acceptodds
Under review as a conference paper at ICLR 2027

Isolating a Language Model’s Judgment with a Label-Free Subspace

Abstract

Language models are often asked to make judgments: whether a proposed answer is correct, whether one statement follows logically from another, or whether a response is safe. Yet changing the instructions that specify how the judgment should be expressed can make the final answer less reliable, even for the same input. We ask whether this reflects a change in internal judgment or a failure to express it. Comparing hidden-state vectors for the same input under different instructions, we find that their differences concentrate in a low-dimensional expression subspace. We identify this subspace without benchmark labels and remove its contribution from a linear predictor to obtain an isolated judgment. Across correctness, entailment, and safety, the isolated judgment remains reliable under instructions not used to identify the subspace, even when the model's final answer and the original predictor degrade. When the judgment question is negated, isolation raises the predictor's AUROC from 0.504 to 0.825 for entailment and from 0.480 to 0.891 for safety. We also find that judgment is readable in earlier layers, before the expression subspace begins to disrupt the predictor. Together, these results provide a way to separate judgment from expression and recover internal judgments even when the model's final answers are unreliable.

Primary area:
interpretability and explainable AI
Keywords:
interpretability, probing, linear probes, representation geometry, prompt sensitivity, large language models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_FCQLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether LLMs represent correct judgments even when their final answers are incorrect due to response instructions. They identify a low-dimensional“expression subspace” to recover more stable judgments. Experiments across correctness, entailment, and safety tasks show improved probe AUROC, suggesting that internal judgments may be more stable than the model’s expressed answers.

Critical strengths
  • Demonstrating that verdict is consistently encoded in a single stable subspace is a potentially interesting result
  • I appreciate the experimentation across multiple model, many datasets, and multiple prompt formats
  • The paper makes clear claims and the evidence and argument made is understandable
  • Causal evidence is convincing
Critical weaknesses
  • I have concerns about the meaningfulness of this result, because as I understand it, the main novelty here is that you can find a single subspace that consistently across formats encodes a verdict. But even under the very constrained, strict, and simple formatting that you use we already see that this result fails to generalize on ANLI, Qwen2.5 negated question. I am concerned that even despite testing on multiple formats, the formats that you use are substantively similar enough that the expression subspace is simply an artifact of the lexical similarity of these formats. In fact, I worry this is critical confound.
  • This result also only extends to binary, objective judgements which meaningfully reduces significance of the result
Questions for authors
  • What is the difference in accuracy between this subspace and a no COT response to the same question/format?
  • If you are interested in understanding the model's verdict why do you only report AUROC to predict the correct answer? Why not predict the model's answer?
2 comments
AuthorsOpenReview ↗

Response to Reviewer FCQL

We thank the reviewer for their thoughtful feedback. We have updated the code and uploaded a revised manuscript, including new experiments and analyses in the appendix.

W1. ANLI performance and generalization beyond lexical similarity.
We thank the reviewer for highlighting the ANLI result and the concern about lexical confounds. The cited ANLI result is in fact a particularly striking example of the phenomenon our work aims to characterize. On Qwen2.5-7B, the model's expressed answer, original readout, and isolated readout achieve 0.398, 0.771, and 0.882 AUROC, respectively. This contrast illustrates our central finding: the model's expression can become unreliable even when its judgment signal remains recoverable.

To directly examine lexical generalization, we tested 20 new ways of asking whether an answer is incorrect (e.g., "contains an error" or "fails to match"), none of which appeared during subspace estimation. Isolation improves AUROC from 0.662 to 0.785, comparable to a wording-specific readout (0.783), with improvements across all 20 wordings and 11 model–dataset settings (Section 5.3 and Appendix C.5).

We further tested generalization to entirely unseen response mappings. We estimated the expression subspace using only non-reversing response-format variations, excluding all negation and response-reversal examples across every task. We then evaluated four transformation types absent from subspace estimation: negation, No-means-correct, False-means-correct, and 0-means-correct. Despite never observing these transformations during estimation, isolation improves AUROC by +0.096 (14/14 settings), +0.106 (14/14), +0.082 (13/14), and +0.021 (10/14), respectively (Section 5.2 and Appendix C.2, Table 7).

Together, these results show that the identified expression subspace generalizes beyond lexical similarity, transferring both to unfamiliar wordings and to response mappings entirely absent during its estimation.

W2. Generality beyond binary judgments.
We thank the reviewer for encouraging evaluation beyond binary, objective judgments. We examined both broader judgment domains and more flexible forms of expression. Across ETHICS, BBQ, and ToxiGen, covering moral, socially sensitive, and safety-related judgments, we evaluated task-specific readouts across multiple response mappings.

Expression isolation improves AUROC in all 12 held-out dataset–format combinations (Section 5.2 and Appendix C.6, Table 18):

DatasetNegationNo-means-correctFalse-means-correct0-means-correctReference
ETHICS+0.585+0.484+0.290+0.243−0.006
BBQ+0.136+0.105+0.071+0.042+0.015
ToxiGen+0.672+0.572+0.366+0.312−0.002

We also tested explanation-then-answer responses, where models explain their assessments before providing a final answer (e.g., "The calculation contains an error because ... Verdict: No."). We apply a readout trained without explanations to the hidden state immediately before the final answer, after the explanation. Under negation, isolation improves AUROC from 0.445 to 0.702, exceeding the model's expressed assessment (0.606; Section 6.3 and Appendix D.5).

Finally, we evaluated free-form critiques, where models assess candidate answers in their own words without prescribed binary response options (e.g., "The reasoning is sound, but the final calculation is incorrect because ..."). The expressed assessment is read by asking the same model whether its critique says the answer is correct (Yes, No, or Unclear). Across four model–dataset settings, our readout measured before critique generation predicts correctness more reliably than the expressed assessment (0.844–0.933 vs. 0.549–0.781 AUROC; Section 6.3 and Appendix D.5, Table 29).

Together, these results show that recoverable judgment information extends beyond purely factual judgments and fixed binary response formats, encompassing human-annotated moral assessments, explanations, and open-ended critiques.

Q1. Accuracy compared with direct responses without chain-of-thought.
We thank the reviewer for requesting a direct comparison with responses generated without chain-of-thought. On matched examples, our isolated readout achieves 76.9% accuracy (71.4% balanced accuracy), compared with 61.8% (58.1% balanced accuracy) for the model's direct response (Section 5.2 and Appendix C.9, Table 25). The readout threshold is calibrated using unlabeled target scores, without access to target correctness labels.

These results show that the model's internal judgment signal can be more reliably recovered than its expressed answer, even without chain-of-thought generation.

AuthorsOpenReview ↗

Response to Reviewer FCQL

Q2. Why not predict the model’s own answer?

We thank the reviewer for raising this distinction. We use benchmark correctness as a fixed reference across instructions to test whether the hidden state retains information about whether a candidate answer is correct. We shared the reviewer’s concern and had also examined the model's own verdict in the original submission (now Section 6.2 and Appendix D.3). Specifically, we train a readout using the model’s verdict under the ordinary Yes/No question as the label.

On held-out questions, we apply this readout after expression-subspace removal to hidden states from the negated prompt. Across 13 model–dataset settings, the isolated readout predicts the ordinary verdict with a median AUROC of 0.939, whereas its AUROC against the verdict expressed under negation is 0.667 (Section 6.2 and Appendix D.3, Table 27). Even on examples where the two responses imply opposite assessments of the candidate answer, the ordinary verdict remains predictable (0.930 AUROC across 12 settings; Table 27). Thus, the hidden state under negation retains information about how the model evaluated the same candidate under the ordinary question.

Reviewer_z2ptOpenReview ↗
Rating
3(+1)
Confidence
1(−1)
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

This paper studies how judgment and response instructions affect linear readouts that predict answer correctness from a language model’s hidden states. The authors find that changing these instructions—for example, requiring “No” to indicate a correct answer—can reduce the performance of a previously fitted readout, even though the candidate answer and its correctness remain unchanged.

They propose identifying a low-dimensional subspace associated with these instruction changes and removing its contribution from the readout. Across multiple models and datasets, the adjusted readout generalizes better to unseen response instructions, performing close to readouts trained specifically for those instructions. The authors interpret this improved robustness as evidence for separating information about the model’s judgment from information about how it expresses that judgment.

Critical strengths
  1. The paper investigates how instruction changes affect the reliability of internal readouts. Holding the candidate answer fixed while varying the judgment question or response mapping provides a useful way to study sensitivity to these instructions.

  2. I found the paper clear and easy to understand.

  3. The evaluation covers multiple models, datasets, and instruction wordings, with additional experiments on entailment and safety. Comparisons with random subspaces, condition-identity directions, pooled readouts, and negation-only subspaces help assess which aspects of the construction contribute to its performance.

Critical weaknesses
  • Upon inspecting the code, I found discrepancies between what is written in the paper and the implmentation, which affect the central subspace-estimation procedure. Appendix C.1 describes rank selection using principal angles between subspaces, whereas the code compares corresponding singular vectors, which is not equivalent. The implementation also centers the concatenated differences globally rather than centering each set separately. Additionally, the internal stability procedure independently partitions the difference blocks, allowing the same underlying question to appear in both halves across blocks; this raises concerns about the independence of the stability estimates. These issues can affect the selected rank and subspace, although their empirical impact remains unknown. The supplied archive also lacks required inputs, including candidate-answer files, prepared task inputs, and regraded TriviaQA labels, without a complete procedure for regenerating them. This prevented me from assessing whether reconciling the implementation with the manuscript would materially change the reported findings.

  • The generality and added value of the projection remain unclear. The proposed projection is primarily evaluated with class-mean readouts. Although other probe estimators are included as baselines, these comparisons do not establish whether applying the same projection benefits those estimators. It therefore remains unclear whether the method’s benefits extend across readout estimators or depend substantially on the chosen class-mean construction. Applying the projection to additional linear readouts under matched training and evaluation conditions would strengthen the evidence for its broader applicability.

Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer z2pt

We thank the reviewer for their thoughtful feedback. We have updated the code and uploaded a revised manuscript, including new experiments and analyses in the appendix.

W1. Implementation consistency and reproducibility.
We appreciate the reviewer's careful code inspection. During submission preparation, an earlier version of the code for certain subspace-estimation details and associated stability analyses was inadvertently retained alongside the updated methodological description. We sincerely apologize for the resulting confusion.

The reviewer’s observations prompted us to compare the two implementations directly and to include additional sensitivity analyses in the revised appendix. We reran the analysis using the procedure described in Appendix C.1 and compared the resulting AUROCs with those obtained using the code released with our submission. Averaged over the fourteen cells, the largest AUROC change on any held-out condition was 0.012 (Appendix C.1, Table 4). The revised appendix reports sensitivity analyses for rank selection, centering, and the use of question-disjoint splits to assess subspace stability. When we varied the subspace rank from 4 to 12, mean AUROC ranged from 0.797 to 0.806 (Appendix C.1).

For reproducibility, the revised supplementary package includes the prepared inputs and regeneration instructions.

W2. Generality across readout estimators.
We thank the reviewer for suggesting an evaluation beyond class-mean readouts. We applied the same label-free expression-subspace projection to logistic regression, ridge regression, linear SVM, and a one-hidden-layer MLP, training each estimator on projected hidden states under matched conditions.

Across 14 model–dataset settings, isolation improves AUROC across all four estimators (Section 5.3 and Appendix C.5, Table 16):

Evaluation formatLogisticRidgeLinear SVMMLP
Negated judgment+0.075+0.085+0.092+0.134
False-means-correct+0.055+0.051+0.061+0.058
No-means-correct+0.083+0.085+0.090+0.090
0-means-correct+0.036+0.034+0.031+0.115
Reference format+0.001+0.002+0.001+0.011

Performance on the reference format remains largely unchanged.

These results demonstrate that the benefits of expression isolation extend beyond class-mean readouts to multiple linear estimators and a nonlinear MLP, without requiring labeled representations from the evaluated expression formats.

I thank the authors for their rebuttal.

Regarding the implementation, during the reviewing period, I invested considerable time in checking whether the code accurately reflected what was described in the paper. In my assessment, the implementation errors were more substantial than just "details and sensitivity analyses." I have not had time to re-examine the code thoroughly following the updates that the authors report having made, and I believe it is important to ensure the correctness of the implementation at the time of submission.

That said, based on the authors' response, I consider my concerns to have been addressed. I am therefore happy to increase my overall score. However, I have decided to lower my confidence score to reflect the fact that I have not independently verified the correctness of the updated implementation.

AuthorsOpenReview ↗

Response to Reviewer z2pt

Thank you for the updated assessment. We are grateful for the time you spent checking our code and the paper. That check found parts we had missed, and we take your point about correctness seriously.

For the updated code and manuscript, we went through the code completely and confirmed that it implements exactly the procedure the paper describes. We will keep the released code and the paper consistent in the final version, and we would welcome any further issue you find.

Reviewer_3StzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether language models preserve an internal judgment when changes in the instructions lead to unreliable final verdicts. The authors propose estimating a low-dimensional “expression subspace” using SVD on paired hidden-state differences obtained by varying the instructions while keeping the evaluated input fixed. They then remove the component of a supervised linear readout lying in this subspace.

The proposed method substantially improves the robustness of linear readouts under negated questions and reversed response mappings. The experiments cover correctness, entailment, and safety judgments across multiple language models and datasets. The paper also provides layer-wise analyses suggesting that judgment-related information becomes linearly accessible before instruction-dependent interference emerges in later layers.

The empirical findings are interesting and generally welly supported. However, I have some concerns about whether the results justify the claimed separation between judgment and expression, the novelty relative to existing representation-based approaches, and the strength of the generalization evaluation.

Critical strengths
  1. Interesting and relevant research question.

The distinction between a model’s internal assessment and its expressed verdict is important for understanding the reliability of language-model evaluations. The paper provides a useful experimental framework for examining this distinction.

  1. Strong empirical results under the evaluated instruction changes.

The improvement over the frozen source readout is substantial. On the negated correctness question, AUROC improves from 0.650 to 0.802. Similar improvements are reported for entailment (0.504 to 0.827) and safety (0.480 to 0.891). These results demonstrate that the proposed projection provides an effective means of improving readout robustness.

  1. Relatively comprehensive experimental controls.

The authors evaluate multiple model families, response formats, and judgment tasks. The appendix includes controls involving random subspaces, PCA, LEACE/INLP, alternative instruction wordings, and subspace rank. The question-disjoint evaluation protocol and additional prospective models strengthen the empirical evidence.

  1. Clear presentation and useful layer-wise analysis.

The methodology and experimental results are generally well organized. The layer-wise experiments provide an interesting observation: instruction-dependent degradation becomes pronounced in later layers, while earlier layers retain relatively stable linear predictability.

Critical weaknesses

W1. Insufficient evidence for judgment–expression disentanglement.

The experiments demonstrate improved readout robustness, but do not establish that the identified subspace specifically represents expression rather than instruction-dependent nuisance variation. Paired hidden-state differences may also capture changes in task interpretation or judgment formation. Without targeted causal interventions, the claimed separation between judgment and expression remains insufficiently supported.

W2. Limited generalization to genuinely unseen transformations.

Although the evaluation conditions are held out, their underlying transformations are not entirely unseen. In particular, negation pairs from other tasks are used during subspace estimation. Table 11 shows that these pairs account for nearly all the improvement under negation (0.801 vs. 0.802 AUROC with the full subspace). This primarily demonstrates cross-task transfer of known transformations rather than generalization to novel instruction changes.

W3. Limited novelty and incremental gains over simpler alternatives.

The method combines established techniques (paired activation differences, SVD, and linear projection). Moreover, the supervised pooled baseline achieves comparable performance (0.794 vs. 0.802 AUROC under negation), while a three-dimensional negation-derived subspace recovers nearly the full improvement. The additional value of the complete method over these simpler alternatives requires clearer justification.

W4. Restricted evaluation scope and incomplete robustness.

The experiments focus on binary judgments with short, structured responses. Furthermore, transfer to unseen instruction wordings shows noticeable degradation in some settings (Table 15). It remains unclear whether the proposed separation extends to more diverse instructions, open-ended judgments, or longer generations.

This paper offers a thoughtful experimental investigation of instruction-dependent failures in language-model judgments. The empirical results are substantial, and the presentation and experimental controls are stronger than those of many submissions.

Nevertheless, the central scientific claim is not yet fully supported. The experiments convincingly establish that removing a low-dimensional, instruction-associated subspace can improve the robustness of linear readouts. They do not conclusively demonstrate that this subspace corresponds to expression independently of judgment.

Furthermore, the held-out evaluations reuse transformation types observed during subspace estimation on other tasks, and the gains over simpler alternatives are relatively limited.

I therefore lean toward rejection at the current stage. A more precise framing around robust readout recovery, or stronger causal and genuinely out-of-transformation validation of the proposed decomposition, would significantly strengthen the contribution.

Questions for authors

Q1. What evidence distinguishes an actual judgment–expression decomposition from the simpler explanation that the method removes instruction-dependent nuisance directions that interfere with a supervised linear probe?

Q2. How well does the method generalize to a transformation type that is entirely absent from subspace estimation across all tasks, rather than merely held out within the target task?

Q3. Can the authors provide a causal intervention on the expression subspace itself, demonstrating a change in expression while preserving the independently measured judgment?

Q4. Given the strong performance of the pooled supervised baseline and the three-dimensional negation-only subspace, what is the specific practical advantage of the complete method under matched data-collection and computational budgets?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 3Stz

We thank the reviewer for their thoughtful feedback. We have updated the code and uploaded a revised manuscript, including new experiments and analyses in the appendix.

W1 / Q1 / Q3. Causal separation of judgment and expression.

We appreciate the reviewer's concern that improved readout robustness alone does not establish a separation between judgment and expression. To directly address this, we conducted targeted causal interventions across 14 model–dataset settings. For the same question and candidate answer, we obtained hidden states under ordinary and negated judgment instructions, then replaced only the component within our identified expression subspace with that from the opposite instruction, leaving the remaining representation unchanged.

Swapping only this component changes the model's answer in 65.9% of cases, nearly matching full hidden-state replacement (67.0%), while equal-displacement complementary-subspace and random-subspace controls yield only 3.1% and 1.9%, respectively (Section 6.1 and Appendix D.4, Table 28). The two swaps also change the same answers: the expression-only swap changes 98.0% of the answers that full-state replacement changes and 0.5% of those it leaves unchanged, and the two post-swap answers agree in 98.5% of cases (33.4% and 33.6% for the complementary and random controls; Table 28). Crucially, a separately constructed, unprojected correctness readout trained on unpatched states remains predictive after the intervention (AUROC 0.736 → 0.800), including on the answers that the swap changed (0.715 → 0.791; Table 28). The reverse intervention yields consistent results, changing 63.7% of responses while correctness prediction remains strong (0.763 → 0.797; 0.741 → 0.780 on the changed responses; Table 28).

These findings provide causal evidence that the identified subspace directly influences the model's answer, rather than merely capturing variation correlated with probe degradation. At the same time, correctness information remains independently recoverable despite substantial changes in expression. We therefore interpret our findings as evidence for a functional dissociation between judgment-related information and its expression, rather than a claim of perfectly disentangled representations.

W2 / Q2. Generalization to genuinely unseen transformations.

We appreciate the reviewer's distinction between held-out examples and genuinely unseen transformation types. To directly address this concern, we re-estimated the expression subspace using only non-reversing response-format variations, excluding all negation and response-reversal examples across every task.

Despite never observing these transformations during subspace estimation, isolation consistently improves AUROC on all four held-out transformation types (Section 5.2 and Appendix C.2, Table 7):

Held-out transformationAUROC gainSettings improved
Negation+0.09614/14
No-means-correct+0.10614/14
False-means-correct+0.08213/14
0-means-correct+0.02110/14

These results demonstrate that the identified expression subspace generalizes to instruction transformations entirely absent during its estimation, rather than merely transferring previously observed transformations across tasks.

AuthorsOpenReview ↗

Response to Reviewer 3Stz

W3 / Q4. Novelty and practical value over simpler alternatives.
We appreciate the reviewer's comparison with simpler alternatives. While paired activation differences, SVD, and projection are established techniques, our contribution lies in the label-free identification and causal characterization of an expression-related subspace. As shown in W1, intervening on this subspace alters the model’s expression while its judgment signal remains decodable, providing causal evidence for a functional separation between judgment and expression.

The strong performance of the three-dimensional negation-derived subspace is itself an important finding of our work. Just three directions recover most of the improvement on an unseen response mapping (0.690 → 0.770 AUROC, versus 0.800 for the full method), revealing that expression-related interference is concentrated in remarkably few directions (Section 7.2 and Appendix C.4, Table 11).

Importantly, unlike the pooled baseline, which relies on labeled representations from multiple expression formats, our approach anchors supervision in a single reference condition and achieves robustness across expression formats through label-free subspace estimation. We further evaluated this distinction across matched numbers of labeled activations (Section 5.2 and Appendix C.2, Table 8):

Labeled representationsPooledOurs
160.7110.744
320.7420.765
640.7650.782
800.7710.786
All0.7880.801

Across these budgets, our method consistently achieves higher average AUROC without requiring labeled representations from additional expression formats. This demonstrates the practical value of label-free isolation for robust judgment readout across different forms of expression.

W4. Generalization beyond binary expressions of judgment.

We appreciate the reviewer's concern about generalization beyond familiar wordings and binary judgments. We tested 20 new ways of asking whether an answer is incorrect (e.g., "contains an error" or "fails to match"), none used in subspace estimation. Isolation improves AUROC from 0.662 to 0.785, comparable to a wording-specific readout (0.783), with improvements across all 20 wordings and 11 model–dataset settings (Section 5.3 and Appendix C.5).

We further extended our evaluation beyond binary response formats. First, we tested explanation-then-answer responses, where models explain their assessments before providing a final answer (e.g., "The calculation contains an error because ... Verdict: No."). We apply a readout trained without explanations to the hidden state immediately before the final answer, after the explanation. Under negation, isolation improves AUROC from 0.445 to 0.702, exceeding the model's expressed assessment (0.606; Section 6.3 and Appendix D.5).

More importantly, we tested free-form critiques, where models express their judgments in their own words without binary response constraints (e.g., "The reasoning is sound, but the final calculation is incorrect because ..."). We read the expressed assessment by asking the same model whether its critique says the answer is correct (Yes, No, or Unclear). Across four model–dataset settings (Qwen2.5-7B and Llama-3.1-8B on GSM8K and TriviaQA), our readout measured before critique generation predicts correctness more reliably than this expressed assessment (0.844–0.933 vs. 0.549–0.781 AUROC; Section 6.3 and Appendix D.5, Table 29).

Together, these results show that recoverable judgment information is not limited to binary forms of expression, extending to explanations and open-ended assessments.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their thoughtful feedback. We have substantially strengthened the evidence for our central finding: judgment-related information and the way a model expresses its assessment can be functionally dissociated.

First, we provide direct causal evidence beyond improved readout robustness. Across 14 model–dataset settings, swapping only the identified expression-subspace component changes the model's answer in 65.9% of cases, nearly matching full hidden-state replacement (67.0%) and changing the same answers (98.5% agreement after the swap), while complementary- and random-subspace controls yield only 3.1% and 1.9% (Section 6.1 and Appendix D.4, Table 28). Crucially, an independently trained, unprojected judgment readout remains predictive after the intervention (Table 28). These findings establish that the identified component causally influences expression while judgment-related information remains recoverable.

Second, we demonstrate generalization beyond previously observed expression formats. Estimating the subspace exclusively from non-reversing format variations, with all negation and response-reversal examples excluded across tasks, still improves AUROC on four entirely unseen response mappings (+0.021 to +0.106, improving 10/14–14/14 settings; Section 5.2 and Appendix C.2, Table 7). Importantly, our method requires labels only for a single reference condition, using label-free subspace estimation to enable robust judgment readout across different forms of expression. The benefits persist across logistic regression, ridge regression, linear SVM, and a nonlinear MLP, as well as across label budgets (Appendix C.5, Table 16; Appendix C.2, Table 8).

Third, we substantially broaden the empirical scope. Beyond 20 unseen wordings, we evaluate explanation-based responses, open-ended critiques, and moral, socially sensitive, and safety-related judgments across ETHICS, BBQ, and ToxiGen (Sections 5.2, 5.3 and 6.3; Appendices C.5, C.6 and D.5). We further extend our evaluation to the newer Qwen3.5-9B model and larger Qwen2.5-32B/72B models, confirming that the benefits persist across model generations and scales (Section 5.2 and Appendix C.7, Tables 21 and 22).

Taken together, these results move beyond improved readout accuracy to show that expression can be causally altered while judgment-related information remains recoverable. This separation also has clear practical value: our label-free expression isolation enables more accurate and robust judgment recovery than both standard probing methods and the model's own direct answers (76.9% vs. 61.8% accuracy with an unlabeled threshold-calibration rule; Section 5.2 and Appendix C.9, Table 25).

AuthorsOpenReview ↗

A Note to All Reviewers

We sincerely thank all the reviewers for their thoughtful and constructive feedback. The suggestions and questions have led to a stronger and more general central argument. Looking at how much the paper has improved reminded us of the value of the review process.

We truly believe that the revised paper is much stronger now. We would greatly appreciate the opportunity to hear the reviewers’ thoughts again on how this work has evolved.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.