Strong yet Controllable LLM Abstention without Learning to Abstain

Junwei Zhang; Yifei Wang; Wanli Yang; Hexiang Tan; Rui Tang; Xin Chen; Jingang Wang; Du Su

cs.CLOct 12, 2026

Strong yet Controllable LLM Abstention without Learning to Abstain

Authors: Junwei Zhang, Yifei Wang, Wanli Yang, Hexiang Tan, Rui Tang, Xin Chen, Jingang Wang, Du Su, +1 more

Organizations: State Key Laboratory of AI Safety, Institute of Computing Technology, CAS · University of Chinese Academy of Sciences · Beijing Institute of Technology

Abstract

Large language models (LLMs) can reduce hallucination by abstaining from questions they would likely get wrong. Abstention can be implemented outside the model (selective answering) using a confidence threshold that controls the answer rate (coverage), or learned inside it (abstention training), which can perform better but fixes coverage once trained. To compare the two, we match their coverage and decompose risk (error rate) into answering ability and distinguishing ability, i.e., how well the model answers and how accurately it abstains, respectively. The decomposition shows that abstention training performs better mainly by answering better: with strong confidence estimators, the two have nearly equal distinguishing ability. This motivates Train for Accuracy, then Post-hoc Select (TAPS), a modular approach that keeps coverage controllable, needs only standard post-training and an off-the-shelf confidence estimator, and benefits from advances in either component. Across three LLMs and three factual QA benchmarks, TAPS with a probing-based confidence estimator attains 1.5--9.4 percentage points lower mean risk than three representative abstention training methods at matched coverage. Within TAPS, accuracy training often weakens the model's own confidence signals, yet refitting the estimator largely preserves distinguishing ability. Together, these results establish TAPS as a strong yet controllable baseline and raise the bar for future abstention training: sacrificing control over coverage should yield gains beyond what can be achieved by optimizing answering and abstaining separately.

Figures & tables

Figure 1: Behavioral decomposition of abstention performance at matched coverage. (a) Illustration of outcome transitions under abstention. (b) On Qwen2.5-7B-Instruct/TriviaQA at matched coverage

c=72.7\%

, stronger selective answering (S.E.) closes the distinguishing-ability gap with TruthRL (

G_{\mathrm{dis}}

), but TruthRL’s edge in answering ability (

G_{\mathrm{ans}}

) still yields lower risk.

Figure 2: Comparison of abstention approaches. (a) Abstention training jointly optimizes answering and abstention decisions inside the model. (b) Selective answering externalizes abstention decision after generation using a confidence estimator and an adjustable threshold

\tau

. (c) TAPS trains the answering model solely for accuracy while retaining the same post-hoc selection procedure.

Figure 3: Risk–coverage comparison with abstention training. Each curve is obtained by varying the acceptance threshold

\tau

for a fixed answering checkpoint and confidence estimator. Results for the other two models are reported in Appendix F .

Figure 4: Effects of accuracy training on answering ability (Acc) and distinguishing ability (AUROC). Values show AccRL minus PreRL in percentage points (pp).

Figure 5: AURC Changes from PreRL to AccRL.

Figure 6: Relationship between AUROC and AURC.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Effect of sweeping the training-time abstention reward. Each point represents a separately trained TruthRL policy based on Qwen2.5-7B-Instruct across NQ, PopQA, and TriviaQA. A common pattern emerges: moving

r_{\mathrm{abs}}

from

-1

-0.9

causes an abrupt drop in coverage, while subsequent changes produce only limited or non-monotonic coverage variation.

Setting	Qwen2.5-7B-Instruct / Llama-3.1-8B-Instruct	Qwen3-8B
Training budget	8 epochs	8 epochs
Questions per rollout batch	128	128
Responses per question	5	5
Rollout temperature / top- $p$	1.0 / 1.0	1.0 / 1.0
Rollout top- $k$ truncation	Disabled	Disabled
Maximum prompt length	512 tokens	512 tokens

Appendix

Table 3: RL training hyperparameters.

			TruthRL		Selective Answering
Model	Dataset	Cov.			Vanilla		P(True)		S.E.
			$G_{\mathrm{ans}}$	$G_{\mathrm{dis}}$	$G_{\mathrm{ans}}$	$G_{\mathrm{dis}}$	$G_{\mathrm{ans}}$	$G_{\mathrm{dis}}$	$G_{\mathrm{ans}}$	$G_{\mathrm{dis}}$
Qwen2.5-7B-Instruct	NQ	50.60	6.46	87.36	0.00	67.93	0.00	82.63	0.00	88.07
	PopQA	19.69	5.30	91.75	0.00	78.91	0.00	88.43	0.00	92.03
	TriviaQA	72.70	6.13	86.23	0.00	43.75	0.00	75.42	0.00	86.83
Qwen3-8B	NQ	50.45	4.74	84.89	0.00	65.34	0.00	78.96	0.25	85.59

Appendix

Table 5: Answering and distinguishing ability at matched coverage for every model–dataset setting. All entries are percentages. Selective answering matches TruthRL’s coverage within each setting. The highest value for each metric within each row is in bold .

Figure 8: Extended risk–coverage comparisons. Results on Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct. TAPS curves vary the acceptance threshold. Markers denote the native operating points of abstention-trained checkpoints.

Figure 9: Extended confidence-estimation results. AUROC–AURC relationships on Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct. Each line connects confidence estimators for one dataset in increasing AUROC order.

Figure 10: Post-hoc selection on TruthRL. The results show Qwen3-8B across three datasets. Because the probe can only filter TruthRL’s native answers, its curve ends at the native coverage (star). Its relative advantage over TAPS varies across datasets and coverage levels.

Explore similar work

Answering the Wrong Question: Reasoning Trace Inversion for Abstention in LLMs
Apr 2, 2026Abinitha Gourabathina, Inkit Padhi, Manish Nagireddy +2LLM Abstention LLM Hallucination Mitigation
MIT · IBM Research · MIT-IBM Watson AI Lab
I-CALM: Incentivizing Confidence-Aware Abstention for LLM Selective Answering
Apr 5, 2026Haotian Zong, Binze Li, Yufei Long +3Selective Prediction LLM Hallucination Mitigation
Department of Applied Mathematics and Statistics, Johns Hopkins University, Baltimore, MD 21218, USA · Department of Computer Science, Johns Hopkins University, Baltimore, MD 21218, USA · School of Government and Policy, Johns Hopkins University, Washington, D.C. 20001, USA +1
Uncertainty-Aware Abstention in Large Language Models with Provable Alignment Guarantees
Jul 5, 2026Sijin Dong, Hiroyuki ShinnouSelective Prediction Confidence Estimation in Language Models
1Ibaraki University, Japan.

Apr 2, 2026J/K move · Enter open · S save

Answering the Wrong Question: Reasoning Trace Inversion for Abstention in LLMs

Abinitha Gourabathina, Inkit Padhi, Manish Nagireddy, Subhajit Chaudhury+1

MIT · IBM Research · MIT-IBM Watson AI Lab

LLM Abstention LLM Hallucination Mitigation LLM Reasoning

Open paper AskPDF arXiv

For Large Language Models (LLMs) to be reliably deployed, models must effectively know when not to answer: abstain. Reasoning models, in particular, have gained attention for impressive performance on complex tasks. However, reasoning models have been shown to have worse abstention abilities. Taking the vulnerabilities of reasoning models into account, we propose our Query Misalignment Framework. Hallucinations resulting in failed abstention can be reinterpreted as LLMs answering the wrong question (rather than answering a question incorrectly). Based on this framework, we develop a new class of state-of-the-art abstention methods called Trace Inversion. First, we generate the reasoning trace of a model. Based on only the trace, we then reconstruct the most likely query that the model responded to. Finally, we compare the initial query with the reconstructed query. Low similarity score between the initial query and reconstructed query suggests that the model likely answered the question incorrectly and is flagged to abstain. Extensive experiments demonstrate that Trace Inversion effectively boosts abstention performance in four frontier LLMs across nine abstention QA datasets, beating competitive baselines in 33 out of 36 settings.

Abstract

Explore similar work

Answering the Wrong Question: Reasoning Trace Inversion for Abstention in LLMs

I-CALM: Incentivizing Confidence-Aware Abstention for LLM Selective Answering

Uncertainty-Aware Abstention in Large Language Models with Provable Alignment Guarantees