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

Abinitha Gourabathina; Inkit Padhi; Manish Nagireddy; Subhajit Chaudhury; Prasanna Sattigeri

cs.AIApr 2, 2026

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

Authors: Abinitha Gourabathina, Inkit Padhi, Manish Nagireddy, Subhajit Chaudhury, Prasanna Sattigeri

Organizations: MIT · IBM Research · MIT-IBM Watson AI Lab

Abstract

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.

Figures & tables

Figure 1: Prior approaches often underperform abstention in reasoning models.

Figure 2: Examples of how distinguishing between a user query $q$ compared to model query $q^{*}$ can reveal hallucination patterns. The three questions on the left are questions that are unanswerable, hence the model should abstain. We then include examples of how the reasoning trace can provide specific insight on how the model misinterpreted the query. Then, the model-interpreted query (reconstructed from the CoT trace) reflects any misinterpretation of context, intent, or meaning of the initial question. Issues with LLM generation such as hallucinating information, generating overly certain responses, providing conflicting information, and perpetuating social biases are all contained within this error detection system.

Figure 3: Overview of our three step approach. We provide an example of how our method particularly detects subtle hallucinations in a reasoning trace by comparing the user query $q$ with the model-interpreted query $q^{*}$ .

Method	Math & Knowledge				Comprehension				Biases & Safety				Overall
Method	MMLU	GSM	UMWP	Overall	KC	HS	Qu	Overall	Mis	Prop	BBQ	Overall	Overall
phi-4
Probs	.477	.509	.488	.491	.451	.666	.303	.473	.512	.624	.332	.489	.484
AskCali	.471	.504	.506	.494	.550	.612	.307	.490	.552	.618	.299	.490	.491
Reflect	.379	.541	.438	.453	.633	.771	.405	.603	.455	.515	.411	.460	.506
Cooperate	.424	.685	.420	.510	.504	.718	.414	.545	.369	.598	.422	.463	.506

Table 1: Abstain Accuracy of abstain strategies on nine datasets and four LLMs. Each number reported is the average of three seeds. We group the nine datasets (MMLU, GSM, UMWP, K-Crosswords, Hellaswag, Quail, Misconceptions, Propaganda, and BBQ) according to domain and provide aggregate results per domain and across all datasets. Best results in bold and second best in underline . Approaches are color-coded per category: calibration , prompting , collaboration , and our method . Trace Inversion achieves the best performance in 33 out of 36 settings.

Method	Math & Knowledge	Comprehension	Biases & Safety
Probs	0.1449	0.1829	0.1460
AskCali	0.1288	0.2099	0.1961
Reflect	0.1040	0.2359	0.1629
Cooperate	0.1796	0.3199	0.0975
Compete	0.1178	0.0746	0.1964
Trace Inversion	0.0347	0.0517	0.0681

Table 2: Performance gap between datasets with only answerable queries and datasets also containing unanswerable queries by domain and method. The gap is calculated as the difference between the average performance on answerable datasets and the average performance on UMWP for Math & Knowledge, Quail for Comprehension, and BBQ for Biases & Safety. For each method, scores are averaged across all four LLMs. Positive values indicate that answerable datasets are easier than those containing unanswerable queries. The bottom row shows the average gap across all methods besides Trace Inversion.

Method	Math & Knowledge				Comprehension				Biases & Safety				Overall
Method	MMLU	GSM	UMWP	Overall	KC	HS	Qu	Overall	Mis	Prop	BBQ	Overall	Overall
Averaged Across 4 LLMs
SE	.848	.841	.839	.842	.533	.721	.605	.620	.555	.617	.616	.596	.686
TrInv-LLM	.749	.785	.720	.751	.704	.788	.707	.733	.569	.598	.624	.597	.694
GROUND	.684	.747	.556	.662	.686	.780	.590	.685	.667	.800	.790	.752	.704
Trace Inversion	.808	.829	.784	.807	.663	.772	.666	.700	.686	.740	.645	.690	.732

Table 3: Ablation of Trace Inversion by looking at the Abstain Accuracy performance of the method with individual distance metrics. Performance is averaged across all four LLMs to show overall patterns. Best results in bold and second best in underline . Model specific results are reported in Appendix C.2 .

Figure 4: Abstain Accuracy of each domain and method averaged across four LLMs.

Method	Variant	Math & Knowledge				Comprehension				Biases & Safety				Overall
Method	Variant	MMLU	GSM	UMWP	Overall	KC	HS	Qu	Overall	Mis	Prop	BBQ	Overall	Overall
Averaged Across 4 LLMs
Probs	Regular	.689	.690	.544	.641	.512	.618	.382	.504	.521	.633	.431	.529	.558
Probs	+ CoT prompt	.673	.645	.504	.607	.474	.618	.344	.479	.531	.593	.417	.514	.533
AskCali	Regular	.668	.694	.552	.638	.533	.565	.339	.479	.526	.632	.383	.513	.543
AskCali	+ CoT prompt	.659	.648	.544	.617	.520	.530	.310	.453	.501	.587	.378	.489	.520

Table 4: Ablation of baseline methods with regular prompting (just the query from the dataset) versus additionally appending a CoT prompt. Abstain Accuracy of abstain strategies on the nine datasets, averaged across four LLMs. Lower of the two values (comparing between Regular and + CoT prompt ) are in red . Model specific results are reported in Appendix C.3 .

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Prompt
AskCali	Provide your best guess for the following question. Give only the guess.
	Example: {most likely guess}.
	Question: {question}.
	[ LLM-generated answer ].
	Provide the probability your guess is correct.
	Probability: [ LLM-generated probability in $[0,1]$ ].

Appendix

Table 5: Exact LLM prompts used for each method.

Method	Math & Knowledge				Comprehension				Biases & Safety				Overall
Method	MMLU	GSM	UMWP	Overall	KC	HS	Qu	Overall	Mis	Prop	BBQ	Overall	Overall
phi-4
Probs	.545	.584	.614	.581	.463	.634	.363	.487	.439	.636	.522	.533	.533
AskCali	.404	.585	.711	.567	.626	.588	.358	.524	.598	.538	.358	.498	.530
Reflect	.412	.466	.669	.516	.591	.757	.550	.633	.449	.422	.522	.464	.538
Cooperate	.428	.682	.634	.581	.532	.731	.656	.640	.272	.677	.601	.516	.579

Appendix

Table 6: Reliable Accuracy of abstain strategies on nine datasets and four LLMs. Each number reported is the average of three seeds. We group the nine datasets (MMLU, GSM, UMWP, K-Crosswords, Hellaswag, Quail, Misconceptions, Propaganda, and BBQ) according to domain and provide aggregate results per domain and across all datasets. Approaches are color-coded per category: calibration , prompting , collaboration , and our method .

Method	Math & Knowledge				Comprehension				Biases & Safety				Overall
Method	MMLU	GSM	UMWP	Overall	KC	HS	Qu	Overall	Mis	Prop	BBQ	Overall	Overall
phi-4
SE	.876	.798	.865	.846	.538	.727	.619	.628	.567	.608	.656	.610	.695
TrInv-LLM	.710	.743	.752	.735	.684	.774	.723	.727	.555	.556	.639	.584	.682
GROUND	.638	.721	.556	.638	.658	.760	.609	.675	.696	.826	.772	.765	.693
Trace Inversion	.712	.733	.757	.734	.663	.830	.694	.729	.649	.710	.614	.658	.707

Appendix

Table 7: Ablation of Trace Inversion by looking at the Abstain Accuracy performance of the method with individual distance metrics. Each number reported is the average of three seeds. Performance is shown for each of the four LLMs. Best results in bold and second best in underline .

Method	Variant	Math & Knowledge				Comprehension				Biases & Safety				Overall
Method	Variant	MMLU	GSM	UMWP	Overall	KC	HS	Qu	Overall	Mis	Prop	BBQ	Overall	Overall
phi-4
Probs	Regular	.477	.509	.488	.491	.451	.666	.303	.473	.512	.624	.332	.489	.485
Probs	+ CoT prompt	.449	.426	.395	.423	.424	.548	.286	.419	.404	.559	.236	.400	.414
AskCali	Regular	.471	.504	.506	.494	.550	.612	.307	.490	.552	.618	.299	.490	.491
AskCali	+ CoT prompt	.390	.394	.494	.426	.475	.565	.169	.403	.574	.613	.220	.469	.432

Appendix

Table 8: Abstain Accuracy of abstain strategies on nine datasets for each of the four LLMs. Each number reported is the average of three seeds. We group the nine datasets (MMLU, GSM, UMWP, K-Crosswords, Hellaswag, Quail, Misconceptions, Propaganda, and BBQ) according to domain and provide aggregate results per domain and across all datasets. Approaches are color-coded per category: calibration , prompting , collaboration , and our method .

Explore similar work

Abstain-R1: Calibrated Abstention and Post-Refusal Clarification via Verifiable RL
Apr 18, 2026Skylar Zhai, Jingcheng Liang, Dongyeop KangLLM Hallucination Mitigation RL for Language Model Reasoning
University of Minnesota
Strong yet Controllable LLM Abstention without Learning to Abstain
Oct 12, 2026Junwei Zhang, Yifei Wang, Wanli Yang +6LLM Abstention LLM Hallucination Mitigation
State Key Laboratory of AI Safety, Institute of Computing Technology, CAS · University of Chinese Academy of Sciences · Beijing Institute of Technology
Two Axes of LLM Abstention: Answer Correctness and Question Answerability
Jul 9, 2026Benedikt J. WagnerSelective Prediction Confidence Estimation in Language Models
City St George’s, University of London

Apr 18, 2026J/K move · Enter open · S save

Abstain-R1: Calibrated Abstention and Post-Refusal Clarification via Verifiable RL

Skylar Zhai, Jingcheng Liang, Dongyeop Kang

University of Minnesota

LLM Hallucination Mitigation RL for Language Model Reasoning LLM Reliability Reinforcement Learning with Verifiable Rewards Refusal Behavior in Language Models LLM Abstention

Open paper AskPDF arXiv

Reinforcement fine-tuning improves the reasoning ability of large language models, but it can also encourage them to answer unanswerable queries by guessing or hallucinating missing information. Existing abstention methods either train models to produce generic refusals or encourage follow-up clarifications without verifying whether those clarifications identify the key missing information. We study queries that are clear in meaning but cannot be reliably resolved from the given information, and argue that a reliable model should not only abstain, but also explain what is missing. We propose a clarification-aware RLVR reward that, while rewarding correct answers on answerable queries, jointly optimizes explicit abstention and semantically aligned post-refusal clarification on unanswerable queries. Using this reward, we train Abstain-R1, a 3B model that improves abstention and clarification on unanswerable queries while preserving strong performance on answerable ones. Experiments on Abstain-Test, Abstain-QA, and SelfAware show that Abstain-R1 substantially improves over its base model and achieves unanswerable-query behavior competitive with larger systems including DeepSeek-R1, suggesting that calibrated abstention and clarification can be learned through verifiable rewards rather than emerging from scale alone.

Abstract

Explore similar work

Abstain-R1: Calibrated Abstention and Post-Refusal Clarification via Verifiable RL

Strong yet Controllable LLM Abstention without Learning to Abstain

Two Axes of LLM Abstention: Answer Correctness and Question Answerability