PaperlayerView exact revision on arXiv

Source-linked AI summary

When Should LLMs Abstain? Chain-of-Self-Questioning for Selective Risk Control

Ali Şenol

arXiv:2609.17516v1cs.CLcs.AI

TL;DR

Fluent factual answers can lack adequate support, while conventional question-answering evaluations may reward commitment without measuring whether answers should be withheld. CoSQ prompts models to assess required information before deciding whether to answer; on TruthfulQA-MC, Grounded-CoSQ at τ = 0.90 lowered wrong commitments and raised answered accuracy versus CoT.

  • Problem

    Factual question-answering evaluations may reward commitment even when a model lacks reliable support, although unsupported answers can be consequential.

    Source: §1 · p. 1

  • Method

    CoSQ is a prompt-only framework that identifies needed information, assesses its support, and answers from accepted units or abstains.

    Source: §1 · p. 2

  • Results

    At τ = 0.90, Grounded-CoSQ lowered mean HR from 0.131 under CoT to 0.089 and raised AA from 0.869 to 0.897, with 0.876 coverage.

    Source: §6 · p. 14

  • Takeaways & Limitations

    Grounded-CoSQ offers a selective operating point when unsupported commitments are more costly than abstention, while Critical-CoSQ favors broader coverage and Adaptive-CoSQ the lowest mean wrong-commitment rate.

    Source: §8 · p. 15

    The common prompt protocol may leave performance on the table for some models, and robustness to paraphrases, instruction order, and confidence wording remains open.

    Source: §7 · p. 15

Abstract

from arXiv · show

Large language models can produce fluent answers when their factual support is weak. This paper introduces Chain-of-Self-Questioning (CoSQ), a prompt-only framework that makes answer commitment conditional on an explicit assessment of the information required to answer a question. We evaluate three CoSQ variants under seventeen conditions on the 817-item TruthfulQA multiple-choice validation set using eleven open-weight and hosted model families. In the final balanced-option protocol, Grounded-CoSQ at {\tau}=0.90 reduces the mean unconditional wrong-commitment rate from 13.1% under chain-of-thought prompting to 8.9%, a 32.1% relative reduction, while increasing answered accuracy from 86.9% to 89.7% and answering 87.6% of questions. Both improvements hold for all eleven models and at every evaluated threshold. Critical-CoSQ and Adaptive-CoSQ provide neighboring operating points with 88.6% and 86.5% coverage, respectively, while remaining more reliable than the baseline. A secondary Natural Questions Short-Answer evaluation provides convergent open-form evidence. These findings show that self-assessment can support explicit, tunable answer-or-abstain decisions when an unsupported commitment is more costly than referral or review.

1 Introduction

The introduction motivates abstention as a way to manage unsupported factual answers and presents CoSQ as a prompt-only framework separating information assessment from commitment. It evaluates selective policies across models and frames risk reduction, reliability, and threshold choice as its central questions.

1 Introduction

  • When factual support is incomplete or wrong, a fluent answer can be unsupported; abstention can instead trigger verification or human review.The introduction notes this trade-off in high-consequence settings including healthcare, law, finance, and public administration.Source: §1 · p. 1
  • Selective evaluation pairs answered accuracy with coverage and wrong-commitment rate, since aggregate accuracy alone may reward guessing.The paper frames rejection as a means of reducing risk on the inputs a system answers [7].Source: §1 · p. 1
  • CoSQ separates information assessment from answer commitment in a prompt-only selective-answering framework.The model identifies required information, assesses its support, then answers using accepted information or abstains.Source: §1 · p. 2
  • The study evaluates CoSQ across eleven model families on deterministic, balanced-option TruthfulQA and compares grounded, critical-item, and adaptive variants.It reports a threshold sweep rather than a single post-hoc threshold.Source: §1 · p. 2

1.1 Hypotheses and research questions

  • The hypotheses test whether Grounded-CoSQ at τ = 0.90 reduces wrong commitments and improves answered accuracy versus forced-choice CoT; the questions examine threshold control and generalization.Generalization is examined across model families and in a secondary Natural Questions Short-Answer evaluation.Source: §1.1 · p. 2

2 Related Work

Related work situates CoSQ among hallucination, uncertainty, grounding, and selective-prediction research. It distinguishes a prompt-level answer-or-abstain decision from answer-generation improvements, retrieval, and post-hoc verification.

2.1 Hallucination and truthfulness

  • TruthfulQA tests whether models reproduce common misconceptions, while multiple-choice scoring avoids fragile open-form semantic parsing.The paper also notes that ordinary accuracy incentives can encourage guessing, motivating abstention as an observable decision.Source: §2.1 · p. 3

2.2 Reasoning, uncertainty, and self-knowledge

  • CoSQ converts an elicited confidence signal into an answer-or-abstain decision without claiming human-like introspective access or calibrated probability.The paper treats the signal as useful when it changes the risk–coverage profile.Source: §2.2 · p. 3

2.3 Grounding, verification, and selective prediction

  • Unlike retrieval or post-hoc verification, CoSQ assesses support before commitment and can be used before retrieval, after retrieval, or alongside verification.The accompanying framework figure depicts required-information identification, support assessment, a confidence gate, and answer or abstention.Source: §2.3 · p. 3, Figure 1 · p. 4
  • The paper adapts reject-option risk–coverage evaluation to LLM prompting and focuses on selective factual answering rather than assuming transfer from unrelated clustering or streaming tasks.Answered accuracy is conditional on commitment, whereas coverage measures how often the system commits.Source: §2.3 · p. 4

3 Chain-of-Self-Questioning

CoSQ prompts a model to identify needed information, assess its support, and either answer from accepted units or abstain. Its variants differ in which confidence checks govern commitment and in how accepted information is used.

3 Chain-of-Self-Questioning

  • CoSQ constructs question-specific information units, scores their support, and applies a decision rule before generating an answer from accepted units.In the grounded variant, rejected premises are not passed to the answer stage.Source: §3 · p. 4

3.1 Grounded-CoSQ

  • Grounded-CoSQ commits when mean item-level confidence meets τ and otherwise abstains; the accepted information units then condition the final answer.The study evaluates τ ∈ {0.50, 0.60, 0.70, 0.80, 0.90} before selecting Grounded-CoSQ at τ = 0.90 as its primary conservative operating point.Source: §3.1 · p. 4

3.2 Critical-CoSQ

  • Critical-CoSQ gates on units labeled critical, reducing the mechanical effect of the number of supporting details.The variant tests whether focusing on information necessary for a correct answer improves selective answering; τ = 0.90 is its principal coverage-oriented comparison.Source: §3.2 · p. 4

3.3 Adaptive-CoSQ

  • Adaptive-CoSQ accepts only when overall mean confidence meets τ, critical-claim mean is at least 0.65, and minimum critical-claim confidence is at least 0.40.After acceptance, claims at or above 0.40 pass to answer generation; the answer mode and contradiction check can still produce abstention.Source: §3.3 · p. 5

3.4 Baselines

  • Direct prompting answers without an explicit reasoning or abstention stage, while CoT reasons before answering; neither baseline receives task-specific fine-tuning or model logits.These prompt-level baselines measure the cost of adding a selective gate.Source: §3.4 · p. 5

4 Experimental Design

The experiments use a balanced-option TruthfulQA-MC protocol across eleven model families, with selective-answering metrics and paired comparisons across models. The design also reports CoSQ’s inference overhead.

4.1 Dataset and scoring protocol

  • The primary benchmark is TruthfulQA’s 817-question multiple-choice validation split, with option order balanced deterministically and abstentions excluded from answered accuracy.Direct and CoT are forced-choice baselines with full coverage; malformed responses count as wrong.Source: §4.1 · p. 5

4.2 Model panel

  • The model panel contains eleven contemporary instruction-tuned or hosted model families, although hosted endpoint labels may not identify immutable weight revisions.The paper treats that endpoint-label uncertainty as a reproducibility limitation, not evidence of a particular model version.Source: §4.2 · p. 5

4.3 Conditions and reproducibility

  • Each of eleven models is evaluated under seventeen conditions using the same questions, balanced option order, generation settings, and prompt versions.Responses and intermediate decisions are cached for re-analysis, and the framework, utilities, and experiment prompts are made available.Source: §4.3 · pp. 5–6

4.4 Metrics

  • Answered accuracy measures correctness among committed answers and must be read with coverage; HR measures wrong commitments across all questions.For parseable outputs, HR = Coverage × (1 − AA); it is not a complete linguistic annotation of hallucination.Source: §4.4 · p. 6

4.5 Statistical analysis

  • The primary comparisons use paired model-level differences, exact one-sided Wilcoxon signed-rank tests, 10,000-resample bootstrap intervals, and Cohen’s dz.Because the model count is small and endpoints are not independent training examples, inferential results are interpreted alongside per-model values and complete threshold curves.Source: §4.5 · pp. 6–7

4.6 Inference cost

  • CoSQ adds model calls because information extraction and confidence assessment precede answer generation.Table 2 summarizes relative inference stages per question, with a stage defined as one prompt–completion interaction.Source: §4.6 · p. 7, Table 2 · p. 7

5 Results

Across the eleven-model TruthfulQA-MC panel, Grounded-CoSQ at τ=0.90 lowers wrong commitments and raises answered accuracy versus CoT while abstaining on some questions. The sweep, secondary NQ-Short evaluation, and paired analyses characterize operating trade-offs and consistency.

5.1 Main model-panel results

  • Grounded-CoSQ at τ=0.90 yields 8.9% HR and 89.7% AA, versus CoT’s 13.1% wrong-commitment rate and 86.9% AA, while answering 87.6% of questions.CoT and Direct have 100% coverage under the forced-choice protocol; this reflects the evaluation contract, not perfect instruction compliance.Source: §5.1 · pp. 7–8

5.2 Abstention quality and lost value

  • 31.3% of Grounded-CoSQ’s referrals coincide with a CoT error; the remaining referrals withhold answers CoT happens to answer correctly.The referral shares describe abstention composition, averaged across eleven models, rather than a separate accuracy definition.Source: §5.2 · p. 8, Table 4 · p. 8

5.3 Threshold sweep

  • At every evaluated threshold, all three CoSQ variants have lower mean HR and higher mean AA than CoT; increasing τ generally lowers coverage and HR.The largest AA gains occur at τ=0.90, making τ an explicit deployment control whose choice depends on the consequences of wrong commitments versus referrals.Source: §5.3 · p. 9, Figure 3 · p. 9

5.4 Cross-dataset generalization: NQ-Short

  • Grounded-CoSQ on NQ-Short lowers mean HR from 0.412 under CoT to 0.274, with 0.830 coverage and AA rising from 0.588 to 0.670.This five-model open-form evaluation provides convergent evidence, not a numerically comparable or second confirmatory estimate.Source: Table 6 · p. 10, §5.4 · p. 10, Figure 4 · p. 10

5.5 Model-level consistency

  • At τ=0.90, all three variants have lower HR and higher AA than CoT for every model; Grounded-CoSQ HR reductions range from 1.84 to 9.79 percentage points.Table 7 reports the model-level values; DeepSeek Flash has the largest absolute Grounded-CoSQ reduction, from 0.177 to 0.080.Source: Table 7 · p. 11, §5.5 · p. 11

5.6 Paired statistical analysis

  • For Grounded-CoSQ, mean HR decreases by 0.0422 and AA increases by 0.0287; both directional Wilcoxon tests yield p=0.00049, with all eleven model differences in the hypothesized direction.The paired directions are consistent within this panel, but the hosted endpoints are not independent samples from a population of all LLMs.Source: Table 8 · p. 11, §5.6 · p. 11

6 Discussion

CoSQ’s selective policies reduce wrong commitments while improving answered accuracy, with distinct coverage–risk profiles. The discussion frames threshold choice around application costs and notes the limits of multiple-choice evaluation.

6 Discussion

  • Grounded-CoSQ at τ = 0.90 reduces mean HR from 0.131 to 0.089 and increases AA from 0.869 to 0.897, with 0.876 coverage.All three variants improved AA and reduced unconditional wrong commitment for every model; Grounded-CoSQ’s abstentions include both prevented CoT errors and withheld correct CoT answers.Source: §6 · p. 14
  • The full threshold sweep shows all 15 selective configurations had lower model-macro HR than CoT, but does not establish one universally optimal threshold.The operating point should reflect the relative costs of wrong commitments and abstentions; the comparison with the earlier open-form pilot is provisional because the protocols differ and were not pooled.Source: §6 · p. 14
  • Model size or recency does not automatically yield substantially higher CoSQ coverage; model families differ in how self-assessed scores map to the gate.The authors argue selective answering should remain an explicit system component even for strong models.Source: §6 · p. 14
  • Balanced, forced-choice TruthfulQA-MC removes fixed-label and accidental-abstention explanations, but does not represent every property of open-ended factual dialogue.NQ-Short is therefore treated as secondary generalization evidence rather than included in the confirmatory estimate.Source: §6 · p. 14
  • In high-stakes domains, CoSQ is an escalation aid rather than a substitute for professional review, and thresholds should reflect domain-specific costs.The relevant costs include false answers, abstentions, and missed opportunities.Source: §6 · p. 14

7 Limitations

The study’s findings are bounded by self-generated confidence, hosted-model reproducibility, inference cost, benchmark scope, and prompt robustness.

7 Limitations

  • CoSQ elicits confidence from the answering model itself, so the signal is not independent and may be miscalibrated.Threshold values are empirical operating points, not literal probabilities.Source: §7 · pp. 14–15
  • Hosted model endpoints may change weights or serving behavior without exposing immutable revisions, making endpoint identifiers insufficient to establish a fixed model version.Run manifests, prompt versions, and aggregate tables are essential to interpreting the results.Source: §7 · p. 15
  • CoSQ adds inference calls and token cost by decomposing questions and assessing information units before answering.The authors consider this cost justified only when the expected cost of a wrong answer is sufficiently high, and call for comparison under an explicit cost budget.Source: §7 · p. 15
  • TruthfulQA-MC is the primary benchmark and NQ-Short is a smaller secondary evaluation, so larger multi-domain evaluations are needed before claiming broad deployment validity.Threshold transfer across datasets and domains remains unresolved.Source: §7 · p. 15
  • A common prompt protocol improves comparability but may leave model-specific performance untapped, while prompt robustness and human assessment remain open issues.The study does not exhaustively test paraphrases, instruction order, or alternative confidence wording.Source: §7 · p. 15

8 Conclusion

The paper concludes that prompt-only CoSQ supports selective factual answering across the evaluated models and benchmark conditions. Its variants offer different risk–coverage operating choices, not universal superiority.

8 Conclusion

  • Across eleven models and 817 balanced-option TruthfulQA-MC questions, Grounded-CoSQ at τ = 0.90 reduced HR by 4.22 points and increased AA by 2.87 points while retaining 87.6% coverage.Both directional effects held for every model in the panel and were supported by paired exact Wilcoxon tests.Source: §8 · p. 15
  • Critical-CoSQ retained broader coverage, while Adaptive-CoSQ achieved the lowest mean HR; all three variants improved AA and reduced wrong commitments for every model.The complete threshold sweep presents these configurations as choices on a risk–coverage frontier rather than claims of universal superiority.Source: §8 · p. 15
  • Reliable LLM evaluation should consider both answered accuracy and how often the model answers, with threshold choice reflecting the application’s trade-offs.Grounded-CoSQ provides the strongest balance in these experiments; Critical-CoSQ offers broader coverage and Adaptive-CoSQ the lowest mean wrong-commitment rate.Source: §8 · p. 15
  • The study’s experimental framework, configuration examples, aggregate tables, and figure-generation scripts are publicly available, alongside versioned source files and aggregate results.Provider credentials, private endpoints, local caches, and operational configuration files are excluded from the public release.Source: §8 · p. 15
  • CoSQ is not a guarantee of factual correctness and should not be deployed in high-stakes settings without domain-specific validation, evidence access, and human oversight.Abstention thresholds should reflect the costs of wrong answers and unanswered questions in the target application.Source: §8 · p. 16

A Prompt Templates

The appendix reproduces the prompt templates used in the experiments, from question wrappers and baselines through information assessment and answer generation.

A.1 Question wrappers

  • The question wrappers distinguish multiple-choice and open-form inputs.The appendix labels these templates as question.v1.txt and question_open.v1.txt.Source: §A.1 · p. 20

A.2 Multiple-choice baselines

  • The multiple-choice Direct baseline requires exactly one option token and forbids an explanation.The CoT baseline also returns one option token, with silent reasoning and no explanation.Source: §A.2 · p. 20

A.3 Open-form baselines

  • The open-form Direct baseline requests a one-short-sentence answer, while open-form CoT requests step-by-step thinking and a short final answer.The final answer must begin with “Answer:”.Source: §A.3 · p. 20

A.4 Grounded-CoSQ information-unit stage

  • Grounded-CoSQ and Adaptive-CoSQ share a stage that lists the individual information pieces needed to answer correctly, without answering yet.The template asks for one numbered information need per line.Source: §A.4 · p. 21

A.5 Grounded-CoSQ claim and confidence stage

  • For each identified information need, the confidence template elicits a factual claim and an integer confidence score from 0 to 100.The prompt instructs the model not to answer the main question yet.Source: §A.5 · p. 21

A.6 Grounded-CoSQ answer stage

  • The multiple-choice answer stage supplies claims that passed the confidence gate and permits either an option token or ABSTAIN.The open-form stage supplies accepted claims as evidence and requests brief reasoning followed by a short answer.Source: §A.6 · p. 21

A.7 Critical-CoSQ information classification

  • Critical-CoSQ classifies each needed information item as critical or supporting before answering.Critical means no correct answer is possible without the information; supporting means it helps but is not necessary to answer.Source: §A.7 · p. 22

A.8 Critical-CoSQ claim and confidence stage

Critical-CoSQ elicits confidence in claims for each critical information need before answering from claims that pass the gate. Adaptive-CoSQ adds role assessment and a post-answer contradiction check, with shared abstention when a gate rejects or a contradiction is found.

A.8 Critical-CoSQ claim and confidence stage

  • Critical-CoSQ separately elicits a factual claim and 0–100 confidence for each critical information need before the model answers the main question.The prompt requires exactly one concise factual claim and a confidence score.Source: §A.8 · p. 22

A.9 Critical-CoSQ answer stage

  • Critical-CoSQ answers from claims that passed the confidence gate, returning one option token or ABSTAIN for multiple choice, and a short claim-grounded sentence for open-form answers.Open-form answers may use general reasoning but must not contradict the accepted claims.Source: §A.9 · p. 22

A.10 Adaptive-CoSQ role-aware assessment

  • Adaptive-CoSQ separately elicits a factual claim, critical-or-supporting role, and 0–100 confidence for each information unit from the shared identification stage.The prompt requests this assessment before answering the main question.Source: §A.10 · p. 23

A.11 Adaptive-CoSQ answer stage

  • Adaptive-CoSQ multiple-choice answers use claims passing adaptive confidence checks and return one option token or ABSTAIN; open-form answers use accepted claims and specify a cautious response mode.The open-form prompt allows general reasoning, bars contradiction of accepted claims, and asks cautious mode to signal uncertainty without refusing when evidence supports a useful answer.Source: §A.11 · p. 23

A.12 Adaptive-CoSQ consistency check

  • Adaptive-CoSQ checks whether the candidate answer contradicts any accepted factual claim, returning CONSISTENT or CONTRADICTORY.The check receives both the candidate answer and accepted claims.Source: §A.12 · p. 23

A.13 Shared abstention output

  • A rejected gate or detected contradiction records an abstention without another answer-generation call.The shared abstention output is used in either case.Source: §A.13 · p. 24

B Supplementary NQ model-level results

Table 9 reports model-level NQ-Short results at τ = 0.90 for five representative models. Its cells give answered accuracy, coverage, and unconditional wrong-commitment rate; the primary claims do not use an NQ threshold sweep.

B Supplementary NQ model-level results

  • Table 9 gives model-level NQ-Short results for five representative models at τ = 0.90, without an NQ threshold sweep for the primary claims.The table supplies the model-level values underlying the aggregate NQ-Short results in Table 6.Source: B Supplementary NQ model-level results · p. 25, Table 9 · p. 25
  • Table 9 provides the model-level values underlying the aggregate NQ-Short results in Table 6.No cell values are stated in the supplied passages.Source: B Supplementary NQ model-level results · p. 25
  • Each Table 9 cell reports answered accuracy, coverage, and unconditional wrong-commitment rate.The table caption abbreviates these metrics as AA / coverage / HR.Source: Table 9 · p. 25

Related papers

Loading 2609.17516v1…