LLM Abstention Can Be a Prompt Artifact, in Addition to Genuine Uncertainty

Zipeng Ling; Shuliang Liu; Yuehao Tang; Junqi Yang; Shenghong Fu; Seonil Son; Chen Huang; Kejia Huang

cs.CLJul 22, 2025

LLM Abstention Can Be a Prompt Artifact, in Addition to Genuine Uncertainty

Authors: Zipeng Ling, Shuliang Liu, Yuehao Tang, Junqi Yang, Shenghong Fu, Seonil Son, Chen Huang, Kejia Huang,

Organizations: The Hong Kong University of Science and Technology (Guangzhou) · University of Alberta · Alberta Machine Intelligence Institute · Huazhong University of Science and Technology ·

Abstract

Large Language Models (LLMs) are increasingly trained to abstain from answering questions they are unsure about. However, this ability is often misapplied: in real-world applications, user prompts sometimes contain elements of uncertainty, which lead LLMs to abstain even on problems they are capable of solving. We argue that LLM abstention is not only an expression of genuine uncertainty; it can also be an artifact largely shaped by prompts. We name this phenomenon Abstention Inflation. We add "Unknown" as an extra option for LLMs to choose from; experiments show serious accuracy drops on True/False Questions (TFQs). Replacing "Unknown" with an unrelated random word produces a similar effect. We argue that LLMs are trained to imitate the surface pattern of abstention, rather than to express genuine uncertainty. Based on ten experimental settings, we support four claims that form a progressive argument: (C1) Abstention Inflation can be triggered by the presence of an extra option, not by genuine uncertainty; (C2) it makes the models deny they can answer, even when they can; (C3) it is a later-layer output override, as the reasoning traces and mid-layer representations preserve correct answers; (C4) it is not stochastic noise: it results from various factors, emerges through instruction tuning, is boosted by problems' higher difficulty, and can be mitigated at larger model sizes.

Figures & tables

Figure 1: Problem background: adding an “Unknown” option makes LLMs abstain heavily, and replacing it with a random word makes LLMs treat the word as a placeholder for “Unknown” and choose it. A detailed explanation and examples are in Section 4.1.4 .

Figure 2: Representative experiments for the four claims about Abstention Inflation , traced from prompt to training: extra-option trigger (C1), introspective gap (C2), later-layer output override (C3), various factors: instruction tuning, difficulty and scale (C4).

Model	Metric	FLD	FLD MCQ	FOLIO	FOLIO MCQ	ARC	MedQA	MMLU	LogiQA
DeepSeek-V4-Flash	Acc (S1)	78.4	78.4	87.0	87.0	98.4	94.2	89.2	73.4
	Acc (S2)	53.0 $\blacktriangledown$ 25.4	51.8 $\blacktriangledown$ 26.6	75.4 $\blacktriangledown$ 11.6	76.2 $\blacktriangledown$ 10.8	98.8 $\blacktriangle$ 0.4	92.6 $\blacktriangledown$ 1.6	88.8 $\blacktriangledown$ 0.4	69.2 $\blacktriangledown$ 4.2
	Abs Rate (S2)	42.4	43.2	19.4	18.2	0.2	0.6	1.4	3.6
GPT-5.4-nano	Acc (S1)	61.9	61.9	84.2	84.2	93.8	82.8	76.1	50.6
	Acc (S2)	23.2 $\blacktriangledown$ 38.7	27.4 $\blacktriangledown$ 34.5	72.0 $\blacktriangledown$ 12.2	71.6 $\blacktriangledown$ 12.6	94.0 $\blacktriangle$ 0.2	79.9 $\blacktriangledown$ 2.9	75.5 $\blacktriangledown$ 0.6	50.2 $\blacktriangledown$ 0.4
	Abs Rate (S2)	69.4	63.1	22.2	21.8	1.0	2.5	2.4	5.6

Table 1: Main results. S1 Acc : baseline (no “Unknown” Option). S2 Acc : Accuracy with “Unknown” in prompts; subscript =

\Delta

from S1 (

\blacktriangledown

decline /

\blacktriangle

gain, %).

Figure 3: S4 Word Content Ablation on both TFQ datasets (top: FLD, bottom: FOLIO). Left: Replacement with synonyms of “Unknown” yields a similar effect. Right: Random unrelated words also yield a similar effect to “Unknown”.

Figure 4: Left: S5, w/o “Unknown” Option Rerun; all accuracies are above the 50% random baseline. Right: S6, Self-Diagnosis; models attribute 94.6–100% of their abstentions to the samples being objectively unanswerable, despite being able to answer correctly.

Figure 5: S7: Reasoning Traces Evaluation on traces generated with and without the “Unknown” option in the prompt. Left: the rate at which the NLI probe fails to recover the gold label from the trace. Right: adding the “Unknown” option causes a large accuracy drop.

Figure 6: S8: Logit-Lens Representation Probe. Left: the logit change

\Delta\log P(\text{``Unknown''})

across all 33 layers for prompts with and without the “Unknown” option. Right: Comparison of

\log P(\text{``Unknown''})

between Abstention Inflation and Wrong Prediction .

Figure 7: S9 Stability of Abstention Inflation . Left: Persistence Across Repeats. Right: Perception of Unknown-labeled Samples.

Figure 8: S10 Factor Analysis. Left: FLD samples with more steps lead to a higher Abs Rate . Right: higher temperature raises Abs Rate only slightly.

Figure 9: S10 Factor Analysis. (A) Abs Rate by Model Size & Alignment. (B) Abs Rate heatmap. (C) Accuracy and Abs Rate bar chart.

Figure 10: S10 Factor Analysis: Positional Biases. Error bars indicate 95% Wilson confidence intervals. Unknown (First), Unknown (Second), and Unknown (Third) place the “Unknown” option in the first, second, and third positions, respectively.

LLM Abstention

Large language model abstention studies when AI systems should answer, ask for more information, defer, or withhold unsupported claims.

See all 101 papers

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
Strong yet Controllable LLM Abstention without Learning to Abstain
Oct 12, 2026Junwei Zhang, Yifei Wang, Wanli Yang +6Factual QA LLM Abstention
State Key Laboratory of AI Safety, Institute of Computing Technology, CAS · University of Chinese Academy of Sciences · Beijing Institute of Technology
Beyond "I Don't Know": Evaluating LLM Self-Awareness in Discriminating Data and Model Uncertainty
Apr 19, 2026Jingyi Ren, Ante Wang, Yunghwei Lai +5LLM Evaluation Selective Prediction
Dept. of Comp. Sci. & Tech., Institute for AI, Tsinghua University, Beijing, China · Institute for AI Industry Research (AIR), Tsinghua University, Beijing, China

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

LLM Abstention

Explore similar work

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

Strong yet Controllable LLM Abstention without Learning to Abstain

Beyond "I Don't Know": Evaluating LLM Self-Awareness in Discriminating Data and Model Uncertainty