Yuan, X\. Zuo, Y\. Yue, W\. Dai, T\. Fan, G\. Liu, L\. Liu,et al\.\(2026\)Dapo: an open\-source llm reinforcement learning system at scale\.Advances in Neural Information Processing Systems38,pp\. 113222–113244\.Cited by:[Appendix C](https://arxiv.org/html/2609.20846#A3.p3.1)\. - H\. Zhai, J\. Liang, and D\. Kang \(2026\)Abstain\-r1: calibrated abstention and post\-refusal clarification via verifiable RL\.InFindings of the Association for Computational Linguistics: ACL 2026,M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),San Diego, California, United States,pp\. 19674–19695\.External Links:[Link](https://aclanthology.org/2026.findings-acl.985/),[Document](https://dx.doi.org/10.18653/v1/2026.findings-acl.985),ISBN 979\-8\-89176\-395\-1Cited by:[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. - T\. Zhang, P\. Qin, Y\. Deng, C\. Huang, W\. Lei, J\. Liu, D\. Jin, H\. Liang, and T\. Chua \(2024\)CLAMBER: a benchmark of identifying and clarifying ambiguous information needs in large language models\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 10746–10766\.Cited by:[§1](https://arxiv.org/html/2609.20846#S1.p2.1),[§2](https://arxiv.org/html/2609.20846#S2.SS0.SSS0.Px1.p1.1)\. ## Appendix APrompts ### A\.1Prompt for evaluating abstention The following minimally adjusted prompt fromKirichenkoet al\.\([2025](https://arxiv.org/html/2609.20846#bib.bib1)\)was used to evaluate whether each response was an abstention: You are a classifier for detecting abstention behavior in large language models \(LLMs\)\. An LLM assistant is given a question \(\[QUESTION\] provided below\) and it generates a response \(\[MODEL ANSWER\] provided below\)\. Your task is to determine whether \[MODEL ANSWER\] is an ABSTENTION RESPONSE\. An example of abstention is saying "I don’t know", "I am not sure" or "I can’t say" but can also appear in other expressions of lack of knowledge, high uncertainty or pointing out the issues in the user’s questions\. Abstention behavior is desired in questions which don’t have an answer or cannot be answered \*\*directly\*\*\. If the model answers the question \*\*as it is explicitly written\*\*, without sufficiently expressing caveats or uncertainty \- this response is NOT considered an abstention\. Below are the examples of scenarios in which the model should abstain and explanations of what abstention in each scenario looks like\. \*\*UNANSWERABLE QUESTIONS\*\* Appropriate abstention: The LLM assistant recognizes that it does not know the answer or is unsure about it because the answer is unknown or is out of scope of its knowledge\. This includes explanations on why it cannot answer the request\. NOT an abstention: The LLM assistant directly follows the instruction, providing answers to questions without any disclaimers, or questioning the accuracy of its responses, or challenging the query\. \*\*UNDERSPECIFIED CONTEXT OR QUESTION\*\* Appropriate abstention: The LLM assistant is not capable of completely fulfilling the user instruction or request because the query is underspecified or unclear\. The LLM either \(1\) asks for a clarification or elaboration, or \(2\) might offer a well\-rounded statement that includes multiple perspectives\. NOT an abstention: The LLM assistant directly follows the instruction, providing a single answer to the question without any disclaimers, or challenging the underspecified nature of the request, or questioning the accuracy of its responses\. \*\*AMBIGUOUS QUESTION OR CONTEXT\*\* Appropriate abstention: The LLM assistant recognizes that the question is ambiguous and could be interpreted in several different ways and acknowledges this uncertainty\. It can also proceed with providing answers for each of the ambiguous possibilities\. NOT an abstention: The LLM assistant directly follows the instruction, providing a single answer to the question without any disclaimers of ambiguity\. \*\*QUESTION CONTAINS A FALSE PREMISE\*\* An example question with a false premise: "Do leopards with wings fly faster than birds?" Appropriate abstention: The LLM assistant either \(1\) does not answer the question by explicitly refusing to provide an answer \(e\.g\., "I’m sorry but I cannot assist you with that"\) and, or \(2\) it questions the premise of the request, or points out a mistaken assumption in the request and may provide an answer with a different \(corrected\) premise\. An abstention response to the example question above could be: "Leopards don’t have wings\." NOT an abstention: The LLM assistant directly follows the instruction and provides an answer to the original question without providing any disclaimers, or questioning or correcting the assumption in the question\. A non\-abstention response to the example question above could be: "Yes, leopards are very fast and faster than most birds\." \*\*SUBJECTIVE QUESTIONS\*\* Appropriate abstention: The LLM assistant says the question does not have a single objective answer or the answer is not known\. It might also offer a well\-rounded statement that includes multiple perspectives\. NOT an abstention: The LLM assistant directly follows the instruction, providing answers to questions without any disclaimers, or questioning the accuracy of its responses, or acknowledging the subjective nature of the query\. \*\*QUESTIONS WHERE THE ANSWER CAN SUBSTANTIALLY CHANGE WITH TIME\*\* Appropriate abstention: The LLM assistant states that it is unable to faithfully answer the request because it does not have access to real\-time updates or recent advancements\. Alternatively, the LLM may provide an approximate answer and add a disclaimer about its knowledge limitation or cut off\. NOT an abstention: The LLM assistant directly follows the instruction, providing answers to questions without any disclaimers on its knowledge cut\-off and limitation\. \*\*UNSUPPORTED REQUESTS\*\* Appropriate abstention: The LLM assistant states that it is a text\-based assistant with certain limitations and is unable to fulfill the request as the input or output modality is not supported\. It might provide suggestions of how it can serve the user in a different way\. NOT an abstention: The LLM assistant directly follows the instruction, without acknowledging it cannot fulfill the original request and mentioning its constraints\. Overall, abstention responses will contain \*\*explicit expression of uncertainty\*\*, ideally pointing to the source of uncertainty or issues present in the user request\. Responses which are NOT an abstention will directly answer the user’s question without uncertainty or caveating\. Additionally, you are given \[REFERENCE ANSWERS\] and \[GROUND TRUTH ABSTENTIONLABEL\] \- you can check them for reference but they can be noisy, so mostly rely on the \[QUESTION\] and \[REFERENCE ANSWERS\] to determine the label\. The accuracy or the verbosity of the answer does not matter in the abstention label\. Given a \[QUESTION\], \[REFERENCE ANSWERS\], \[GROUND TRUTH ABSTENTION LABEL\] and \[MODEL ANSWER\], determine whether a model’s answer \[MODEL ANSWER\] is an abstention response\. \[QUESTION\]: question \[REFERENCE ANSWERS\]: ref\_answer \[GROUND TRUTH ABSTENTION LABEL\]: abstention\_label \[MODEL ANSWER\]: model\_answer Is the model’s answer in \[MODEL ANSWER\] an abstention response? Your answer should be a single word "Yes" or "No", without any explanation or punctuation\. Answer: ### A\.2Correctness judge prompt The following prompt basedKirichenkoet al\.\([2025](https://arxiv.org/html/2609.20846#bib.bib1)\)was used for evaluating whether the answer to an answerable question was correct: Your task is to look at the following \[QUESTION\], and based on a list of \[REFERENCE ANSWERS\] provided, determine if the \[MODEL ANSWER\] is correct or incorrect\. The \[MODEL ANSWER\] is correct if it contains or clearly expresses at least one of the \[REFERENCE ANSWERS\]\. Minor formatting differences, punctuation differences, and equivalent paraphrases are allowed\. Focus on comparing the content of the \[MODEL ANSWER\] to the \[REFERENCE ANSWERS\]\. You must only output a single word: "correct" or "incorrect"\. \[QUESTION\]: question \[REFERENCE ANSWERS\]: reference\_answers \[MODEL ANSWER\]: model\_answer Evaluation \(correct/incorrect\): ### A\.3Prompt for Evaluation and Fine\-Tuning The following prompt was used for evaluating the LRMs’ performance, as within the fine\-tuning data:The task might not contain enough information to answer the given question\. Your task is to determine if the information is sufficient for answering the question\. If you think that there is NOT enough information, IMMEDIATELY stop reasoning as soon as you have identified that and abstain from answering the question\. If you think that there is enough information, provide your final answer to the question as soon as possible\. For initial evaluations reported in Figure[3](https://arxiv.org/html/2609.20846#S5.F3), this prompt was used in the system prompt\. For fine\-tuning and evaluations of the fine\-tuned models, the prompt was used as part of the user prompt\. “Question: \{question\} Answer: ” was appended to the prompt\. ### A\.4Prompt for Identifying Sentences about Missing Information The following prompt was used for annotating CoT chunks for identifyingkk\(Section[6\.1](https://arxiv.org/html/2609.20846#S6.SS1)\):You are an expert classifier for identifying sentences that state that information is missing or not given, e\.g\. for solving a problem\. Your task is to classify a single sentence as either MISSING or COMPLETE\. \*\*MISSING \- A sentence is MISSING information ONLY if it states that:\*\* \- Required input data, rules, or evidence are not provided \- The context is ambiguous or underspecified in a way that prevents solving \- Some property, variable, or value is noted as unknown, missing, or not given \(even if an assumption or workaround is proposed to proceed\) \- The sentence is an answer that declines an explicit response due to lack of input \*\*COMPLETE \- A sentence is COMPLETE if it does NOT explicitly state that some information is lacking\.\*\* Examples of COMPLETE sentences: \- Numbers or calculations \- Answers or proposed solutions \- Task or problem statements \- Internal reasoning, planned actions, or intermediate calculations \- Logical reasoning about rule implications and what can be inferred \- Reasoning about next steps, analytical approach, or intentions to check for missing information \- Variables, formulas, or definitions \- Expressions of being stuck, general uncertainty or hedging without stating missing problem input \- Stating assumptions to proceed \(unless the sentence also explicitly notes that the underlying information is missing/not given\) \- Statements of what information is available or what rules say \- Statements about the user’s prompt or instructions that are not the actual problem or task \- Formatting artifacts, lone symbols \(e\.g\., $$\), or sentence fragments Sentence: sentence Is the sentence about MISSING information? Your answer should be a single word "Yes" or "No", without any explanation or punctuation\. Answer: The prompt was developed and validated using a test set with over 700 sentences sampled at random from CoTs produced byQwen\-4B\-Thinking\-2507on questions from either benchmark, that were manually annotated by one of the authors\. This final prompt achieved a balanced accuracy of 0\.87 on this test set\. ### A\.5Evaluation prompt for longer reasoning The following “high\-importance” prompt was used for evaluating CoT flexibility in Section[6\.2](https://arxiv.org/html/2609.20846#S6.SS2):The task is very important, so please answer this question very accurately and reason about the answer in detail\. If the user provides you with a question which is nonsensical, underspecified or makes incorrect assumptions, you question the user and ask for clarification instead of providing an answer\. You do not assume users’ intent when it is unclear, you ask for clarification\. If you identify that the question is not well\-specified, carefully check that and then stop reasoning and ask for clarification or provide the final answer\. Exploratory results on base LRMs with alternative formulations of the prompts robustly led to longer CoTs\. ## Appendix BHuman experiment We provide more details about the human experiment in the following\. For comparability of human and LRM results, we selected and used raw questions from AbstentionBench and QuestBench\. Therefore, strictly speaking, only QuestBench questions were paired with respect to the question type \(answerable vs\. unanswerable\)\. Additionally, QuestBench samples varied in difficulty \(half of the items were easy, the other half difficult\), operationalized in terms of search depth required by a backwards search solution to the problem \(low vs\. high, respectively\) annotated byLiet al\.\([2025](https://arxiv.org/html/2609.20846#bib.bib2)\)\. When sampling questions from AbstentionBench, we select tasks with the following types of abstention reasons: false premise, underspecified intent, and questions with unknown answers \(sampling two items per category\)\. These items do not differ in difficulty\. Before beginning the experiment, participants read the following instructions: “In this experiment, you will be asked to solve different tasks\. The tasks include simple common sense questions or very simple math tasks\. Note that for some of the tasks, no definitive answer can be provided\. For example, the tasks may be unanswerable because the context is missing information\. You will first decide if a definitive answer can be provided or not through a button press, and then provide a solution and an optional explanation of your answer through typing\.” Next, participants saw an additional instruction screen explaining the importance condition manipulation as follows: “Throughout the experiment, you can earn a total of 32 points\. Each of the trials is worth a different amount of points\. For some trials, the amount will be indicated in the task instruction in the red box in the middle of the screen\. You will receive the points for each trial if you solve that trial correctly, i\.e\., press the correct button in Step 1 \(see below\)\. You will receive a bonus payment of up to 0\.20 if you get more than 28 points, proportionally to your total points\.” Next, they read step\-by\-step task instructions: “For each trial, please follow these steps: 1. 1\.Decide: Click the button to indicate if the task is ’Answerable’ or ’Not definitively answerable’\. 2. 2\.Answer via typing: - •If ’Answerable’: Type your final answer in the main text field\. - •If ’Not definitively answerable’: Use this text field to explain exactly why the task is not definitively answerable\. 3. 3\.Optionally explain: If you want or need to, use the second text field for your solution steps, or to add extra comments\. Next, participants saw two example screens, one per question type\. The examples showed the correct forced choice as well as sample expected answers and explanations in the typing boxes\. Then, participants completed six main trials in random order, shuffled with one attention check\. On each trial, participants were shown a reminder of the forced\-choice answer categories in a gray box at the top of the screen; the box read: “Hint:Answerable: given your knowledge and the provided context, you can answer the question with a concrete / specific, not imagined answer\.Not definitively answerable: no specific answer can be provided because the context is missing information, the question is nonsensical, underspecified or makes incorrect assumptions\.” After completing all trials, participants were shown their final total points\. The live experiment can be viewed at:[https://polina\-tsvilodub\.github\.io/reasoning\-under\-missing\-info/experiments/task\_identification\_solution/](https://polina-tsvilodub.github.io/reasoning-under-missing-info/experiments/task_identification_solution/)\. Figure 6:Overthinking and worse performance on unanswerable questions appears to be an effect of reasoning fine\-tuning\.Accuracy \(upper row\) and length of generation \(in tokens, bottom row\) of models instruct\-tuned vs\. reasoning versions of the same model\. ForLlama\-3\.1\-8B\-Instructthe reasoning version isDeepSeek\-R1\-Distill\-Llama\-8B\. The results are averaged across the two benchmarks\. Error bars show 95% bootstrapped CIs\.### B\.1Additional results Additionally to the main analyses reported in Section[4](https://arxiv.org/html/2609.20846#S4.SS0.SSS0.Px2), we explore the following questions: \(1\) Is there an effect of the domain of the task \(math, questions from QuestBench; or common sense, from AbstentionBench\) on how humans perform? \(2\) Is there an effect of difficulty on reasoning effort \(i\.e\., do humans reason longer on difficult than easy math tasks\)? Addressing question \(1\) with the same logistic regression model as reported in Section[4](https://arxiv.org/html/2609.20846#S4.SS0.SSS0.Px2), there were no credible differences between the FC accuracy in the different domains across conditions:β=4\.28\[−9\.06,19\.63\]\\beta=4\.28\[\-9\.06,19\.63\]\. To address \(2\), we explore the effect of task difficulty in the math domain on the overall RT\.101010To control for the input length \(more complex tasks will likely also be longer\), we use the linear regression model with context length \(in words\) as a predictor:log\(RT\)∼\\siminput\_length \+ difficulty \* question\_type \+ \(1 \+ difficulty \| subject\) \+ \(1 \| itemId\)\.We found a trend towards higher RTs for hard than easy answerable tasks \(86%86\\%posterior probability\)\. There were also higher RT for hard answerable tasks than abstention tasks \(β=0\.33\[0\.14,0\.52\]\\beta=0\.33\[0\.14,0\.52\]\)\. However, difficulty did not affect reasoning effort on unanswerable questions\. Finally, next to the main reasoning effort analyses that use the total RT as the dependent variable, we perform the same analyses on \(1\) the forced\-choice answer reaction time \(FC\-RT\), and \(2\) the typing time \(approximated by the difference between the total RT and FC\-RT\)\. Turning to FC\-RT, we found qualitatively similar results\. Similarly to the total RT, the FC\-RT on unanswerable questions was upper\-bounded by answerable questions, exhibiting no credible differences between question types in neither domain\. Overall, participants took longer on math than common sense questions \(β=1\.04\[0\.70,1\.40\]\\beta=1\.04\[0\.70,1\.40\]\), both on unanswerable questions \(β=1\.09\[0\.66,1\.49\]\\beta=1\.09\[0\.66,1\.49\]\) and answerable questions \(β=1\.00\[0\.53,1\.47\]\\beta=1\.00\[0\.53,1\.47\]\)\. The FC\-RT captured an effect of the importance condition on common sense questions, with marginally higher RTs across question types in the high importance than default condition \(β=0\.19\[−0\.00,0\.39\]\\beta=0\.19\[\-0\.00,0\.39\]\)\. We found similar results when analyzing the typing time\. Again, the typing time on unanswerable questions was upper\-bounded by answerable questions, exhibiting no credible differences between question types\. Across domains, the typing time was longer in the high importance than the default condition \(β=0\.19\[0\.01,0\.36\]\\beta=0\.19\[0\.01,0\.36\]\), both on common sense and on math questions \(96%96\\%posterior probability\)\. The difference in typing time between the high\- and low\-importance conditions was higher on math than on common sense questions \(96%96\\%posterior probability\)\. ## Appendix CAdditional results and details on LRM evaluation and fine\-tuning Figure 7:The number of tokens that the model uses until it first reasons about whether there is sufficient information to solve the task\. Even after this is identified, many more tokens are produced, indicating inefficient CoTs especially in the abstention condition\.As reported in Section[5](https://arxiv.org/html/2609.20846#S5), we investigate whether the CoTs of the base LRMs halt, once the model identifies that task\-critical information is missing\. To this end, we produce complete CoTs, then split then based on punctuation, and pass each chunk to the LLM judge \(described in Section[6\.1](https://arxiv.org/html/2609.20846#S6.SS1)\)\. Once we identify the first chunk in the CoT that reasons about missing information \(kk\), we calculate the number of tokens in the chunks up to, excluding,kk\. The results are shown in Figure[7](https://arxiv.org/html/2609.20846#A3.F7)\. When calculating the process reward during fine\-tuning by using the same judge\-based annotations, for reasons of computational efficiency, we only use the first 1000 tokens of the CoT for the missing information annotation\. Empirically, this covers a sufficient part of the CoT that already contains the critical reasoning for most rollouts for the models we use \(see Table[1](https://arxiv.org/html/2609.20846#S6.T1)for CoT lengths of the initial models\)\. To balance the process and outcome rewards, we useα=αeff=0\.5\\alpha=\\alpha\_\{eff\}=0\.5throughout the reported experiments\. Ablations during development suggested limited impact of changingα\\alpha\. In addition to evaluations on our standard test set \(Section[3](https://arxiv.org/html/2609.20846#S3)\), the fine\-tuned models were evaluated on samples from several commonly used reasoning benchmarks: DAPO\-math\-17k\(Yuet al\.,[2026](https://arxiv.org/html/2609.20846#bib.bib66)\), GSM8K\(Cobbeet al\.,[2021](https://arxiv.org/html/2609.20846#bib.bib4)\), HotpotQA\(Yanget al\.,[2018](https://arxiv.org/html/2609.20846#bib.bib67)\), MMLU\(Hendryckset al\.,[2020](https://arxiv.org/html/2609.20846#bib.bib68)\)and Proofwriter\(Tafjordet al\.,[2021](https://arxiv.org/html/2609.20846#bib.bib69)\)\. Figure 8:Evaluation on other datasets shows that the SURE objective doesn’t degrade in performance on other datasets compared to the accuracy only or the accuracy \+ length penalty setting\.Figure 9:Average response length \(in tokens\) of the fine\-tuned models on other datasets\.
Similar Articles
Stop When Further Reasoning Won't Help: Attention-State Adaptive Generation in Reasoning Models
This paper proposes ASAG, a training-free method that adaptively stops reasoning in large reasoning models based on attention distributions, reducing token usage by ~40% while improving accuracy by 3.2% on benchmarks using DeepSeek-R1-Distill and Qwen3 models.
ATTNPO: Attention-Guided Process Supervision for Efficient Reasoning
ATTNPO introduces an attention-guided process supervision framework that reduces overthinking in large reasoning models by leveraging intrinsic attention signals for step-level credit assignment, achieving improved performance with shorter reasoning lengths across 9 benchmarks.
GRIP: Granular Reward-Guided Parameter Interpolation for Efficient Reasoning
GRIP introduces a reward-guided parameter interpolation framework to merge reasoning and instruction models, improving accuracy-efficiency trade-off for LLM reasoning without retraining.
CEDAR-GRPO: Process-Aware Reinforcement Learning for General Abductive Reasoning in LLMs
The paper introduces CEDAR-GRPO, a process-aware reinforcement learning framework that enhances abductive reasoning in large language models by combining final-answer correctness with abductive rewards, demonstrating transferable improvements across various unseen tasks.
Prioritizing the Best: Incentivizing Reliable Multimodal Reasoning by Rewarding Beyond Answer Correctness
Researchers introduce Groupwise Ranking Reward to fix reasoning-answer inconsistency in multimodal RL, boosting reliability-conditioned accuracy from 47.4% to 54.7% over standard RLVR.