Ask a model "how many settlements did the Normans control?" and it answers that the passage does not say. Tell it "given that the Normans controlled 12 settlements of 400 inhabitants, what is the total?" and it computes 4,800. Nothing was added to the passage; the unsupported fact merely became usable. Benchmarks for unanswerable questions are built almost entirely by omission, deleting the fact a question needs and scoring refusal, so they never probe this. Holding source support fixed and varying only whether an unsupported premise is usable, flagging drops 8-30 points and commitment to an answer rises 1.5-2.3x on GSM8K across four models spanning two families and 2B-26B; all 16 paired contrasts remain significant after Holm correction. Two further manipulations reproduce it, one also on SQuAD: a fact asserted rather than asked about, and a corrupted intermediate value rather than a removed one. Models articulate the exact gap they then ignore, stating that "there is no financial information provided" for an inert premise and silently adopting a fabricated price for the same item. We read these as one behaviour: whether a model checks depends substantially on whether usable material is present, even when source support is held fixed. Because the main design never instructs models to reject premises introduced in the question, adopting such a premise can also be read as conditional reasoning, and we delimit what the design does and does not show. In the judging seat the pattern differs: asked whether the answer follows, judges pass solutions with a required step deleted 78-99% of the time at every scale tested. A natural prompt-level mitigation lowers accuracy by 22 points without improving detection.
Figures & tables
flagged unanswerable
committed to answer
Model
Entity
inert
action.
p
inert
action.
p
Gemma-4-E2B
present
0.853
0.553
10−12
0.367
0.547
< 0.001
Gemma-4-E2B
absent
0.867
0.640
10−7
0.293
0.513
10−6
Qwen2.5-3B
present
0.673
0.473
< 0.001
0.467
0.753
10−10
Qwen2.5-3B
absent
0.693
0.553
0.009
0.280
0.653
10−13
Qwen2.5-7B
present
0.887
0.740
10−5
0.187
0.400
10−7
Table 1: Actionability. Identical framing, entity and prompt; only whether the unsupported premise can be used varies. Rates over 150 items per cell (95% confidence interval at most ± 8 points); p from paired exact McNemar tests, all 16 of which survive Holm correction. Metrics are the audited ones (§ 3 ). q 4-bit quantised.
flagged unanswerable
confabulated
Domain
Model
quest.
assert.
p
quest.
assert.
p
GSM8K
Gemma-4-E2B
0.867
0.487
10−13
0.033
0.267
10−8
GSM8K
Qwen2.5-3B
0.673
0.393
10−8
0.220
0.433
10−6
GSM8K
Qwen2.5-7B
0.827
0.720
0.020
0.127
0.273
0.001
SQuAD
Gemma-4-E2B
0.947
0.527
10−16
0.020
0.280
10−10
SQuAD
Qwen2.5-3B
0.980
1.000
0.250
0.020
0.000
0.250
Table 2: Assertion vs. interrogation. The same missing fact posed as a question versus asserted as a premise, entity held present. Confabulated : the response commits to an answer that the source does not support.
flagged INVALID
passes with
Judge
intact
step deleted
fact invented
p
step deleted
Gemma-4-E2B
0.033
0.220
0.213
1.000
78.0%
Qwen2.5-3B
0.000
0.013
0.007
1.000
98.7%
Qwen2.5-7B
0.000
0.080
0.073
1.000
92.0%
Gemma-4-26B q
0.020
0.047
0.293
10−10
95.3%
Table 3: The judging seat. Rate at which a judge marks a solution INVALID. The predicted asymmetry is absent at ≤ 7B ( p=1 ) and inverts at 26B. Solutions with a deleted step pass at a high rate at every scale.
Large language model reliability studies whether outputs remain correct, calibrated, robust, and controllable across prompts, contexts, and deployment conditions.
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.
Millions of users turn to AI models for their information needs. It is conceivable that a large number of user queries contain assumptions that may be factually inaccurate. Prior work notes that large language models (LLMs) often fail to challenge such erroneous assumptions, and can reinforce users' misinformed opinions. However, given the recent advances, especially in model's reasoning capabilities, we revisit whether large reasoning models (LRMs) can reason about the underlying assumptions and respond to user queries appropriately. We construct queries with varying degrees of presuppositions spanning health, science, and general knowledge, and use it to evaluate several widely-deployed models When compared to non-reasoning models, we find that reasoning models achieve a slightly higher accuracy (2-11%), but they still fail to challenge a large fraction (26-42%) of false presuppositions. Further, reasoning models remain susceptible to how strongly the presupposition is expressed.
Multimodal reasoning requires models to combine multiple pieces of evidence into decisions that satisfy all relevant constraints. Evaluations centered on final-answer accuracy leave unclear whether failures arise from reading individual facts or using them together. We study this question through the multi-evidence integration gap, linking local fact recovery to global decision validity on controlled subset-selection and path-finding tasks. Our evaluation combines conditional performance given correct local answers with same-grid local queries, text-oracle inputs, retrieval controls, and constraint ablations. We further examine hidden-state probes and activation replacement across network depth, token positions, and model families. Across the evaluated models, accurate local recovery coexists with substantial global failure. For example, Gemma-3-27B recovers all evaluated isolated item facts correctly but achieves only 19.75% global validity on eight-item visual instances. Probes decode information about solution membership from global-prompt states, while corresponding clean-state replacement recovers failed decisions over model-dependent depth ranges. These findings show that recovering facts and representing solution-related information are insufficient for reliable joint decisions. They motivate evaluations and model improvements that explicitly target how available evidence is combined, alongside advances in perception and model scale.
Yuyang Dai, Yuxia Wang, Zhuohan Xie
Florida State University · INSAIT, Sofia University “St. Kliment Ohridski” · MBZUAI