Source-linked AI summary
Knowing When Not to Answer: Cross-Domain and Multi-Turn Generalization of Latent Underspecification Signals
Jerzy Kamiński, Ilya Galyukshev, Artem Kuznetsov, Danil Fedorov, Kirill Redko, Sergey Chuprin, Aidar Shumbalov, Stanislav Chumakov, Anna Kalyuzhnaya
TL;DR
It is unclear whether different forms of unanswerability share a hidden-state representation, or whether probes can detect when dialogue is still underspecified. The paper tests transfer across datasets and dialogue structure, then evaluates a calibrated clarification gate; transfer is strongest within grounds, while cross-structure detection and end-task gains remain limited.
Problem
It remains unclear which forms of unanswerability share a hidden-state representation, and prior work does not detect turn by turn when dialogue remains underspecified.
Source: §1 · p. 1Method
The paper evaluates hidden-state probes across datasets and dialogue structure, using a turn-labeled benchmark and an answering-user harness to test a calibrated probe gate.
Source: §1 · p. 1, §2.4 · p. 3, §2.5 · p. 4Results
Probes transfer within shared grounds, but not reliably across structure; the gate comes within 0.08 of a true-label gate in end-task success yet does not beat consolidation.
Source: §5 · p. 8Takeaways & Limitations
The evidence supports transfer within a ground, while separate representation of epistemic unknowability and practical gains from probe-gated clarification remain unsettled.
Source: §5 · p. 8The benchmark uses generated rather than naturally occurring dialogue, and probe results are decoding claims because no activation-steering ablation was run.
Source: Limitations · p. 9
Abstract
from arXiv · showhide
Large language models routinely answer questions that cannot be answered from the information given, and in dialogue they answer before enough has been said. Unanswerability is linearly decodable from hidden states, but it is unclear which of its forms share a representation and whether the signal is useful in dialogue. We contribute a turn-labeled multi-turn benchmark (423 conversations, 1,661 labeled turn-states) and an evaluation harness with a simulated user who answers clarifying questions, and use them with six datasets and six open-weight LLMs to test how far probes for unanswerability carry. Probes transfer robustly between datasets that share a ground of unanswerability: missing information in math (AUROC 0.77-0.97) and in a passage (SQuAD 2.0<->MuSiQue, 0.77-0.90). Probes for epistemic "known-unknowns" transfer poorly to math, but this separation weakens under lexical controls and changes with layer and coordinate system, so it remains unresolved. Single-turn probes fail zero-shot to detect when a conversation becomes answerable; in-structure probes recover it, but no better than a bag-of-words classifier. A gate on the calibrated probe, with no model fine-tuning, fires on underspecified turns far more precisely than chance, and its end-task success comes within 0.08 of a gate given the true labels. Yet across four models it does not reliably beat vanilla generation or prompted consolidation. The remaining gap lies mostly in how models use a clarification, not in detection.
1 Introduction
Prior work establishes that unanswerability can be read from hidden states, but it remains open whether different grounds share a representation and whether the signal works across dialogue turns. This paper tests that premise with cross-dataset probes, a turn-labeled benchmark, and a clarification harness, finding robust transfer within grounds but limited generalization across structure and deployment.
Motivation
- The paper addresses uncertainty about whether missing information in a question, missing information in context, and epistemic unknowability share a usable hidden-state direction.All three forms call for abstention, but need not share a latent direction readable without fine-tuning.Source: §1 · p. 1
- Incrementally revealed instructions degrade model performance, but prior work leaves open the scope and lacks a turn-by-turn detector of underspecification.The paper tests whether a domain- and structure-general signal could gate generation without model fine-tuning.Source: §1 · p. 1
RQ1: Across datasets
- Probes transfer robustly between datasets sharing a ground of unanswerability, while whether epistemic unknowability has a separate representation remains unresolved.The apparent separation changes with lexical controls, layer, and coordinate system.Source: §1 · pp. 1–2
RQ2: Across structure
- Single-turn probes fail zero-shot on turn-level dialogue states; in-structure probes recover answerability only to the level of a TF-IDF classifier on user turns.The contribution presents this as a turn-labeled multi-turn benchmark of 423 conversations and 1,661 turn-states.Source: §1 · p. 2
RQ3: At deployment
- At deployment, the probe gate fires on underspecified turns more precisely than chance and comes within 0.08 of an oracle gate, but does not beat prompting workarounds.The end-task endpoint fixed in advance remains negative.Source: §1 · p. 2
Contributions
- The evaluation harness uses a simulated answering user and oracle, position, and rate-matched random gates to separate detector error from intervention cost.The user answers from unrevealed constraints without leaking beyond them.Source: §1 · p. 2
2 Method
The methods evaluate frozen hidden-state probes across datasets and structures, construct a repaired turn-labeled dialogue benchmark, and test a calibrated probe gate with simulated clarification and comparison gates.
2.1 Probing for Unanswerability
- Frozen, layer-specific logistic-regression probes use prompt-token hidden states; cross-domain tests apply the training-domain scaler without test-side fitting.The study uses six open-weight models from four families, with multi-turn roll-outs limited to four 8–12B models.Source: §2.1 · p. 2
2.2 Domains and Datasets
- Six frozen datasets cover missing information in math, missing information in extractive contexts, and epistemic unknowability, with source labels and versioned, deduplicated sets.The datasets include SUM and UMWP; SQuAD 2.0 and MuSiQue; and SelfAware and KUQ.Source: §2.2 · p. 3
2.3 Transfer and Geometry
- Transfer and geometry analyses compare train-test domains and class-separation directions, with lexical residualization and training-only layer selection among the preregistered robustness checks.The analyses also include bootstrap confidence intervals, split-half null angles, and a whitened common-coordinate variant.Source: §2.3 · p. 3
2.4 A Turn-Labeled Multi-Turn Benchmark
- The benchmark labels each conversation prefix underspecified until its predefined sufficiency set is revealed, then answerable; its audited, repaired release contains 423 conversations and 1,661 turn-states.Constraints are revealed one per turn; generated conversations were audited by eight annotators and repaired under a fixed protocol.Source: §2.4 · p. 3
2.5 Probe-Gated Intervention
- A calibrated probe gate triggers abstention in single-turn tasks or a clarifying question in dialogue, where a simulated user returns an unrevealed constraint; oracle, position, and rate-matched random gates serve as references.The comparisons also include vanilla generation, prompted ask-when-unsure, prompted abstention, and RECAP-style consolidation.Source: §2.5 · pp. 3–4
3 Results
Probes transfer across datasets sharing an unanswerability ground, but whether known-unknowns form a distinct representation remains unresolved. In dialogue, single-turn probes fail zero-shot, and probe-gated clarification does not reliably improve end-task success.
Is known-unknowns represented separately?
- SelfAware and KUQ are both lexically separable (TF-IDF 0.94), while lexical residualization leaves within-dataset AUROC at 0.81–0.92 but weakens cross-ground separation.SelfAware→SUM rises above 0.65 in 3/6 models after residualization.Source: Is known-unknowns represented… · pp. 4–5
3.1 RQ1: Cross-Dataset Transfer
- Cross-ground transfer varies by model and layer (math↔SQuAD 0.38–0.94; math↔MuSiQue 0.17–0.90), leaving known-unknowns separation unresolved.Lexical controls, whitening, layer choice, or extractive-dataset changes remove the apparent separation in some models.Source: §3.1 · p. 4, Figure 1 · p. 5, Is known-unknowns represented… · p. 5
- 0.77–0.97 AUROC for transfer between the two math sets and 0.77–0.90 for SQuAD 2.0↔MuSiQue transfer show robust within-ground generalization.Both pairs transfer in both directions across all six models.Source: §3.1 · p. 4, Table 1 · p. 5
3.2 RQ2: Cross-Structure Transfer to Multi-Turn
- 0.20–0.55 AUROC at mid/late layers marks zero-shot failure; in-structure probes reach 0.925–0.935, but within-turn performance matches TF-IDF.Within-turn, TF-IDF reaches 0.79, on par with the probe ceiling averaged over turns 2–4 (0.79–0.81 across models).Source: Table 3 · p. 6, §3.2 · p. 6
3.3 RQ3: Probe-Gated Abstention and Clarification
- Probe-gated multi-turn success is mixed: +0.083 for Granite and +0.078 for Gemma, −0.071 for Qwen, and −0.144 versus consolidation for Qwen and Granite.The endpoint was fixed before gated runs were scored.Source: Table 4 · p. 7, §3.3 · p. 7
- Firing precision is 0.74–0.88 versus 0.54–0.58 for random, but the true-label gate is at most 0.08 higher in success and clarification use remains limiting.At the same firing rate, probe-gated asking beats randomly timed asking by +0.14 to +0.28 in every model.Source: §3.3 · p. 7
4 Related Work
Related work establishes linear probes for truthfulness and answerability, cross-task limits, and prompting-based abstention and dialogue workarounds. This paper varies the ground of unanswerability and follows the signal into turn-level dialogue and a deployed gate.
Probing internal states for answerability and truthfulness
- Truthfulness probes recover linear structure, while Liu et al. (2026) show answerability decodability within math and Slobodkin et al. (2023) report transfer across extractive benchmarks.Other work finds hidden-state error detectors do not transfer universally, and searches for a shared truthfulness hyperplane across tasks and domains.Source: §4 · pp. 7–8
- This paper varies the unanswerability ground—missing question information, missing context information, or epistemic unknowability—and follows the signal across dialogue and deployment.The related-work framing concludes that transfer is robust within a ground, while separation between grounds remains unresolved.Source: §4 · p. 8
Abstention and selective prediction
- Wen et al. (2024) frame abstention as a possible cross-task meta-capability; this paper reports transfer across shared grounds but leaves epistemic separation unsettled.The related-work passage identifies SUM, UMWP, SQuAD 2.0, SelfAware, and KUQ as examples of distinct grounds.Source: §4 · p. 8
Multi-turn underspecification and clarifying questions
- Laban et al. (2025) study recap and snowball consolidation; this paper adds labeled turn-states and a hidden-state gate, which does not convert precise firing into end-task gains.CLAM (Kuhn et al., 2022) and Zhang and Choi (2025) also study clarification with simulated users or expected benefit.Source: §4 · p. 8
Calibration and confidence
- The gate uses Platt scaling fit on training-side validation data, while label-shuffle, TF-IDF, and position controls test whether probe signals exceed confounds.The model itself is never fine-tuned.Source: §4 · p. 8
5 Conclusion
The paper finds that unanswerability probes transfer across datasets sharing an unanswerability ground, but do not transfer from single-turn inputs to dialogue states zero-shot. A leakage-free probe gate detects underspecified turns better than chance, yet does not reliably outperform prompting or consolidation.
Across datasets
- Probes transfer robustly when datasets share a ground of unanswerability, while whether epistemic unknowability is represented separately remains unresolved.The evidence on epistemic unknowability changes with layer, lexical controls, and coordinate system.Source: §5 · p. 8
Across structure
- Single-turn probes fail zero-shot on turn-level dialogue states; in-structure probes recover answerability no better than a classifier on user turns.The paper reports the in-structure comparison as matching TF-IDF.Source: §5 · p. 8
At deployment
- A leakage-free gate fires on underspecified turns more precisely than chance and comes within 0.08 of a true-label gate in end-task success, but still does not beat consolidation.The paper reports the benchmark and simulator as resources it releases alongside the result tables.Source: §5 · p. 8
Limitations
The paper’s evidence is bounded by its model coverage, synthetic dialogue setting, and choices in probe evaluation. The authors also identify limits on causal interpretation, annotation independence, and the known-unknowns benchmarks.
Scale and architecture coverage
- Only Qwen and Gemma appear in both size tiers, and the Qwen pair changes model generation, so the set is not a controlled scaling comparison.Layer indices are not comparable across architectures; layer selection is per-model.Source: Limitations · p. 9
Synthetic multi-turn data
- The multi-turn benchmark uses generated, human-audited conversations rather than naturally occurring dialogue, and its eight annotators are paper authors.The benchmark labels follow deterministic sufficiency-set rules, but the audit is not independent.Source: Limitations · p. 9
Simulated user
- The simulated user supplies scripted constraints without improvising or misunderstanding, and its validation covered 60 exchanges with one annotator.The authors note that real users would answer less cleanly, affecting ask-based conditions alike.Source: Limitations · p. 9
Correlational probes
- Probe results are decoding claims, not causal evidence, because the planned activation-steering ablation was not run.The paper therefore does not show that the probed direction causally mediates abstention.Source: Limitations · p. 9
Known-unknowns datasets
- Both known-unknowns benchmarks source answerable and unanswerable examples differently, confounding label with genre; lexical residualization cannot remove the issue.The authors identify a lexically matched known-unknowns set as necessary to address the confound.Source: Limitations · p. 9
Layer selection and multiple comparisons
- Best-layer sweeps expose results to selection effects, although the reported weak transfer cannot be created by the optimistic held-out-domain layer selection.The RQ2 in-structure ceiling is layer-insensitive across the evaluated models.Source: Limitations · p. 9
Ethical Considerations
The paper flags risks in using probe gates as safeguards and describes its data and AI-assistant practices. It also documents decision-rule timing and threshold sensitivity.
Ethical Considerations
- Because gates do not beat simple prompting baselines, the authors do not recommend deploying them as a safeguard.A gate that fires too often withholds needed answers; one that fires too rarely can give false assurance.Source: Ethical Considerations · p. 9
- The study uses public research datasets under their licenses or synthetic conversations without personal information; paper authors volunteered as annotators.AI assistants also helped with code, drafting, and grammar, while authors designed the study and take responsibility for its content.Source: Ethical Considerations · p. 9
A Decision Rules Committed Before Each Run
- Table 5 records rule commits and first-output timestamps, but the record establishes order rather than independence.The table notes that the headline was fixed in the project specification and that only one generation output had been inspected before the answering-user scoring commit.Source: Table 5 · p. 10, A Decision Rules Committed Before Each… · p. 10
B Threshold Sensitivity
- The single-turn threshold sweep compares correct behavior, abstention on unanswerable inputs, false abstention on answerable inputs, and answerable accuracy across nine threshold settings.For Qwen, correct behavior reaches 0.603 at τ=0.9 but does not reach the abstain prompt’s 0.636.Source: Table 6 · p. 11, B Threshold Sensitivity · p. 11
C Multi-Turn Dataset Datasheet
The datasheet describes how the synthetic multi-turn benchmark was constructed, audited, repaired, and run. It also reports checks on the robustness of layer selection and transfer results.
C Multi-Turn Dataset Datasheet
- The benchmark covers English arithmetic, small Python-function, and data-to-text conversations, with invented data-to-text records and deterministic labels from predefined sufficiency sets.Its initial construction generated 504 conversations, revealing atomic constraints one per user turn in randomized order.Source: C Multi-Turn Dataset Datasheet · p. 11
- The audit produced 728 verdicts from eight annotators and found a recurring defect: default-confirming constraints padded into sufficiency sets.Flags concentrated on math and code rather than data-to-text conversations.Source: C Multi-Turn Dataset Datasheet · p. 11
- A shrink-only repair by a judge outside the generator’s model family yielded a frozen set of 423 conversations and 1,661 turn-states.Conversations were dropped when flags could not be explained, decompositions conflicted, or the answerability flip depended on the final question utterance.Source: C Multi-Turn Dataset Datasheet · p. 11
- The eight benchmark annotators were paper authors, while model roll-outs used greedy decoding with a 384-token cap and recorded truncation rates.The paper estimates about 200 GPU-hours, with roughly 100 spent on the answering-user evaluation; code and probe software versions are pinned in the release.Source: C Multi-Turn Dataset Datasheet · p. 11
D Layer Selection and Multiple Comparisons
- The weakest headline within-turn ceiling survives Bonferroni correction, and the RQ2 in-structure ceiling ranges from 0.885–0.934 across all non-embedding layers.The reported layer-selection checks conclude that no layer choice drives that ceiling.Source: D Layer Selection and Multiple… · p. 11
- Every reported leave-one-domain-out transfer value is treated as an optimistic upper bound, because selecting a best layer per held-out domain can inflate transfer.That selection direction cannot create the weak transfer reported.Source: D Layer Selection and Multiple… · p. 12
E Additional Ablations and Controls
The controls show that conversational position carries substantial signal, while removing model replies improves transfer for Granite. Position and truncation controls qualify what can be inferred from the probes.
Position controls
- 0.934 AUROC is reached by relative position alone, while within-turn-index scores remain 0.70, 0.88, and 0.86 at turns 2, 3, and 4.A metadata model using conversation length reaches 0.89, but length is not observable from a turn-state prefix; observable metadata reaches 0.68 and user-turn TF-IDF reaches 0.79.Source: E Additional Ablations and Controls · p. 12
Excluded-replies read-out
- 0.829 vs 0.789 transfer AUROC for Granite follows removal of model replies, which also shrinks the single-turn-to-multiturn rotation from 81.2° to 78.6°.The authors note that the model's own reply tokens are part of what rotates the direction.Source: E Additional Ablations and Controls · p. 12
Shuffle nulls
- 0.486–0.496 ceiling label-shuffle nulls and 0.47–0.52 single-turn nulls remain near chance across the reported models and datasets.Truncation effects and infrastructure details are reported in released CSVs, with extraction dtype recorded in cache metadata.Source: E Additional Ablations and Controls · p. 12
F Worked Example
The worked example illustrates how the calibrated probe gate interacts with a user who supplies requested constraints. The associated sections describe pooled transfer evaluation and a cost-sensitive gate threshold.
F Worked Example
- 68–100% of the probe gate's premature answers across models occur when it fails to fire, as in the example's missed underspecified turn and premature $144 answer before correction to $108.At turns 1–2, the gate fires and the simulated user supplies the requested constraints in the order asked.Source: F Worked Example · p. 12
H Pooled and Source-Only Transfer
- Table 8 reports, on the four core datasets, a probe pooled over all but one domain and evaluated frozen on the held-out domain, alongside dissociation verdicts using each training dataset’s independently chosen layer.Source: H Pooled and Source-Only Transfer · p. 12
I RQ3: Scripted User and Cost-Sensitive Gate
- The cost-sensitive gate sets its threshold from extra-exchange cost κ=0.1, asking-on-sufficient-turn cost λ=0.7, and answerable accuracy.Its closed-form threshold is τ* = (κ + λ − 1 + acc)/(acc + λ).Source: I RQ3: Scripted User and… · p. 12
J RQ2: Reply Truncation
Reply truncation affects within-turn probe AUROC materially under the authors' threshold, with direction varying by model; it does not explain the zero-shot failure.
Reply truncation
- 16.6% of Qwen replies, 30.7% of Granite replies, 39.3% of Ministral replies, and 63.0% of Gemma replies hit the 384-token cap.Table 10 splits within-turn AUROC by whether an earlier reply in context was truncated.Source: J RQ2: Reply Truncation · pp. 12–13
- |∆| ≥0.05 makes reply truncation material within turns 2–4, with a model-specific sign, but it does not account for zero-shot failure.The zero-shot probe is at or below chance in mid/late layers.Source: J RQ2: Reply Truncation · p. 13
K Transfer Matrices on Six Datasets
The six-dataset transfer matrices show train-by-test AUROC across six models. The accompanying annotator instructions specify turn-label criteria and a review process for the multi-turn benchmark.
L Annotator Instructions
- Annotators checked whether turn labels switch from underspecified to answerable only when all essential constraints appear, and flagged mistimed labels or broken and leaking tasks.Eight annotators, all paper authors, received disjoint shards plus a shared calibration set; optional turns were not to change answerability.Source: L Annotator Instructions · pp. 13–14
K Transfer Matrices on Six Datasets
- Figure 3 presents train-by-test AUROC transfer matrices for all six models and six datasets, including MuSiQue and KUQ.The figure uses each model's six-dataset common best layer and reports means over seeds 13, 42, and 77, with color clipped at 0.5.Source: K Transfer Matrices on Six Datasets · p. 13, Figure 3 · p. 15