--- title: Epistemic Abstention in AI Systems url: https://www.emergentmind.com/topics/epistemic-abstention type: topic --- # Epistemic Abstention in AI Systems Searching arXiv for recent papers on epistemic abstention and closely related formulations. Search query: epistemic abstention LLM abstention benchmark reject option deference AI site:arxiv.org Epistemic abstention is the withholding of an answer, judgment, endorsement, vote, or epistemic standing when the available evidence, validated scope, or reliability conditions do not justify categorical commitment. Recent research uses the term across several levels of analysis: as selective refusal in large language models and multimodal systems; as a reject-option in Bayesian prediction under limited data; as a defeasible response to artificial epistemic authority; and, in human–AI interaction, as a stance in which AI is used while being denied any epistemic contribution [2510.21043] [2511.04855] [2508.18760] [2508.03673]. Across these literatures, abstention is typically not treated as default skepticism. It is a calibrated response to insufficiency, conflict, ambiguity, or missing warrant. ## 1. Conceptual scope and major senses The contemporary literature does not use a single, uniform definition of epistemic abstention. In the survey literature on large language models, abstention is defined as the refusal to answer a query, and is analyzed from three perspectives: the query, the model, and human values. On this view, abstention is appropriate when the query is not answerable, when model confidence is too low, or when the response would violate human-value constraints. The same survey distinguishes **full abstention**, **partial abstention**, and non-abstention failure modes such as ignoring or reframing the question [2407.18418]. A distinct philosophical usage appears in work on deference to AI. There, abstention concerns whether a person should withhold full endorsement of an AI output when the system’s authority is not securely established. This treatment makes abstention a defeasible response to unreliable or insufficiently grounded AI authority rather than a blanket refusal to rely on AI. The relevant question is not whether AI should ever be trusted, but how much independent judgment a user should retain when confronted with AI output [2510.21043]. In human–computer interaction, the term names an epistemic relationship rather than a model behavior. The category **Epistemic Abstention** is defined as **“Uses AI but denies its epistemic contribution”**. It is characterized by the metaphor *Tool*, the trust type *Distrust*, the assessment perspective *Outcome-based*, the human epistemic status *Both*, and the tasks *Summarization, General search*. In this sense, abstention is not non-use. It is active refusal to credit AI with a knowledge-producing role [2508.03673]. These senses are related but not identical. One concerns when a system should not answer; another concerns when a human should not fully endorse a system’s answer; a third concerns refusing to recognize AI as a knower at all. This suggests that “epistemic abstention” is best understood as a family of practices organized around one common norm: do not convert inadequate grounds into authoritative commitment. ## 2. Formal decision structures A central formalization comes from reject-option prediction. In the classical setting, a predictor \(q:\mathcal{X}\to \mathcal{Y}\cup\{\text{reject}\}\) incurs the reject-option loss \[ \ell^\epsilon(y,d)= \begin{cases} \ell(y,d) & \text{if } d\in\mathcal{Y},\\ \epsilon & \text{if } d=\text{reject}, \end{cases} \] with rejection cost \(\epsilon\). The classical optimal rule abstains when conditional risk exceeds that cost. In the epistemic reject-option framework, however, the relevant trigger is not total predictive uncertainty alone, but the expected performance gap caused by limited data. The optimal predictor is \[ Q_E(x,D)= \begin{cases} H_B(x,D) & \text{if } E(x,D)\le \delta,\\ \text{reject} & \text{if } E(x,D)>\delta, \end{cases} \] where \(E(x,D)\) is conditional regret relative to the Bayes-optimal predictor with full knowledge of the true parameter. This makes abstention a response to data insufficiency rather than merely to aleatoric ambiguity [2511.04855]. Related selective-prediction formulations appear in scientific reasoning. There the predictor is a pair \((F,g)\), where \(F\) produces a label and \(g(x,E)\in\{0,1\}\) decides whether to answer. Selective risk \[ R(F,g)=\frac{\mathbb{E}[l(F(x,E),y)\,g(x,E)]}{\mathbb{E}[g(x,E)]} \] is evaluated together with coverage \[ \phi(F,g)=\mathbb{E}[g(x,E)]. \] The system abstains whenever confidence falls below a threshold \(T\). The core evaluation object is therefore the risk–coverage trade-off rather than unconditional accuracy alone [2602.14189]. Prompt-based factual QA work proposes another explicit threshold model. If the model believes an answer is correct with probability \(p\), answer utility is \((R+\beta)p-\beta\), abstain utility is \(\gamma\), and the Bayes-optimal rule is to answer iff \[ p \ge \tau := \dfrac{\gamma+\beta}{R+\beta}. \] Under this formulation, abstention is governed by an explicit reward scheme rather than by an implicit preference for coverage. The operational effect is a tunable abstention–hallucination frontier [2604.03904]. These formulations differ in what counts as the relevant insufficiency. In regret-based reject-option theory it is uncertainty due to limited data; in scientific verification it is insufficiency of evidence; in prompt-based factual QA it is insufficient expected utility under an answer-versus-abstain scheme. A plausible implication is that epistemic abstention is not one decision rule but a class of decision rules whose common structure is selective commitment under explicit costs. ## 3. Language models, unanswerability, and abstention failure Recent LLM work treats abstention as a first-class capability rather than a peripheral safeguard. **AbstentionBench** evaluates abstention across 20 diverse datasets involving answer unknown, underspecification, false premise, stale information, subjective questions, and related conditions. Its central result is that abstention is an unsolved problem and that scaling models is of little use. It also reports that reasoning fine-tuning degrades abstention by \(24\%\) on average, even on math and science domains on which reasoning models are explicitly trained [2506.09038]. A more focused study on large reasoning models examines unanswerable math problems and identifies three response types: **correct abstention**, **hallucinated answer**, and **cognitive fixation**. Across models, more than half of the unanswerable questions are not handled with correct abstention. The study further shows that answerability is linearly recoverable from hidden states, with probe accuracy often exceeding **0.8** by the end of the reasoning trajectory, indicating a mismatch between internal recognition and external response. Its proposed two-stage method combines cognitive monitoring with inference-time intervention and significantly improves abstention rate while maintaining overall reasoning performance [2508.18760]. On factual questions with verifiable ground truth, prompt-only interventions can also induce abstention. The I-CALM framework elicits verbal confidence, announces an explicit reward scheme for answer versus abstain, and adds concise norms emphasizing truthfulness, humility, and responsibility. In the main PopQA setting with GPT-5 mini, answered false-answer rate drops from **52.3%** under Pure Eval to **34.2%** under Scheme B + norms, while coverage falls from **96.5%** to **55.3%**. The paper’s interpretation is explicit: selective reliability improves, but the underlying forced-answer competence remains largely unchanged [2604.03904]. Training-time approaches pursue the same problem more directly. **Reinforced Hesitation** replaces binary reward with a ternary reward: \(+1\) for correct answers, \(0\) for “I don’t know,” and \(-\lambda\) for errors. The paper reports that frontier models almost never abstain despite explicit penalty warnings, and on MedQA records exactly zero abstentions across 11 models, 5 penalty conditions, and 1,273 questions. In controlled training, varying \(\lambda\) yields a Pareto frontier from aggressive answerers to conservative abstainers [2511.11500]. **Abstain-R1** extends this by rewarding not only explicit abstention on unanswerable queries but also semantically aligned post-refusal clarification. On Abstain-Test, the base Qwen2.5 3B Instruct model has **U-Ref = 9.4** and **U-Clar = 0.6**, whereas Abstain-R1 reaches **U-Ref = 68.1** and **U-Clar = 55.1** while also improving answerable accuracy [2604.17073]. A separate line of work argues that abstention should occur before answer commitment rather than after output inspection. **Aspect-Based Causal Abstention** distinguishes **Type-1** abstention, triggered by knowledge conflict when aspect effects are inconsistent, from **Type-2** abstention, triggered by knowledge insufficiency when aspect effects consistently support abstention. Its main claim is that hallucination is not only an output problem but an internal-knowledge-structure problem [2511.17170]. ## 4. Deference, oversight, and human epistemic roles In philosophical work on AI authority, epistemic abstention is tied to the rejection of **AI Preemptionism**. That position holds that the output \(O\) of a recognized artificial epistemic authority should function as a pre-emptive reason that replaces a user’s independent reasons. The alternative **Total Evidence AI Deference** instead treats \(O\) as a contributory reason integrated into the user’s total evidence. On this view, users should defer conditionally while remaining alert to defeaters. The paper’s practical principle, **Critical Deference with Oversight**, states that deference should be withheld or revisited in cases of domain mismatch, reliability-undermining evidence, conflicting authority, or novel evidence not considered by the AI [2510.21043]. This account explicitly links abstention to human oversight. The three claimed advantages of the total-evidence view are that it mitigates expertise atrophy, provides an epistemic basis for meaningful human control, and explains justified mistrust when reliability conditions fail. High-stakes settings such as medicine, finance, and aerospace are emphasized because opacity, self-reinforcing authority, and lack of failure markers make automatic endorsement especially risky [2510.21043]. Interview-based HCI work describes a different but related form of abstention at the level of user stance. **ER-5**, Epistemic Abstention, captures cases where users utilize AI tools but actively deny them any epistemic authority. The paper stresses that this is stronger than pragmatic non-authoritativeness: unlike **Instrumental Reliance**, which treats AI as useful for task efficiency, ER-5 is an explicit epistemic refusal. It is associated with minimal or absent trust, outcome-oriented assessment, and a normative commitment to human interpretive labor, especially in research and teaching [2508.03673]. The **Epistemic Alignment Framework** situates abstention within a broader design problem: aligning knowledge delivery with user preferences, evidence standards, and trust needs. One of its ten challenges is **Well-Calibrated Abstention**, defined as text that guides when the model should refuse to answer or acknowledge uncertainty. The framework represents user and system epistemic profiles as \(E_u=\langle r_u,p_u,t_u\rangle\) and \(E_s=\langle r_s,p_s,t_s\rangle\), where \(r\) captures the error–ignorance tradeoff. The paper’s central tension is that too little abstention risks harmful outputs while excessive abstention degrades utility [2504.01205]. ## 5. Scientific, multimodal, and interpretable-system abstention In scientific reasoning, abstention is framed as evidence-sensitive verification. Claims or questions are decomposed into minimal conditions, each condition is audited against evidence using NLI, and the system selectively decides whether to support, refute, or abstain. The aggregation rules are deliberately conservative: contradiction dominates support, and abstention occurs when confidence derived from condition-level evidence margins is too low. Across SciFact and PubMedQA, the paper argues that the main challenge is not selecting a single best model, but determining when available evidence is sufficient to justify an answer [2602.14189]. Multimodal work extends the same issue to vision–language models and multi-agent systems. **MM-AQA** defines **effective abstention** as recognizing evidence insufficiency and refraining from answering. Across **2,079 samples**, it finds that under standard prompting frontier VLMs rarely abstain; on A-MMMU, base unanswerable accuracy is **1.2%** for Sonnet 4.5, **4.9%** for GPT-5, and **4.0%** for Qwen 2.5. Multi-agent systems improve abstention, but at the cost of an accuracy–abstention trade-off, and sequential designs match or exceed iterative variants, suggesting that the bottleneck is miscalibration rather than reasoning depth [2604.14799]. Concept Bottleneck Models introduce another operationalization. **CREDENCE** decomposes concept uncertainty by construction into epistemic uncertainty from disagreement across concept heads, \[ U_{\text{epi}}^{(k)}=\mathrm{Var}_h[p_h^{(k)}], \] and aleatoric uncertainty from a dedicated ambiguity head, \[ U_{\text{ale}}^{(k)}=\sigma_\theta^{(k)}(\mathbf h). \] The resulting policy is quadrant-based: high epistemic and low aleatoric suggests data collection; low epistemic and high aleatoric suggests human review; high–high suggests abstain or escalate; low–low suggests automate. The paper’s claim is that abstention should be reserved for cases where both model ignorance and inherent ambiguity are high, not for all uncertain cases [2604.24170]. Across these domains, abstention is coupled to structured routing rather than simple refusal. The system is not merely silent; it declines commitment while directing the case toward additional evidence, human review, or model improvement. ## 6. Collective, governance, and measurement settings Epistemic abstention also appears in collective decision theory. A jury-theorem formulation with confidence-calibrated agents models each agent’s reliability belief as a Beta posterior and defines a publication decision \[ D_{\text{vote},i}(t)= \begin{cases} 1 & \text{if } \mathcal{C}_{i,t} > \tau_{\text{abstain},i},\\ 0 & \text{otherwise.} \end{cases} \] Only sufficiently confident agents vote. The paper derives a non-asymptotic lower bound on the probability that the majority is correct and proves an asymptotic Condorcet-style convergence result under positive average competence and a nondegenerate gate. Its interpretation is that abstention functions as epistemic filtering, reducing collective hallucination by suppressing low-confidence votes [2602.22413]. Decentralized-governance work reaches a related conclusion. **Partial abstention** means that a voter with total voting rights \(t_i\) abstains from \(a_i\) of them and exercises only \(\bar t_i=t_i-a_i\). Under the paper’s epistemic tests, partial abstention is a strong governance method because it can reproduce the optimal weighting rule \[ w_i=\ln\left(\frac{p_i}{1-p_i}\right) \] without requiring explicit transfer delegation. Proposition 3.2 states that if exercised voting rights are adjusted so that \(\frac{w_i}{t_i-a_i}=R\), then the probability of a correct collective decision is maximized. The paper argues that abstention is epistemically cleaner than transfer delegation because it reduces one’s own weight without changing the relative weights among others [2505.04136]. An older but conceptually parallel result comes from quantum metrology. There, abstention is represented by an inconclusive POVM outcome \(\Pi_0\), with abstention probability \(Q\) and conditioned fidelity \(F(Q)\). The central claim is that allowing inconclusive outcomes can drastically improve the precision of accepted estimates. For phase states, the no-abstention scaling is \[ 1-F=\frac{1}{2N+2}, \] whereas with any nonzero abstention \(Q>0\), \[ F=1-\frac{\pi^2}{16\,Q\bar Q\,N^2}, \qquad 0