Grounded question answering systems should answer only when the supplied evidence supports the answer. In multi-hop QA, this requirement is difficult because partial evidence can make an unsupported answer appear plausible. We study selective answering through evidence sufficiency boundaries: for the same question, a model should abstain under unsupported or partially supported context, answer when the context first becomes sufficient, and keep the answer stable when redundant evidence is added. We introduce Evidence Sufficiency Boundary Training, a generation-native training framework that constructs ordered evidence chains and supervises the abstain-to-answer transition directly. The method combines level supervision, a boundary flip margin, post-boundary stability, and answer recall protection. We build evidence chains from HotpotQA, 2WikiMultiHopQA, and MuSiQue, then evaluate models with chain metrics, raw QA utility, and unsupported-answer rates on external non-answerable sets. With Qwen2.5-3B-Instruct and LoRA adaptation, Evidence Sufficiency Boundary Training gives the strongest boundary localization among the tested systems, with flip accuracy of 0.807 compared with 0.781 for a token-level abstention baseline. It also achieves the lowest overall unsupported-answer rate on external non-answerable evaluation, 0.095 compared with 0.101 for the same baseline, while retaining competitive raw QA F1. The results show that grounded selective answering improves when training marks the evidence level where refusal should give way to answering.
Vision-language models are evaluated by aggregate accuracy on multimodal benchmarks, a practice that implicitly assumes the model uses its visual input. We show this assumption fails on 40%--97% of samples across six VLMs and three perceptual benchmarks: blurring the question-relevant visual region leaves the next-token distribution nearly unchanged. We name this phenomenon the Visual Insensitivity Gap and quantify it with a per-sample Visual Sensitivity Index (VSI). The gap is a property of samples, not of models: VSI ranks correlate across models (grand-mean Spearman rho=+0.40, permutation p<10^-3), so the same samples are flagged insensitive by VLMs sharing no architectural detail beyond a contrastively pretrained vision tower. The mechanism is concrete: on the insensitive samples, a linear probe on each model's own vision tower distinguishes perturbed from clean images at 0.72--0.79 accuracy, yet the model's argmax token changes on only 2%--11% of the same samples, an encoder--LLM gap above 0.65 on every model. Mapping VSI's diagnostic utility cell by cell surfaces a strong regime (multi-choice reasoning on capable VLMs: AUROC=0.85--0.87) and a weak regime (well-calibrated factuality, where softmax confidence already leads). VSI is not a universal best abstention signal; it is a sample-intrinsic indicator of vision-ignoring failure, best used as a conditional ensemble component.
KV-cache compression reduces LLM inference memory by evicting context tokens, but when the evicted tokens contain answer-bearing evidence, the model may hallucinate instead of recognizing that the compressed context is insufficient. We address this failure from a behavioral perspective: to our knowledge, this is the first work to formulate compression-aware abstention as a learning problem, in which a model learns to answer when supporting evidence survives compression and abstain when it does not. We construct supervision from compressor survival masks and tight answer-bearing spans, labeling examples as Confident when evidence survives and Abstain when it is removed. A 10.1M-parameter LoRA adapter trained on ~2.6K MuSiQue 2-hop QA examples reduces base-model hallucinations by 97% under prompt-style truncation while preserving correct answering on evidence-retaining examples. Unlike prompt-only abstention baselines, which over-abstain on many answerable high-retention examples, the trained adapter learns a conditional policy. We also evaluate the method under actual compressed-cache decoding, where multi-compressor training yields a 6-22x relative lift over the unaided base on evidence-retaining examples. Controlled-deletion experiments show that the learned behavior is driven by evidence content rather than input length alone.
Knowledge-intensive reasoning requires Large Language Models (LLMs) to ground answers in provided evidence. When evidence is insufficient, it is desirable that models abstain rather than confidently generating unsupported answers. Existing abstention methods rely on uncertainty estimation or evidence sufficiency checks, but neither tests whether the reasoning process for generation, driven by the interaction of provided evidence and the model's internal memory parameters, is actually grounded in the evidence. A key contributing factor is that entity mentions in context activate memorised associations, causing models to generate plausible responses ungrounded in evidence. We propose Twin Worlds (TW), a framework for improving reliability in knowledge-intensive reasoning through equivariance-based abstention: unlike invariance, which requires outputs to remain unchanged, equivariance requires outputs to transform correspondingly under entity substitutions. A model grounded in the evidence should produce answers that shift consistently when entities are substituted while their relations are preserved. TW constructs multiple worlds via typed substitutions of the original input that preserve relational structure while reducing parametric priors, and uses equivariance violations as an abstention signal. Across four benchmarks and three model backbones, TW identifies when answers are not reliably grounded in the provided evidence and outperforms uncertainty- and sufficiency-based baselines.
Determining whether large language models derive answers from context or prior knowledge remains a fundamental challenge. Self-contained linguistic olympiad puzzles provide a controlled setting where all answers derive solely from expert-designed context examples without external knowledge. Removing individual context examples can eliminate information needed for specific questions while leaving the rest of the puzzle unchanged. We leverage this to introduce a diagnostic framework for analyzing individual context examples. Using 53 UK Linguistics Olympiad puzzles, we generate two modified variants by deleting a single context example: (1) uniform random deletion, and (2) targeted deletion (inspired by error-correcting codes) to remove a structurally load-bearing example uniquely carrying necessary information. We formalize this impact using a Question Damage Score to classify puzzles as fragile or robust. Evaluating three frontier LLMs under instructions to abstain when information is insufficient, we find they rarely abstain, often continuing to produce correct answers after load-bearing context is removed. These findings motivate further investigation into context-based reasoning, prior knowledge, memorization, and linguistic inference. Beyond abstention, the framework enables fine-grained analyses of context reliance, including causal interventions, stopping-set analysis, targeted contamination studies, and mechanistic interpretability.
Alden Do Rosario, Hussein Younes, Felipe Pirescs.CL cs.AI
Volume-based accuracy rewards retrieval-augmented generation (RAG) systems for guessing: a system that answers everything outscores one that declines when its knowledge base cannot support an answer. Building on the confidence-target analysis of Kalai et al. (2025), we present a penalty-aware evaluation framework for deployed RAG products, combining (i) asymmetric scoring (correct +1, wrong -4, abstain 0), (ii) knowledge-gap canaries, questions whose answers are verifiably absent from the knowledge base, so that any answer constitutes ungrounded generation from parametric memory, and (iii) a failure-attribution pipeline that separates retrieval, generation, and abstention-policy failures. Applying the framework to three commercial RAG systems and a no-retrieval baseline on SimpleQA-Verified (1,000 questions x 3 repeats, graded blind by a cross-family three-judge panel with 98.9% unanimity), we find that accuracy when answering is closely clustered across systems (97.0-98.0%), while canary violation rates differ roughly sixfold (16.7% vs. 98.1%). The systems are separated less by what they answer correctly than by whether they answer at all when they should not, and penalty-aware scoring reorders the volume-based ranking accordingly; the reordering is stable across penalty settings from k=1 to k=9. All code, configurations, transcripts, and judge votes are released for independent audit.
Large language models (LLMs) are increasingly used as evaluators to assess output quality and preference alignment, yet providing reliable guarantees of agreement with human judgments remains challenging. Recent work introduces confidence-thresholding methods that provide such guarantees for pairwise comparisons, relying on the assumption that higher estimated confidence implies lower disagreement risk with humans. However, this assumption can break down when the number of candidate responses increases, since distributing probability mass across many alternatives can distort confidence estimates. To address this issue, we propose a Localize-Then-Decide framework. First, conformal prediction localizes a small shortlist that contains the human-preferred response with high probability. Then, a calibrated confidence-based rule selectively chooses a single response from this shortlist or abstains. This design restores the monotonic relationship between confidence and disagreement risk and enables high-probability agreement guarantees. Experiments with multiple candidate sizes across several datasets and judge LLMs demonstrate that our framework consistently achieves higher guarantee success rates and substantially higher coverage than single-stage baselines.
Many agentic systems must repeatedly choose between acting and abstaining, making faithful reasoning important for oversight: an explanation is useful only if it reflects the computation that produced the action. We study this problem through intervention timing in multi-party conversation, where an assistant must decide whether to speak or remain silent. This setting exposes class imbalance, asymmetric action costs, and the possibility that exposing reasoning changes the policy being audited. Using Qwen3-8B, decoded with or without chain-of-thought reasoning, we compare direct decision policies, reasoning policies, supervised fine-tuning, and reinforcement learning. We find a capability-auditability tradeoff: the strongest direct policy achieves higher quality but exposes no reasoning to inspect, while the reasoning policy provides a trace at the cost of lower performance, particularly recall of true intervention opportunities. Supervised fine-tuning either suppresses reasoning or preserves it without improving decision quality, while reinforcement learning also fails to improve the reasoning policy. We identify one mechanism underlying this failure: group relative objectives provide no learning signal on confidently wrong prompts when sampled rollouts all select the same action. Controlled activation probes and behavioral ablations show that standard faithfulness methods can overstate evidence that exposed reasoning reflects the underlying decision process. Probability-based metrics saturate under confident decisions, probes are vulnerable to class imbalance and textual leakage, and reasoning ablations can confound reasoning content with changes in inference mode. Together, these results show that exposing reasoning can change an agent's action policy rather than simply make it observable. We provide controls for evaluating reasoning-based oversight of agents that can act or abstain.
Utilities lose a substantial share of treated water to leakage, yet rarely trust artificial-intelligence localizers to dispatch crews: guessing everywhere cannot justify excavation. The gap is accountability, not accuracy: no method proves when it should not act. Here we recast leak localization as decision-making under verifiable abstention. A physics-grounded executor agent falsifies hypotheses (leak, demand, sensor, valve) against a digital twin; an independent supervisor agent, with a large-language-model (LLM) auditor, checks evidence against a code-verifiable contract, then certifies a dispatch, requests evidence or abstains. Under field-grade noise, a 32% forced baseline becomes 96% decision precision on acted events. On an independently generated benchmark it acts on only 4 of 33 leaks, all correct. A 194-event register of audited real leak locations with twin-simulated pressures and flows yields five excavation dispatches, three correct, and 44% survey recovery at full district precision. Accountable abstention offers a defensible route to autonomous water-infrastructure operation.
Mohaimenul Azam Khan Raiaan, Nur Mohammad Fahadcs.CV
Medical vision-language models (VLMs) can achieve high accuracy but remain unreliable: they are systematically overconfident, benefit little from test-time reasoning, and lack the ability to reliably calibrate trust in their own responses. We introduce EVADE (Evidence-Verified Agentic Diagnosis with Escape), an inferential, non-training method that enhances the safety of deploying a single frozen VLM. EVADE responds and, when uncertain, localises the region most diagnostically relevant, re-answers on a zoomed view, and commits only when both the entire image and the zoomed view responses agree; otherwise, it abstains. To directly address verification hallucination in single-model self-checking, our main idea is to verify gate consistency across different image views rather than re-reading the model's own text. Experimental evaluation on VQA-RAD, SLAKE, and PathVQA using Qwen2.5-VL-7B reports that EVADE is the only method that simultaneously improves both calibration and selective risk while maintaining accuracy, reducing expected calibration error (ECE) by up to 45% compared to zero-shot. Chain-of-thought, self-consistency, and self-verification all fail at least one axis. A grounding analysis reports that self-proposed regions perform better at diagnostic structure localisation than centres or random crops. However, a 7B VLM cannot use this localisation to revise answers. Therefore, reliability gains come from the consistency gate and calibrated abstention.
Kirill Borodin, Vasiliy Kudryavtsev, Ivan Viakhirev +1cs.CL cs.LG cs.SD
Modern encoder-decoder systems can produce fluent text even when their input contains no recoverable message. We study this failure in ASR and NMT through the models' reserved null tokens, asking whether the score for ending generation already carries a usable abstention signal. Across speech recognizers and translation models, we audit native null-token scores and scalar logit shifts. In Whisper, we additionally probe decoder states and compare supervised row edits with conventional external gates. The evaluated models often expose a useful abstention signal, but stock decoding does not reliably act on it. Raising the null-token score can sharply suppress fabrication, but aggressive intervention also deletes valid speech or shortens legitimate translations. These findings turn the null token into a diagnostic lens on hallucination and motivate evaluating abstention methods by both suppression and deletion costs, rather than by hallucination reduction alone.
Sentiment classifiers are increasingly applied to social media content that is either sarcastic or AI-generated --- two distributional regimes where standard evaluations offer little guidance. We present a three-part empirical study of sentiment classifier behaviour under these conditions. First, we find that confidence scores on sarcastic text are significantly lower than on non-sarcastic text (Mann--Whitney $p = 2 \times 10^{-6}$), confirming that classifiers sense their own uncertainty on ironic content even without explicit uncertainty modelling. Second, and counterintuitively, we show that sentiment classifiers achieve higher accuracy on AI-paraphrased reviews than on the original human-authored text (RoBERTa: $+5.8$ pp for Qwen3.5-4B paraphrases, $+3.7$ pp for Gemma4-E4B), revealing a cross-domain stylistic alignment effect: AI paraphrases remove distributional noise that confounds Twitter-trained classifiers, producing cleaner, more prototypical sentiment text. Third, we demonstrate that a lightweight abstention wrapper --- flagging the $14\%$ of inputs with confidence below $0.6$ --- improves accuracy from 82.2\% to 88.9\% ($+6.7$ pp) on the retained set. We further compare Semantic Entropy and MC-Dropout-style disagreement as uncertainty signals and find near-identical AUROC ($0.650$ vs.\ $0.646$) on sarcastic text, suggesting that for short social media inputs, both methods are interchangeable. Our results motivate a shift from confident single-label prediction to uncertainty-aware abstention in high-stakes sentiment applications such as mental health flagging and content moderation.
A frozen language model on reasoning tasks has two coupled weaknesses: it under-uses evidence its own residual stream already encodes, and it fails to detect when the input is insufficient to answer, so it confabulates. This paper consolidates two research lines that address these on the same residual stream: a conditional steering probe writes the stream at mid-stack layers and recovers reasoning accuracy from a frozen backbone, and a zero-shot sufficiency direction reads the stream and abstains when information is insufficient. Deployed in one forward pass they interfere: the steering write shifts the state the direction reads, costing up to 8 AUROC points of cross-domain transfer on small models; a separate clean pass doubles inference cost. We keep the direction fixed and train a small network to reconstruct the pre-steering residual from the steered one -- mean-squared error on (steered, clean) pairs, no sufficiency labels -- and read the direction on the reconstruction. The resulting system, YOPO (You Only Pass Once), answers, steers, and abstains in one forward pass of a frozen Qwen2.5 backbone (1.5B/3B/7B). End to end, three-way accuracy more than doubles the frozen baseline (0.375->0.798 on 1.5B alphaNLI) and one pass beats the two-pass reference at every scale (0.798/0.830/0.893 vs 0.753/0.790/0.863) and on ten backbones across six model families. We chart the capacity-transfer frontier quantifying the principle that abstention should not be trained in; a source-side audit catches our own alphaNLI construction leaking a surface artifact, so architectural claims are anchored on native-label replications (SQuAD2, RepLiQA, MuSiQue); and on the standard four-domain suite we contribute, to our knowledge, the first answer-or-abstain benchmark, where our gate tops every in-domain dataset and the label-free direction is the only gate family to survive domain transfer.
Large language models have made natural language interfaces to databases (NLIDB) newly credible, but LLM text-to-SQL systems fail in a way that matters for deployment: a hallucinated column or a mis-aggregated total yields a fluent wrong answer, indistinguishable at the point of use from a right one. Where the consumer cannot inspect the generated query, as in enterprise AI deployments and operational dashboards, and increasingly where the consumer is a tool-using agent rather than a person, accuracy alone is insufficient: nothing marks which answers to distrust. This is a reliability problem before it is an accuracy problem. We propose an architectural pattern for such systems, a trusted kernel with a generative shell, resting on one invariant: a component that can fabricate may influence which question the system answers, never which value it returns. A generative shell interprets underspecified input and phrases replies; a deterministic kernel matches fully specified questions against a bounded set of answerable question shapes and compiles them to queries by deterministic execution. The two meet at a confirmation the user reads before any value is computed, and requests the kernel cannot express are declined rather than approximated. We call this structural abstention, and distinguish it from the statistical abstention of selective prediction and calibrated confidence: refusal here needs no confidence estimate, because unanswerable requests are unrepresentable. We specify the pattern implementation-independently, give a five-decision recipe and work it across three domains, extend the invariant from returned values to the actions of agentic systems, and report a two-year production case study alongside two generative alternatives, a fine-tuned parser and a tool-retrieval agent. We close against enterprise and reliability benchmarks published since.
Fnu Pramono, John Cai, Sourabh Kulkarnics.CV cs.AI cs.CL cs.LG
When visual evidence is occluded or chaotic, models should abstain. In this paper, we show that Vision-Language Models (VLMs) can internally distinguish when abstention is required, but fail to express it anyway. We introduce TRAPSBench, a procedurally generated video benchmark of 1,404 matched physics pairs in which a single targeted change renders the outcome undeterminable from the visual evidence. Furthermore, we introduce Penalized Epistemic Calibration Score (PECS), a new robust metric that requires models to both answer correctly when the outcome is knowable, and abstain when the outcome is not. Across 16 VLMs spanning five families, spontaneous restraint is poor: the best PECS is 0.292. The bottleneck is expression, not perception: linear probes decode answerability from hidden states at up to 0.91 AUROC across physics domains; steering a single-layer void direction causally induces or suppresses abstention. Our results replicate across three open-weight families (Qwen, Gemma, LLaVA). The failure is also more pronounced in visual than textual uncertainty: models detect textual impossibility about 4x more readily than missing visual evidence. Closing this representation--output gap likely requires output-stage interventions.
Machine learning models should be robust, in the sense of remaining predictively consistent under permissible variations. A model's predictions should ideally remain unchanged when it is replaced by a functionally equivalent one, or when its inputs are subject to minor, admissible perturbations. If such changes alter a prediction significantly, then the prediction is "ambiguous" with respect to the model. Models should abstain from making such ambiguous predictions and/or should flag them for human inspection, especially in high-stakes decision-making scenarios. However, in practice, such ambiguity is not easy to identify once a model is deployed. Here, the Robust Ambiguity Detection (RAD) framework is advanced for quantifying predictive ambiguity using two complementary metrics: Model-Space Consistency and Feature-Space Consistency. These two scores, the RAD Score-Pair, visualised through the RAD Plot, provide an interpretable characterisation of the sources of ambiguity and the actions a user may consider in response. RAD is evaluated on synthetic datasets with systematically controlled overlap, as well as several real-world datasets where the level of ambiguity cannot be directly inspected. Finally, we demonstrate a downstream application of RAD where samples are ranked by their RAD Pareto-Rank and the most ambiguous are abstained from prediction, achieving performance comparable to existing rejection-based approaches.
Uma Ranjan, Kunal Tilaganji, Aditya Koul +9cs.CV cs.SC
Large language models (LLMs) often rely on shortcuts rather than systematic reasoning, raising safety concerns in medical applications. Allowing models to abstain when uncertain improves reliability but introduces a coverage accuracy tradeoff. We propose a two-stage framework for medical hypothesis verification in multiple-choice settings that manages this tradeoff through targeted ontology grounding, applied only when the model abstains. We show that abstention is not random but reflects genuine uncertainty, with abstained predictions associated with lower confidence. Across two frontier models (GPT-5.5, accessed via the Azure OpenAI API, and DeepSeek-R1), the proposed framework improves question-level accuracy by 9.6 percentage points (82.9% to 92.5%) and hypothesis-level accuracy by 4.2 percentage points (92.0% to 96.2%). Our experiments conducted on MedReason and MedQA show that abstention can be repurposed as a control signal for selective reasoning refinement, achieving knowledge-graph-level performance without explicit knowledge graph construction.
Large reasoning models (LRMs) are prone to hallucination, which undermines their reliability and poses challenges for safe deployment. Hallucinations in LRMs arise from two distinct failure sources: reasoning hallucination, where flawed inference steps propagate to an incorrect conclusion, and knowledge hallucination, where the model lacks the requisite factual knowledge to answer the query. To address reasoning hallucination, we propose REIN, an alignment framework that trains LRMs to produce a structured reasoning sequence, $\texttt{<think>} $$\rightarrow$ $\texttt{<reflection>} $$\rightarrow$ $\texttt{<answer>}$, enabling explicit self-reflection before committing to a final answer. To address knowledge hallucination, REIN introduces a reward mechanism that encourages explicit abstention (e.g., "I don't know") when none of the sampled reasoning chains yields a correct answer, allowing the model to refrain from unsupported predictions. Extensive evaluations on mathematical and commonsense reasoning benchmarks show that REIN consistently improves selective accuracy, reduces incorrect-but-self-endorsed responses, and maintains high coverage compared with competitive baselines. Notably, REIN achieves these gains within a single forward pass, without requiring process supervision, inference-time controllers, external search, or multi-round critiques. Experiments on multiple backbones show that REIN reduces the hallucination proxy by $58\sim72\%$ relative to the base models while maintaining $86\sim91\%$ average coverage, and improves selective accuracy on attempted questions by $6.6\sim14.2\%$.
Agent systems increasingly expose execution traces, yet telemetry that reveals a failure may still be inadequate for identifying where that failure originated. We introduce TelemetrySuffBench, a controlled benchmark that separates failure detection, fault-origin localization, and safe abstention under insufficient evidence. The benchmark constructs canonical multi-component traces with delayed-binding faults and renders them as paired coarse views, seven-factor telemetry masks, and exact-equal ambiguous origin pairs. We evaluate five frontier language models using unified protocols, explicit candidate sets, invalid-output accounting, subgroup analyses, and a frozen blind holdout. With full telemetry, origin-step Top-1 accuracy ranges from 33.8% to 97.2% across models. Metadata, OpenTelemetry-compatible, and OpenInference-compatible views retain 99.5% to 100% detection F1 while limiting origin-step accuracy to at most 0.5%, exposing a robust detection-localization gap. Factor ablations further show that removing decision content reduces origin-step accuracy to zero for every model, while provenance removal also causes large model-dependent losses. On rich ambiguous inputs that require abstention, evidence gating reduces unsupported unique-origin answers by 12.5 to 48.6 percentage points for three models, whereas two models still answer every case, revealing strong model dependence in safe abstention. Results on the frozen holdout reproduce the central pattern within the same generator family. These findings show that terminal status can support detection, whereas reliable causal attribution requires explicit decision-to-provenance links and abstention safeguards that remain effective across models. The dataset and benchmark implementation are available at https://anonymous.4open.science/r/TelemetrySuffBench-E635/README.md.
Javier Fumanal-Idocin, Javier Andreu-Perezcs.LG cs.AI
Interpretable classification often requires more than accurate predictions for real-life deployment: models should be transparent about the evidence behind their decisions and abstain when they cannot decide reliably. We introduce Fast Evidential Rule Learning (FERL), a method that learns interpretable, accurate fuzzy rule models whose outputs are evidential. Unlike post-hoc calibration, FERL's belief, plausibility, and abstention capabilities arise directly from the fuzzy memberships in a single deterministic pass, with no auxiliary head, held-out set, or repeated inference. Our theoretical analysis further shows that FERL is Lipschitz stable, which means that its evidential outputs vary smoothly with the input. Against state-of-the-art rule learners, FERL is statistically significantly more accurate across a 30 tabular-dataset benchmark ($+2.6\%$ average accuracy over the second best). Its native set predictions attain the best utility-discounted accuracy among credal classifiers ($u_{65}/u_{80}=0.80/0.83$ vs.\ $0.79/0.80$ for the naive credal classifier), at higher set coverage ($0.92$ vs.\ $\le0.82$). FERL also matches dedicated out-of-distribution detectors on tabular near-OOD detection ($77.7$ vs.\ $77.4$ AUROC for the strongest baseline). Under detector-class-disjoint concept-bottleneck evaluation, its it is within $2.3$ AUROC points of the strongest dedicated detector on both CUB and AwA2, while attaining the best AwA2 AUPR-Out ($68.3$) and novel-class rejection ($57.2$), while being able to name which attributes are anomalous.
Vision-language systems combine images with retrieved text, but these sources can disagree or jointly fail to support an answer. Reliable models must identify the trustworthy source and abstain when neither is adequate. Existing post-training objectives score instances independently and therefore do not enforce coherent behavior under counterfactual evidence changes. We introduce CARGO-VL, a group-relative framework that optimizes matched variants covering aligned, image-correct, text-correct, and both-wrong (A/V/T/N) evidence states as one bundle. Its objective couples condition-wise correctness with transition rewards for answer invariance, source equivariance, and answer-to-abstention switching, while a primal-dual controller balances unsafe answers against excessive deferral. We also contribute XMC (eXtended Modal Conflict), a four-condition conflict training resource, and evaluate transfer on CMC-Bench and Modality-Bias. Across multiple seeds, CARGO-VL improves conflict handling, unsupported-answer avoidance, and modality balance over pointwise baselines. Ablations identify complementary benefits from relational transition signals and adaptive risk control, supporting counterfactual consistency as a practical objective for reliable multimodal evidence arbitration.
Error-penalized scoring rules ($+1$ for a correct answer, $-λ$ for a wrong one, $0$ for abstaining) are increasingly prescribed against hallucination: a rational agent facing such a rule answers exactly when its correctness probability exceeds Chow's threshold $t^\ast=λ/(1+λ)$. We prove that a KL-anchored gradient learner can do the opposite. When abstention is a discrete action, the reward gradient and the anchor's restoring force are throttled by the same gate-saturation factor and die together: under explicit conditions (among them, blanket answering loses score in expectation and prompts share a bounded readout) the model drifts toward refusing everything, its mean training reward rising to zero like $1/t$ in training time $t$, so the curve reads as improvement while coverage collapses. The advantage estimator compounds the failure: in its sparse-answer regime, group normalization silently replaces every designed penalty with an effective penalty of one, moving the learned threshold from $t^\ast$ to $1/2$. The repair is structural: train a mandatory confidence report with a strictly proper score plus a correctness reward, and abstain only at deployment by thresholding the report. The always-emitted report has no gate to saturate, so no shared factor can kill its reward gradient and its anchor together, and its calibrated optimum is attracting. Simulations confirm every prediction, and experiments on language models at two scales confirm the mechanism live: the rule silences questions the models demonstrably still solve within ten optimizer steps, an ablation isolates the cause, and report-level training raises coverage, accuracy, and calibration together.
High-stakes decision systems in credit scoring, fraud detection, healthcare, and industrial safety require reliable uncertainty quantification under severe class imbalance and asymmetric error costs. Standard marginal conformal prediction (CP) provides valid overall coverage guarantees; however, we show that it severely under-covers rare, costly minority classes, with minority-class coverage dropping to as low as 0.5% on certain datasets. To characterize and address this limitation, we conduct a comprehensive benchmark comparing marginal CP, class-conditional (Mondrian) CP, and cost-controlled abstention mechanisms across 15 real-world imbalanced tabular datasets, 7 classification models, 3 probability calibration techniques, and 10 random seeds, resulting in 3,150 experimental runs. Our results show that Mondrian CP restores valid minority-class coverage, achieving an average minority-coverage improvement of 61.7 percentage points over marginal CP (p < 1e-80). Furthermore, combining Mondrian CP with cost-controlled abstention significantly reduces expected decision cost compared with standard decision boundaries, confidence-based rejectors, and risk-controlled rejectors under realistic human review budgets. We further quantify dataset-specific break-even thresholds at which deferring ambiguous instances to human experts becomes cost-effective. These findings provide practical guidance for deploying distribution-free, cost-aware uncertainty quantification in high-stakes decision support systems.
Model merging combines independently trained or fine-tuned models, but pairwise alignability does not imply globally consistent alignment. We formulate merging as a finite descent problem in which checkpoints are local objects, alignment maps are transitions, and cycle products are residuals. TwistedMerge is a conservative certification pipeline that separates fixed-chart averaging, synchronization-removable gauge inconsistency, a certified central obstruction on a specified comparison complex, and nonabelian holonomy. A residual is promoted to a cohomology class only after inverse-consistency, coefficient-identification, centrality, and closure tests; otherwise the method abstains and returns an ordinary or synchronized fallback. We prove a constant-edge no-go result, frozen-complex three-way and predeclared-family error-control theorems, and a refinement test for comparison-complex sensitivity. A planted neural alignment defect is removed by cycle-consistent synchronization, showing that a nonzero cycle score alone is not a higher obstruction. Controlled central systems recover the predicted non-coboundary and projective-rank behavior, while noisy estimates move from certification to abstention without false lifts on the tested controls. A trained low-rank-adapter audit shows that naive factor averaging depends on the chosen GLr representative, whereas global factor synchronization and dense-delta SVD are stable. On natural checkpoint collections, cycle residuals do not predict merge degradation and no natural central or period-index class is certified. The results position descent theory as a falsifiable certification and abstention framework.
Sagnik De, Sreenija Pavulurics.AI cs.ET cs.LG eess.AS
Hallucination and abstention benchmarks rarely establish that a model could not have known the correct answer, making it difficult to distinguish appropriate abstention from an unsupported prediction. Seven large language models were evaluated on the TAME Pain speech corpus. Participants read phonetically balanced Harvard Sentences while one hand was immersed in cold or warm water and reported pain only during periodic pain statements. This protocol generated 5,750 no signal Harvard Sentence utterances whose transcripts contained no lexical pain information and 1,294 signal pain statement utterances in which the pain rating was explicitly spoken. In the no signal arm, pain was recoverable from acoustic features (AUC 0.622, 95% CI 0.553 to 0.662), whereas transcript based prediction was near chance (AUC 0.489, 95% CI 0.418 to 0.504). Because automatic speech recognition removes the acoustic pain cues, any pain score inferred solely from the transcript is unsupported by the available evidence. Under cooperative prompting, six models abstained on nearly all no signal transcripts, correctly extracted spoken pain ratings in the positive control task with accuracies ranging from 0.939 to 1.00, and maintained an expected calibration error of at most 0.100. Under authority framed prompts, abstention became prompt dependent, with the same model ranging from 0.18 to 1.00 across equivalent prompt phrasings. Most models produced low confidence estimates when forced to answer, whereas Gemini 2.5 Flash and Llama 3.1 8B consistently generated confident pain scores with confident fabrication rates of 0.53 and 0.76, compared with at most 0.15 for all other models. No significant demographic effects were observed in forced responses, with all $p$ values greater than or equal to 0.20.
Recent advances in equipping Large Language Models (LLMs) with search tools and outcome-reward reinforcement learning (RL) have achieved new state-of-the-art results on open-domain QA tasks. However, we argue that current training paradigms harbor a critical vulnerability: they predominantly reward correct answers but fail to penalize fabricated ones when retrieval fails, thereby implicitly exacerbating hallucinations. To address this, we propose Abstention-Aware Reinforcement Learning (AWA-RL), which dynamically shapes the abstention reward utilizing the model's query-specific prior capabilities and continuous on-policy training observations. We also introduce a novel metric, RA-F1, to measure the capability-reliability trade-off. Compared to non-abstaining baselines, AWA-RL boosts absolute precision by up to 10.3% and overall RA-F1 by 2.9%, with only marginal sacrifice in raw accuracy. These results confirm that AWA-RL successfully yields highly capable and reliable search agents. The code, data, and model weights are publicly available at https://github.com/zfj1998/AWA-RL.
Agent systems based on large language models (LLMs) are increasingly deployed for autonomous tasks, yet existing evaluations mostly focus on task success rather than whether agents know when to abstain. This gap poses real risks: under ambiguity, conflicting constraints, or tool failures, agents may execute unintended and irreversible actions. To close this gap, we present the first systematic evaluation framework for agentic abstention: the calibrated ability of tool-using LLM agents to recognize when not to act. At its core, AgentAbstain is a paired-task benchmark built on an agent-native taxonomy of 8 abstention scenarios across pre-execution reasoning and runtime discovery. It contains 263 paired tasks across 42 executable sandbox environments, where each pair consists of a should-act task and a should-abstain variant produced through a controlled perturbation to the instruction, tool, or environment state. To scale this paired design and resist data contamination, we propose AbstainGen, a fully automated pipeline that synthesizes sandbox environments and generates paired tasks end-to-end, validated by deterministic replay and semantic LLM judges; fresh task instances can be regenerated on demand, and three independent annotators rate 94-98% of sampled tasks as well-designed. Across 17 frontier LLMs in 4 agent harnesses, the best agent (Gemini 3.1 Pro) achieves only 59.5% paired accuracy (correct on both the act and abstain sides of each paired task). More importantly, abstention capability is largely independent of general task-solving capability, indicating that scaling task-solving alone will not close this gap. We further identify failure modes such as post-hoc abstention, in which agents execute irreversible actions before recognizing abstention triggers. Our code and dataset are open-sourced at agentabstain.github.io.
A model should refuse two different things: answers it would get wrong, and questions it should not answer at all, such as unanswerable ones or ones resting on a false premise. The usual recipe thresholds a single confidence score, which cannot tell these apart. Across five instruction-tuned models from three families (2B to 14B), we find they are separate axes. Ordinary answer-confidence tracks whether an answer is right but is nearly blind to whether the question is answerable; a linear probe on hidden states does the reverse. The blind spot does not shrink with scale. It is worst on naturally occurring false-premise questions (CREPE). There, answer-confidence, P(IK), P(True), and even asking the model outright whether a premise is false all stay near chance, while a hidden-state probe reaches 0.69 to 0.77 AUROC: the model represents a problem it will not report. This turns out to be fixable. Instructing a model to check premises backfires, because it then disputes sound and false premises alike (57% false challenges), unable to tell them apart; routing the same instruction with the probe roughly triples challenge precision. We turn the two axes into a calibrated policy that answers only when an answerability score and a correctness score each clear a separately certifies behave differently: the unanswerable-answer rate is controllable at every scale, while the wrong-answer rate is capped by model accuracy, so the guarantee tightens as threshold policy certifies both budgets at 0.75 coverage of correct answers, against 0.31 for a single threshold; at 14B it is the only policy that certifies at all.
Confidence is an estimate of the probability that a chosen answer is correct. Verbal confidence reports are widely used as uncertainty measures in large language models, but whether they are best understood as estimates of correctness is unclear. We test this with a two-stage abstention paradigm from the neuroscience of perceptual decision making: a model first answers and reports its confidence, then decides whether to commit it to a user or abstain. Across four non-reasoning models, prompt framings, and confidence formats, verbal confidence predicted the commit/abstain decision substantially better than whether the answer was correct. Calibrated token log-probabilities showed the opposite profile, with abstention-prediction coupled to correctness discrimination, the signature of an answer-evidence signal. After removing the variance verbal confidence shared with log-probabilities, the residual stayed aligned with commitment while its link to correctness fell to near chance. The dissociation generalised to four reasoning models across four benchmarks of varying difficulty, from hard multiple-choice to frontier-level freeform questions. Mechanistic analyses in Gemma 3 and 4 were convergent: a post-answer state known to causally support verbal-confidence generation already encoded the future abstention decision before the abstention prompt, organised mainly by that decision rather than by correctness, the two lying in approximately orthogonal directions in activation space. Steering along a verbal-confidence-specific direction causally shifted abstention. Verbal and log-probability confidence are thus not interchangeable: log-probabilities track answer evidence and correctness, whereas verbal confidence is better understood as a behaviour-facing readout of an internal commit-readiness state, challenging the practice of treating verbal reports as proxies for reliability.
Yuqi Huang, Yunlong Hou, Vincent Y. F. Tancs.LG cs.IT stat.ML
We study the Bayesian fixed-budget best-arm identification problem in which a learner can abstain from making a terminal recommendation. Subject to an abstention budget $α$, we analyze the probability of undetected error--the risk of recommending a suboptimal arm without abstaining. Our central finding is that abstention induces a phase transition: without abstention, the error probability decays polynomially in the sampling budget $T$; in contrast, introducing any small positive abstention budget shifts this to an exponential decay. For Gaussian priors and rewards, in the regime $T\to\infty$ followed by $α\downarrow0$, we establish exact matching information-theoretic lower bounds and algorithmic upper bounds on the optimal error exponent, which takes the form $\exp(-\frac{α^{2}T}{8κ_ν^{2}})$. The hardness parameter $κ_ν$ represents the prior density of the top-two gap at zero, highlighting that nearly tied instances drive the fundamental error. We introduce an adaptive algorithm, PGWS, that successfully achieves this optimal exponent by expending its abstention budget on statistically ambiguous instances. We further demonstrate that this polynomial-to-exponential improvement is exclusively a Bayesian phenomenon--in the frequentist setting, abstention only affects lower-order exponent terms. We also extend our results beyond the Gaussian model.