acceptodds
Under review as a conference paper at ICLR 2027

GNOSIS: Consistency-aware Probes for Selective Prediction over Internal States

Abstract

Selective prediction determines whether an LLM's output should be trusted or deferred, which is critical for building reliable LLM systems. Existing confidence estimators fall into three groups: multi-sample methods requiring repeated generation, post-generation scoring over fully decoded outputs, and pre-generation hidden-state probes. While pre-generation probes yield confidence estimation before token generation and avoid decoding overhead, existing approaches rely on a narrow readout paradigm: they typically preselect a single readout layer and optimize against only one training objective, limiting their overall performance. We present GNOSIS, a suite of lightweight single-pass pre-generation probes which operates on the prompt-final hidden state of a frozen LLM. It comprises three core designs: (1) The Direct Correctness Probe (DCP) infers answer correctness and adaptively selects the optimal readout layer via cross-validation instead of manual specification. (2) Anchor-Regression Distillation (ARD) distills multi-dimensional cross‑model consistency signals from a stronger teacher model into student hidden states, where the teacher is only used for target construction and is never invoked during inference. (3) A Fusion module jointly optimizes both probe heads on a shared trunk to exploit complementary error patterns. Across 11 LongBench tasks spanning eight base models (3B-72B), GNOSIS achieves competitive or state‑of‑the‑art AUROC and PRR. When deployed as an out‑of‑distribution HotpotQA cascade router, GNOSIS allows escalated queries to fully bypass student‑model generation, substantially lowering latency, token consumption, and API cost. Ablations confirm that DCP and ARD capture distinct failure modes and the joint multi‑objective Fusion outperforms naive post‑hoc score averaging. Our results demonstrate that adaptive readout-layer selection together with multi-faceted training targets is essential for high-performance pre-generation selective prediction.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, hallucination detection, selective prediction, uncertainty estimation, hidden-state probing, representation learning, model cascades, LLM agent routing

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_BVP2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies pre-generation reliability estimation for LLMs using the prompt-final hidden state of a frozen model, with applications to selective prediction and model cascade routing.

GNOSIS comprises three components. Direct Correctness Probing (DCP) predicts answer correctness and selects the readout layer through cross-validation. Anchor-Regression Distillation (ARD) predicts a seven-dimensional target constructed from teacher support, student stability, and auxiliary correctness information. Fusion jointly learns correctness and anchor targets through a shared network and combines their standardized outputs.

The evaluation covers eleven LongBench tasks and eight models, together with transfer from TriviaQA training to HotpotQA cascade routing. The results show competitive reliability ranking and deployment savings from avoiding student decoding on escalated queries. The relative benefits of DCP, ARD, and Fusion vary considerably across settings.

Critical strengths
  1. The work addresses a practically meaningful systems problem and measures deployment benefits.

    Making escalation decisions before student decoding can avoid generating answers that will subsequently be discarded. The paper evaluates this benefit beyond offline detection metrics: Table 2 reports 20.3% time savings and 13.1% generated-token savings for DCP-MLP at teacher-level accuracy, while Table 3 reports 17.1% teacher API cost savings. These measurements give the work practical relevance.

  2. The complementarity analysis provides useful evidence beyond aggregate performance comparisons.

    Figure 3 examines error concentrations, differences in anchor characteristics, and the incremental information in ARD after conditioning on DCP. The reported conditional AUROC improvement of 1.81 percentage points, with a 95% interval of [0.36,3.29], supports the presence of additional predictive information in the studied setting. This provides an empirical motivation for combining the objectives.

Critical weaknesses
  1. The independence of layer selection and final evaluation is unclear, potentially affecting the central empirical comparisons.

    Equation (1) selects the layer that maximizes AUROC over out-of-fold predictions, but the paper does not clearly describe a subsequent independent test set or an outer evaluation loop. If the tables report performance on the same OOF predictions used to select the layer, question-level cross-fitting prevents direct training contamination for each candidate probe but does not eliminate selection-induced optimism.

    This matters because each model–task cell contains only 200 questions, and adaptive layer selection is itself a central contribution. The statement that hidden-state methods receive layer selection also needs reconciliation with the claim that DCP-LR differs from P(IK) only in its selected layer. Without an explicit separation of selection and evaluation, together with uncertainty estimates for the main comparisons, it is difficult to assess whether the smaller improvements are robust.

  2. The incremental contribution over prior work, and the necessity of the teacher-anchor design, are not sufficiently isolated.

    The characterization of previous probes as using fixed layers and correctness-only supervision is too broad. SEPs already learn semantic entropy and select high-performing layer ranges; No Answer Needed uses cross-validation to select layers for predicting correctness before generation. The distinctive contribution therefore rests primarily on teacher anchoring and joint learning. SEPs, Appendix B.3, No Answer Needed, Section 4.3

    However, DCP-MLP outperforms Fusion in both deployment settings: Fusion has route AUROC 0.627 versus DCP’s 0.721, and lower two-tier AUC under both teachers. Figure 5 supports joint training in one setting, but does not isolate the contributions of the seven-dimensional structure, teacher agreement, student statistics, and gold-dependent auxiliary coordinates. The evidence supports complementarity in some settings, but does not yet establish the stronger claim that the proposed multi-faceted teacher targets are essential for high-performance pre-generation prediction.

Questions for authors
  1. Evaluation protocol clarification. Are the results in Tables 1 and 5 evaluated independently of layer selection? Please specify which samples are used for outer evaluation, inner layer selection, and standardization. If the same OOF predictions are currently used for selection and reporting, do the main advantages survive nested evaluation? Do P(IK), SAPLMA, and SEPs receive exactly the same layer-selection budget?
  2. Critical validation of contribution attribution. How does Fusion perform when using only the scalar anchor target, removing the gold-dependent cᵐᵃˣ and mᶜᵒʳʳ coordinates, or constructing auxiliary targets solely from student-side statistics? These controls would help determine whether the gains specifically require teacher-consistency information or arise from more general auxiliary supervision.
  3. Deployment protocol clarification. Equation (10) computes standard deviations on the evaluated split, making the relative weights of the two heads dependent on the test batch. Do the HotpotQA quality and cost benefits persist when normalization parameters and routing thresholds are fixed using training or independent validation data? How is the operating point used to report savings at teacher-level accuracy selected?
  4. Ablation setting clarification. Figure 5 reports Qwen-7B AUROC values of 0.694/0.672 for Fusion/DCP, whereas the simple averages across the eleven corresponding tasks in Table 5 are approximately 0.738/0.715. Which dataset, readout layer, and aggregation procedure are used in Figure 5? This distinction matters for assessing whether the ablation explains the main experimental gains.
Reviewer_FzEmOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces GNOSIS, a lightweight pre-generation probe for selective prediction in LLMs. It avoids the overhead of multi-sample generation. The framework consists of an adaptive readout-layer selection mechanism and two joint objectives, correctness and cross-model consistency. Evaluated across 11 LongBench tasks and multiple backbones, the fused model achieves reasonable good performance.

Critical strengths

Teacher consistency signals for LLM routing: The method amortizes cross-model consistency into student pre-generation states. It effectively tackles hard-to-detect hallucinations and with zero inference overhead.

Critical weaknesses

Over-claimed performance gain: the paper's major contributions of teacher-consistency distillation and dual-head fusion does not clearly beat baseline DCP-MLP, especially in the out-of-distribution (OOD) cascade setting. This degradation undermines the claim that teacher-anchor distillation yields robust, generalizable signals for real-world LLM routing.

Questions for authors
  1. Can authors provide more empirical study, showing the advantages of the proposed method over baselines?

  2. On LongBench, if you choose the readout layer on a separate validation split, do your gains still hold?

2 comments

We thank the reviewer for the comments. Following the reviewer's suggestions, we have added the experiments and analyses below.

Comparison with DCP-MLP

We respectfully clarify that DCP-MLP is not an external baseline: it is design (1) of the GNOSIS suite, which comprises three core designs, DCP, ARD, and Fusion (abstract and Section 3). The submission reports all three, and DCP-MLP and Fusion lead in different settings:

  • OOD cascade: DCP-MLP. As reported in Section 4.2, DCP-MLP gives the strongest route AUROC (0.721) under the TriviaQA-to-HotpotQA shift (Tables 2 and 3). The teacher-consistency distillation also transfers: ARD-MLP reaches 0.670, above every baseline except HaMI (0.679), a post-generation gate that needs the student's answer before scoring, whereas all GNOSIS gates score before decoding.
  • Same-task evaluation: Fusion. When the probes are trained and tested on the same task, Fusion improves over DCP-MLP by +1.19 AUROC points (95% CI [0.41, 1.99]) on the submission's 84 LongBench cells with the readout layer chosen on held-out folds (Q2), and by +2.70 [1.63, 3.80] on FRAMES (Q1); with the recent models on LongBench, the two are within 0.6 points. In Figure 3 (Qwen2.5-7B), ARD and DCP reject different errors: those rejected only by ARD have higher student-only stability than those rejected only by DCP, and adding ARD to a cross-fitted DCP stack yields +1.81 [0.36, 3.29], while shuffling ARD within DCP deciles removes the gain. Q1 evaluates all methods on consistent questions, where stable yet incorrect answers occur.

We will state in the abstract and Section 4.2 that DCP-MLP is the strongest gate in route AUROC under the TriviaQA-to-HotpotQA shift, and Fusion the strongest probe in same-task evaluation.

Q1. More empirical study of the advantages over baselines

We added two recent models, Gemma-4-31B and Qwen3.8-27B, and FRAMES (Krishna et al., NAACL 2025), a recent and challenging multi-hop QA benchmark. In total, our evaluation now covers 12 datasets across five task types, 10 models from four model families (3B to 72B parameters), two cascade teachers including GPT-5.5, and 14 baselines. We compared GNOSIS with the 11 baselines of Table 1 and PTrue-greedy under the nested evaluation of Q2, first on consistent questions and then on all questions.

(1) Consistent questions. We call a question consistent when the student's K = 4 sampled answers fall into one semantic cluster; errors on these questions are the stable yet incorrect answers targeted in Section 1. Table A reports the share of errors on consistent questions and the AUROC of each method on them.

Table A. Consistent questions (number of model–task cells in parentheses). First row: share of all errors that occur on consistent questions. Other rows: AUROC on consistent questions (mean over cells and 3 seeds) with nested evaluation.

MethodSubmission: LongBench (84)Recent models: LongBench (18)Recent models: FRAMES (2)
Share of errors87.2%87.5%65.8%
GNOSIS (ours)Fusion0.7040.7600.774
DCP-MLP0.6930.7530.743
ARD-MLP0.6520.7200.755
ARD-Ridge0.6540.6940.734
Pre-gen.P(IK) (= DCP-LR with selection)0.6900.7540.737
SAPLMA†0.6370.6910.745
SEPs-Ridge0.5320.5700.587
Post-gen.MSP-mean0.6700.6540.673
MSP-min0.6560.6480.672
PTrue-greedy§0.6070.6410.679
Multi-samplePTrue-trained§0.6400.6500.716
PTrue-Kbrain§0.6470.6400.660
LexSim0.6890.6300.645
SelfCk-Ngram1-avg0.6780.6400.598
SelfCk-NLI-avg§0.5670.5440.504
SemanticEntropy‡0.5000.5000.500

Groups follow the submission. †As in the submission, SAPLMA is trained on the other tasks of the same model, with its layer selected on those tasks (submission's cells: 1 seed). ‡Semantic entropy is zero on every consistent question. §As in the submission, the PTrue scores are unavailable for GPT-OSS-20B and SelfCk-NLI-avg for Gemma-2-27B; the first column averages the remaining 73 and 75 cells.

Consistent questions contain most errors: 87% on LongBench and 66% on FRAMES. On these questions, semantic entropy is zero and cannot rank them; the other multi-sample methods stay below 0.72 AUROC, the post-generation methods below 0.68, and SEPs-Ridge below 0.59. Fusion has the highest AUROC in all three settings, and all its gains over the baselines have 95% confidence intervals above zero (paired bootstrap over questions), except over P(IK) on LongBench with the recent models (within 0.6 points). GNOSIS obtains this from the prompt-final hidden state before decoding, whereas the multi-sample and post-generation methods need decoded answers.

Q1. More empirical study of the advantages over baselines (continued)

(2) All questions. Table B reports all methods on all questions with the recent models.

Table B. Selective prediction with the recent models: AUROC (mean over model–task cells and 3 seeds).

MethodLongBenchFRAMES
GNOSIS (ours)Fusion0.7600.802
DCP-MLP0.7550.775
ARD-MLP0.7230.794
ARD-Ridge0.6980.772
Pre-gen.P(IK) (= DCP-LR with selection)0.7570.767
SAPLMA†0.6940.774
SEPs-Ridge0.5770.646
Post-gen.MSP-mean0.6710.740
MSP-min0.6640.738
PTrue-greedy0.6510.711
Multi-samplePTrue-trained0.6600.759
PTrue-Kbrain0.6500.694
LexSim0.6580.727
SelfCk-Ngram1-avg0.6510.653
SelfCk-NLI-avg0.5730.612
SemanticEntropy0.5560.645

LongBench: 18 cells, following the setting of the submission. FRAMES: 2 cells; each question is given passages from its Wikipedia pages in the LongBench multi-document format. Groups follow the submission. †As in Table A.

Fusion achieves the highest mean AUROC on both benchmarks. On FRAMES, it improves over every baseline by 2.8 to 19.0 AUROC points; on LongBench, it improves over every baseline except P(IK) by 6.6 to 20.4 points. All these gains have 95% confidence intervals above zero (paired bootstrap over questions).

Q2. Readout layer chosen on a separate validation split

Yes. We ran nested evaluation, as in our response to Reviewer BVP2: in each of 5 outer folds, the readout layer is selected by 3-fold inner cross-validation on the outer-training questions only, and the probe is evaluated on the held-out outer-test questions. Table C also reports every probe at a fixed layer, without selection.

Table C. LongBench AUROC (mean over 84 model–task cells and 3 seeds; SAPLMA: 1 seed). All scores are computed on outer-test questions. Fixed layer: the deepest candidate layer.

ProbeNested selectionFixed layer
Fusion (ours)0.7030.703
DCP-MLP (ours)0.6910.696
P(IK) (= DCP-LR with selection)0.6890.697
ARD-Ridge (ours)0.6550.654
SAPLMA†0.6380.625
SEPs-Ridge0.5350.547

†As in the submission, SAPLMA is trained on the other tasks of the same model; its layer is selected on those tasks, and its fixed layer is the second-deepest candidate.

Layer selection affects only the hidden-state probes; among them, Fusion has the highest mean AUROC under both protocols. Under nested evaluation, it improves over DCP-MLP by +1.19 AUROC points (95% CI [0.41, 1.99]) and over P(IK) with the same layer-selection budget by +1.41 [0.65, 2.13] (paired bootstrap over questions). All results with the recent models in Q1 use the same nested evaluation. Therefore, with the readout layer chosen on a separate validation split, Fusion's gains over DCP-MLP and P(IK) still hold.

We are grateful for the reviewer's constructive comments, which have been very helpful for our work; we will revise the paper accordingly.

Reviewer_pNRrOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary
  • This paper proposes GNOSIS, a pre-generation selective method for accepting target model’s generation. GNOSIS employs the prediction of target model failure from prompt latent embeddings as well as alignment with teacher, and fuse them, which outperform existing selective methods in predicting failures.
Critical strengths

This paper tackles the important blind-spot where the student model gives out consistent (certain) failure. The proposed method directly tackles this issue, outperforming existing baselines in LongBench across different models.

Critical weaknesses
  • The models / datasetes used in evaluation are quite out-dated: Qwen2.5, Gemma-2, they should be replaced with recent counterparts, and the tasks are quite old as well (TriviaQA and HotpotQA). Is there any specific reason why authors only tested with those old evaluation suites?
  • I’m not convinced how significant the main bottlenecks this paper tries to mitigate. Are existing baseline catastrophically fail without the proposed method? If so, does the overall experiments in this paper highlight that? If so, the overall writings needs revision to make the contributions more clear and impactful, and how the experimental setups highlight that.
Questions for authors

Please see weaknesses.

2 comments

We thank the reviewer for the comments. Following the reviewer's suggestions, we have added the experiments and analyses below.

Q1. Recent models and datasets

We added two recent models, Gemma-4-31B and Qwen3.8-27B, and FRAMES (Krishna et al., NAACL 2025), a recent and challenging multi-hop QA benchmark. In total, our evaluation now covers 12 datasets across five task types, 10 models from four model families (3B to 72B parameters), two cascade teachers including GPT-5.5, and 14 baselines. Table A reports the new results under the protocol of the submission, with nested evaluation as in our response to Reviewer BVP2.

Table A. Selective prediction with the recent models: AUROC (mean over model–task cells and 3 seeds).

MethodLongBenchFRAMES
GNOSIS (ours)Fusion0.7600.802
DCP-MLP0.7550.775
ARD-MLP0.7230.794
ARD-Ridge0.6980.772
Pre-gen.P(IK) (= DCP-LR with selection)0.7570.767
SAPLMA†0.6940.774
SEPs-Ridge0.5770.646
Post-gen.MSP-mean0.6710.740
MSP-min0.6640.738
PTrue-greedy0.6510.711
Multi-samplePTrue-trained0.6600.759
PTrue-Kbrain0.6500.694
LexSim0.6580.727
SelfCk-Ngram1-avg0.6510.653
SelfCk-NLI-avg0.5730.612
SemanticEntropy0.5560.645

LongBench: 18 cells, following the setting of the submission. FRAMES: 2 cells; each question is given passages from its Wikipedia pages in the LongBench multi-document format. Groups follow the submission. †As in the submission, SAPLMA is trained on the other tasks of the same model, with its layer selected on those tasks.

Fusion achieves the highest mean AUROC on both benchmarks. On FRAMES, it improves over every baseline by 2.8 to 19.0 AUROC points, and over DCP-MLP by 2.7. On LongBench, it improves over every baseline except P(IK) by 6.6 to 20.4 points; P(IK) and DCP-MLP are within 0.6 points of Fusion. All these gains have 95% confidence intervals above zero (paired bootstrap over questions). We will add Table A to Section 4.

Q2. Significance of the main bottleneck

The bottleneck is the failure mode stated in Section 1: the student produces stable yet incorrect answers. As noted in Appendix A, an answer that the student reproduces across samples has low semantic entropy and looks reliable to self-consistency checks, whether or not it is correct. We call a question consistent when the student's K = 4 sampled answers fall into one semantic cluster, the clustering used for semantic entropy and for the student-only stability (Appendix A). Table B reports the share of errors that occur on consistent questions and the AUROC of each method on these questions.

Table B. Consistent questions (number of model–task cells in parentheses). First row: share of all errors that occur on consistent questions. Other rows: AUROC on consistent questions (mean over cells and 3 seeds) with nested evaluation.

MethodSubmission: LongBench (84)Recent models: LongBench (18)Recent models: FRAMES (2)
Share of errors87.2%87.5%65.8%
GNOSIS (ours)Fusion0.7040.7600.774
DCP-MLP0.6930.7530.743
ARD-MLP0.6520.7200.755
ARD-Ridge0.6540.6940.734
Pre-gen.P(IK) (= DCP-LR with selection)0.6900.7540.737
SAPLMA†0.6370.6910.745
SEPs-Ridge0.5320.5700.587
Post-gen.MSP-mean0.6700.6540.673
MSP-min0.6560.6480.672
PTrue-greedy§0.6070.6410.679
Multi-samplePTrue-trained§0.6400.6500.716
PTrue-Kbrain§0.6470.6400.660
LexSim0.6890.6300.645
SelfCk-Ngram1-avg0.6780.6400.598
SelfCk-NLI-avg§0.5670.5440.504
SemanticEntropy‡0.5000.5000.500

For the submission's cells, the predictions are those of the nested evaluation in our response to Reviewer BVP2. Groups follow the submission. †As in Table A (submission: 1 seed). ‡Semantic entropy is zero on every consistent question. §As in the submission, the PTrue scores are unavailable for GPT-OSS-20B and SelfCk-NLI-avg for Gemma-2-27B; the first column averages the remaining 73 and 75 cells.

Consistent questions contain most errors: 87% on LongBench and 66% on FRAMES. On these questions:

  • Multi-sample methods. The K samples agree, so semantic entropy is zero and cannot rank these questions. SelfCk-NLI-avg stays below 0.57 AUROC, and the other multi-sample methods below 0.72.
  • Post-generation methods. MSP-mean, MSP-min, and PTrue-greedy, which need the decoded answer, stay below 0.68.
  • Pre-generation probes. SEPs-Ridge, which is trained to predict semantic entropy, stays below 0.59, and SAPLMA below 0.75.

Fusion has the highest AUROC in all three settings, and all its gains over the baselines have 95% confidence intervals above zero (paired bootstrap over questions), except over P(IK) on LongBench with the recent models (within 0.6 points). GNOSIS obtains this from the prompt-final hidden state before decoding, whereas the multi-sample and post-generation methods need decoded answers.

In the submission, Figure 3(b) shows that errors rejected only by ARD have higher student-only stability than errors rejected only by DCP. We will add the analysis above to Section 4 and refer to it in Section 1, so that the size of this failure mode and the behavior of each method on it are stated explicitly.

We are grateful for the reviewer's constructive comments, which have been very helpful for our work; we will revise the paper accordingly.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of responses and request for feedback

Dear Area Chair and Reviewers pNRr, FzEm, and BVP2,

As the author–reviewer discussion closes on November 18 and we have not yet received feedback, we briefly summarize the new results in our responses. Each addresses a question raised in the reviews.

  • Nested evaluation (BVP2 Q1, FzEm Q2). The readout layer is now selected on outer-training questions only (84 LongBench cells, 3 seeds). Fusion has the highest mean AUROC among the probes and improves over DCP-MLP by +1.19 AUROC points (95% CI [0.41, 1.99]) and over P(IK) with the same layer-selection budget by +1.41 [0.65, 2.13].
  • Contribution attribution (BVP2 Q2). With the architecture, labels, folds, and initialization fixed, the gain over DCP-MLP is retained with the scalar anchor target alone (+2.18 [0.65, 3.73]) and without the gold-dependent coordinates (+1.84 [0.30, 3.31]); it comes from the anchor readout rather than from general auxiliary supervision.
  • Deployment protocol (BVP2 Q3). With the normalization statistics and the routing threshold fixed before any test question is scored, the cascade savings persist.
  • Recent models and datasets (pNRr Q1, FzEm Q1). We added Gemma-4-31B, Qwen3.8-27B, and FRAMES. With these models, Fusion has the highest mean AUROC among 16 methods on both LongBench and FRAMES; on FRAMES, it improves over every baseline by 2.8 to 19.0 AUROC points.
  • Consistent failures (pNRr Q2, FzEm Q1). Questions on which all sampled answers agree contain 87% of the errors on LongBench. On these questions, semantic entropy is zero, the other multi-sample and post-generation methods stay below 0.72 AUROC, and Fusion has the highest AUROC in every setting.
  • Comparison with DCP-MLP (FzEm). DCP-MLP is design (1) of GNOSIS. It is the strongest gate in route AUROC under the OOD cascade shift (Section 4.2), and Fusion is the strongest probe in same-task evaluation.

In total, the evaluation now covers 12 datasets, 10 models from four families (3B to 72B parameters), two cascade teachers including GPT-5.5, and 14 baselines.

We would be grateful if each reviewer could let us know whether our responses address their concerns. If any concern remains, please let us know and we will provide further analyses before the discussion closes. If the concerns have been addressed, we would appreciate it if the reviewers could consider updating their scores accordingly.

Thank you for your time.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.