acceptodds
Under review as a conference paper at ICLR 2027

Latent-Space Counterfactual Verification for Robust Reasoning Agents

Abstract

A reasoning trajectory can produce the correct answer while relying on a brittle shortcut. We show that nominal accuracy substantially overestimates strict robust accuracy under semantics-preserving edits, by 21.9% on average and up to 65.4% across 25 model–task configurations and can even reverse model rankings on MATH-500. To detect such failures, we introduce GAUGE, a counterfactual verifier that tests whether model representations respond consistently to controlled input transformations with known answer-level invariances and equivariances. GAUGE learns a contrastive robustness metric and uses conformal calibration to provide statistically valid score thresholds, enabling reference-free scoring at test time. Across 23 configurations, GAUGE predicts which nominally correct outputs fail under held-out edits—a property, we show, predominantly of the prompt rather than the trajectory—with a conditional AUROC of 0.770, outperforming 14 published baselines. Its probe-free distillation reads the candidate’s own hidden state, adding +0.021 AUROC over confidence and length and matching pool self-consistency at zero extra generation. Using GAUGE for self-training selection improves strict robust accuracy while maintaining nominal accuracy, recovering 60–67% of available oracle gains.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM reasoning, reasoning verification, robust reasoning, latent representations, counterfactual verification, conformal prediction

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_DJN1OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GAUGE, a verifier that measures model responses to typed prompt transformations with known answer mappings. It learns a latent metric, combines representational drift with answer agreement, and calibrates score thresholds using split conformal calibration. Supervised variants fit robustness predictors, while distilled variants score the candidate's own activations without extra generation. Experiments compare verifiers, diagnose prompt-level reliability, and use GAUGE to select self-training data. The reported robustness largely concerns repeated policy behavior on a prompt rather than the validity of a specific reasoning trajectory.

Critical strengths
  1. The paper includes extensive comparisons and informative controls, including unedited resampling, within-prompt evaluation, and matched feature ablations. These clarify what the measurements actually capture.

  2. The distilled predictor offers a modest but supported benefit without extra generation: hidden states add about 0.021 AUROC over confidence and length. This is a useful observation within the evaluated settings.

Critical weaknesses
  1. The central verification story exceeds the demonstrated scope. The primary label mostly measures prompt-level reliability, and ranking correct trajectories within a prompt is near chance (AUROC 0.516). Unedited resampling reproduces 83% of the measured robustness gap, so most of that gap cannot be attributed specifically to sensitivity to edits or brittle shortcuts. The conformal result controls score coverage, not correctness or robustness among accepted candidates. These results provide limited support for verifying sound reasoning trajectories.

  2. The added complexity has insufficient demonstrated payoff. With matched supervision and tuning, predictors with and without latent features both reach 0.783 AUROC; the latent increment is 0.000. The distilled benefit is real but small. Self-training gains over outcome filtering are only 0.8-1.0 percentage points and are not significant after multiple-comparison correction. Probe-based scoring also needs about 7.6 extra rollouts per candidate. The evidence does not establish enough value beyond simpler consistency-based prediction to justify the full procedure.

Questions for authors
  1. Can GAUGE reliably distinguish better reasoning trajectories within the same prompt under candidate-linked labels and matched unedited-continuation controls? What evidence supports the shortcut-verification interpretation beyond general sampling instability?

  2. What practical benefit requires the learned latent metric when a matched predictor using probe answers, confidence, and length performs equally well? Please clarify the contribution beyond counterfactual consistency testing and its distillation.

  3. Do the self-training gains remain with the probe-free predictor, under a matched total-compute budget? More repeated runs and broader agent experiments would help establish whether the modest gains are reliable and practically useful.

2 comments

We thank the reviewer for a thorough and technically precise review. Two of the reviewer's points led to results that change the picture:

  • Conformal selection now bounds the share of fragile trajectories among accepted outputs. GAUGE-LR+ accepts 1.5× as many robust trajectories as self-consistency under the same bound.
  • The distilled readout removes the probe cost: with no probe generated, it matches the probe-based verifier.

Section, table and appendix numbers are unchanged from the submission; the three new tables are Tables 35–37 in a new Appendix X.

Changes in the revision

  • Conformal selection (W1): §3.5, Appendix A.8, Table 36. The original threshold-coverage theorem is unchanged and still described as covering the threshold only.
  • Resampling decomposition (W1): abstract, §1 and §4.2.
  • Payoff of the latent construction (W2): §4.4, new paragraph "Zero-generation verification" (teacher control, Table 35; zero-generation verifier; best-of-8). §4.5, §5 and Appendix M state that, with labels, probe answers match the learned metric.
  • Self-training with the probe-free selector, compute measured (Q3): §4.6 and Table 37.
  • Reference-free cues: affected probe rollouts regenerated and all evaluations rescored (Appendix R).

W1 / Q1. Within-prompt discrimination, shortcuts, and what is accepted

Within a prompt. We agree that the primary label is mostly prompt-level. Only 24% of prompts contain correct candidates whose labels differ, and GAUGE-LR+ ranks within a prompt at 0.516.

We also tested a candidate-linked label: fix the first half of a candidate's reasoning and test its continuations (17 configurations, 654 informative prompts).

  • Prefix-conditioned probes reach 0.681 [0.649, 0.717] within-prompt concordance, against 0.489 for ordinary probes.
  • The matched unedited-continuation control reaches 0.665 [0.639, 0.691].

Within-prompt discrimination is therefore real, and it comes from prefix stability. The revision describes GAUGE as an estimate of counterfactual reliability read from a trajectory, not as a check of a specific shortcut. That is the quantity data curation and output acceptance act on.

Edits beyond resampling. Edits are a distinct failure mode. Among initially correct candidates, 18.5% fail held-out edits, against 15.3% under an equal number of unedited resamples: +3.2 pp [+2.2, +4.1], positive in 19/23 configurations, with a median ratio of 1.6×. The abstract, §1 and §4.2 now report this decomposition.

A guarantee on what is accepted. We added conformal selection (Jin and Candès, 2023; §3.5, Appendix A.8, Table 36). Unlike the original score-coverage guarantee, conformal selection controls the expected proportion of fragile trajectories among accepted outputs. Under exchangeability, it guarantees

E[Vmax⁡(1,R)]≤q\mathbb{E}\left[\frac{V}{\max(1,R)}\right]\leq q

where VV is the number of fragile accepted trajectories, RR is the total number of accepted trajectories, and qq is the target false discovery rate.

To account for dependence among candidates sharing a prompt, calibration and evaluation each use one randomly selected candidate per prompt group. We average results over 100 representative draws per configuration, across 23 configurations.

Score (target q = 0.05)Realized fragile share among acceptedRobust trajectories accepted
GAUGE-LR+0.0540.499
GAUGE (label-free)0.0540.456
Zero-generation verifier0.0460.415
P(True), 30B judge0.0530.379
Pool self-consistency0.0550.331

GAUGE-LR+ accepts +0.168 [+0.082, +0.265] more of the robust trajectories than self-consistency (16/23), i.e. 1.5× as many. The label-free score adds +0.125 [+0.057, +0.204], and the zero-generation verifier +0.084 [+0.026, +0.151]. Every realized share is within two standard errors of its target. The advantage is at strict tolerances; at q = 0.10 the scores tie (Table 36).

W2 / Q2. What requires the latent construction, and is it worth the probes?

We agree with the matched ablation: with labels, predictors with and without the latent features both reach 0.783. The revision says so. What the construction contributes is the following:

  1. Typed interventions make a better target and teacher than plain consistency. GAUGE-LR+ exceeds paraphrase agreement by +0.073 [+0.043, +0.102] (21/23). Adding answer-changing and diagnostic edits to meaning-preserving ones raises the label-free score from 0.695 to 0.744.
  2. The robustness signal is readable from the candidate's own hidden states, whatever the teacher. Distilling into those hidden states adds +0.021 to +0.025 AUROC over confidence and length, whether the teacher is GAUGE-LR+, the matched predictor without latent features, or raw probe agreement (Table 35).
  3. That readout removes the probe cost. Ranked together with the self-consistency of the already sampled pool (an unweighted rank sum), it reaches 0.782 conditional AUROC with no probe generated, matching the probe-based predictor (0.783).
    • It beats pool self-consistency by +0.056 [+0.026, +0.087] (19/23).
    • At best-of-8 it submits a strictly robust answer 69.7% of the time, against 68.7% for self-consistency (+1.0 pp [+0.1, +2.0]) and 70.0% for GAUGE-LR+.

The 7.6 probe rollouts per candidate are therefore an offline cost of building the teacher, not a deployment cost.

Reference-free in practice. Only the misleading-answer decoy ever read the reference. Rebuilding it from each candidate's own answer, and regenerating the 684 affected rollouts, moves conditional AUROC by ≤ 0.0001 and unconditional AUROC by ≤ 0.0002 (Appendix R).

Q3. Self-training with the probe-free predictor, and its compute

We ran it (Qwen3-4B; the same 1,500-prompt mixed pool spanning ALFWorld, τ-bench, math, code, tool calls and ARC; 3 seeds × 3 rounds of LoRA rejection fine-tuning). All arms use the same 4 samples per prompt and keep one gold-correct candidate per eligible prompt, so sampling, data and training are matched; we measured total compute rather than equalizing it. The probe-free head was distilled once, from typed-probe agreement, and frozen.

Selector (3 seeds)Extra generations at selectionGPU-h / seedPaired Δ vs. outcome (Table 1 test)
Outcome-RFT03.5—
Probe-agreement-RFT7.6 / candidate17.8+0.2 pp [−2.6, +3.0]
GAUGE-RFT, probe-free04.4+0.4 pp [−2.2, +3.0]

GPU-hours cover sampling, scoring, selection and training on one GPU; the probe-free arm includes its one-off head fit (2.1 h), spread over the three seeds. Per-suite results, including ALFWorld and τ-bench, are in Table 37.

The probe-free selector matches the probe-based one at a quarter of the compute: +0.4 vs. +0.2 pp over outcome filtering, at 4.4 instead of 17.8 GPU-h per seed, and only 1.26× the cost of outcome filtering. As in Table 1, the self-training effect of either selector is small and not resolved at three seeds, and we present it that way. The reason is leverage: on the base-policy pool, typed-probe agreement differs among the correct candidates of only 21% of eligible prompts (232 of 1,130), so any within-prompt selector can change the kept candidate's robustness on few prompts. Where the score acts directly on the output, its gains are resolved: at best-of-8 (+1.0 pp over self-consistency, zero extra generations) and in conformal acceptance (1.5×).

We thank the reviewer. These questions sharpened the paper: GAUGE estimates counterfactual reliability from a trajectory's representation, bounds the fragile share of what it accepts, and needs no probes at deployment.

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

This paper presents a new method to test how stable/robust an LLM’s representation is under perturbations to the input question. Their method, GAUGE, is a contrastive robustness metric implemented to quantify model hidden state changes under typed perturbations. GAUGE uses conformal calibration using a 90th percentile threshold. They show the usefulness of the GAUGE method by using it to augment answer selection during self-training.

Critical strengths

The experimental methodology is sound, prompts are separated across splits. Comparisons uses clustered bootstrap with multiple-testing correction.

Comparison to other existing methods seems thorough. There are 14 published works that are compared to GAUGE.

The experiment/result seem extensive.

Critical weaknesses

The writing in this paper reads quite dense to me, a lot of jargon and notation are given early with little effort to explain their meaning. There are multiple variants of the GAUGE method that use different supervision and cost. It seems the -LR and -D can be moved to appendix without significant impact to the messaging in the main paper.

The impact messaging seems inflated. As the paper states: “we show that nominal accuracy substantially overestimates strict robust accuracy under semantics-preserving edits, by 21.9% on average…”. Yet Appendix H.3 on page 34, in section “Control: how much of the gap is the edit”. It says “Across 23 configurations the measured gap among correct evaluation-split candidates (Table 3 instead uses one rollout per instance) is 18.5% under edits against 15.3% under plain resampling: 83% of it is reproduced without editing anything, and the edit-specific part is +3.2 pp ([+2.2, +4.1], 19/23 configurations).” This is information that significantly limits the motivation, and it should be elaborated on more in the main text. Is this an issue that previous methods faced similarly? It would be helpful to have this information for a reader who is not thoroughly familiar with existing studies.

The abstract and parts of the introduction frames GAUGE as a solution to the issue of brittle reasoning path, which is a trajectory level problem. While the results remain at the prompt level. The authors do concede that the robustness GAUGE measures belongs mostly to the prompt. My question is, why introduce the framing as a trajectory level issue at all? If GAUGE does not adequately address the trajectory level problem, consider removing that implication from the abstract and introduction (example locations: first sentence of abstract, figure 1, etc.)

Post-Rebuttal update: The authors adequately addressed the 3 critiques I provided in their rebuttal and update

Questions for authors

detailed critiques & questions/suggestions included in the weakness section.

2 comments

We thank the reviewer for the positive assessment and for three concrete suggestions. We adopted all three. Section, table and appendix numbers are unchanged from the submission.

Changes in the revision

  • Clarity (W1): §3 opens with a plain-language overview and a short notation block. The main text uses three variants (GAUGE, GAUGE-LR+, GAUGE-D+); GAUGE-LR and GAUGE-D are reported in Appendix F.
  • Resampling vs. edits (W2): the decomposition, including the 1.6× edit-specific ratio, is now in the abstract, §1 and §4.2. §2 adds one sentence on prior perturbation benchmarks.
  • Framing (W3): the abstract's first sentence and Figure 1's title no longer describe a brittle reasoning path.

1. Density and variants

§3 now begins with the idea in plain words. Edit a question in controlled ways and check whether the model responds as each edit requires. Then learn to predict that behavior from the original answer's hidden states, so that no extra generations are needed at test time.

Each edit family has a known answer map, which is what makes scoring reference-free:

  • S keeps the answer (e.g., rewording).
  • V changes it by a known rule (e.g., a changed quantity).
  • C adds a persuasive but irrelevant cue that should be ignored.

The main text now uses three variants, distinguished by supervision and cost:

  • GAUGE: label-free and probe-based.
  • GAUGE-LR+: supervised, for when robustness labels exist.
  • GAUGE-D+: reads the candidate's own hidden states, with no test-time generation.

As the reviewer suggested, GAUGE-LR and GAUGE-D are reported in Appendix F. A short notation block precedes the formal setup, and the full glossary remains in Appendix H.5 (Table 20).

2. The 21.9% headline and resampling

We agree that this belongs in the main text, and it is there now. The matched control separates the two components among initially correct candidates (23 configurations):

Initially correct candidates, evaluated underFailure rate
Held-out meaning-preserving edits18.5%
An equal number of unedited resamples15.3%
Edit-specific increment+3.2 pp [+2.2, +4.1], 19/23

Edits raise the failure rate by a median 1.6× over plain resampling. The accurate reading is two failure modes: a correct answer may not reproduce on a fresh sample, and controlled edits add a consistent further failure. The 21.9% headline aggregates differently (25 configurations, one rollout per instance), so it should not be compared directly with the 18.5% figure.

Does this affect prior work? Yes, in general. Any strict metric that requires correctness under several sampled edits also measures resampling instability unless it includes an equal-budget unedited control. The same caution applies to declines reported on GSM-Symbolic, GSM-IC, GSM-Plus and MATH-Perturb; §2 now says this in one sentence. Within our paper, GAUGE and all 14 baselines are scored against the same labels, so every comparison shares one target.

3. Trajectory vs. prompt framing

We followed the suggestion:

  • The abstract now opens: "A model can produce the correct answer yet fail once its prompt is edited in ways that should not change it."
  • Figure 1 is retitled "Fragile answers become visible under the right latent metric."

The accurate description is that GAUGE's input is a trajectory and its target is mostly a property of the prompt. This combination is what makes the method useful:

  • Curation and acceptance rank across prompts. Under conformal selection with a 5% bound on the fragile share among accepted outputs, GAUGE-LR+ accepts 1.5× as many robust trajectories as self-consistency (§3.5; Table 36).
  • The signal is in the trajectory's representation. A head distilled into the candidate's own hidden states adds +0.021 to +0.025 AUROC over confidence and length, whichever teacher it is distilled from (Table 35). Combined with pool self-consistency, it matches the probe-based verifier (0.782 vs. 0.783) with no extra generation.
  • Within a prompt, the paper is explicit about scope. The candidate-linked analysis shows discrimination driven by prefix stability (Appendix H.3), and we do not present GAUGE as verifying an individual reasoning path.

We thank the reviewer. These changes make the paper easier to read and its claims more precise.

Thank you for the revisions and for taking careful consideration of my review. The changes adequately address the 3 concerns I had in my initial review. I have updated the critical weakness section of the review with a post-rebuttal update to reflect this. I am also increasing the presentation score from a 2 to a 3.

I am keeping the final rating at a 3, which already reflected a positive assessment. My reviews and the authors' edits focused on improving the soundness and presentation of the narrative. My remaining reservation has to do with the scope of the paper's contribution.

Reviewer_kKRfOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

he paper studies whether a model’s correct answers remain correct under controlled changes to the prompt. It introduces GAUGE, which uses transformations with known answer-level relationships and measures the resulting changes in hidden representations. The paper evaluates several versions, including a supervised combination of latent features and answer agreement, and a distilled predictor that requires no additional generations at inference time. Split-conformal calibration provides marginal coverage for the score threshold. The experiments examine prediction of robustness, selection among candidate outputs, and selection of data for self-training.

Critical strengths

I find the motivation important. Correctness on the original prompt gives only a limited picture of a model’s reliability, and testing its behavior under controlled transformations is a sensible approach. The use of known answer maps provides a clean formulation of the consistency that the verifier is trying to measure.

The experimental study is broad, and I particularly appreciate the controls in Appendix H.3. The authors examine ordinary resampling, distinguish prompt-level reliability from trajectory-level robustness, and compare against a supervised predictor with the latent features removed. These are consequential controls, and the paper deserves credit for reporting the negative findings as well as the positive ones.

Critical weaknesses

The experiments do not yet establish the value of the proposed latent geometry.

The most relevant comparison appears in Appendix H.3, p. 35. The full supervised predictor and the otherwise identical predictor without the four latent features both obtain AUROC 0.783; the incremental contribution is 0.000, with confidence interval [−0.007, 0.007]. These comparisons use the same labels, splits, and tuning budget. They provide stronger evidence about the latent contribution than comparisons against less closely matched verifiers. The authors acknowledge this and point to distillation as the regime where hidden representations help. However, the distillation experiment answers a different question: it establishes that the student benefits from receiving hidden states. It does not establish that the teacher needs the proposed latent metric. An otherwise identical student distilled from the matched predictor using only probe answers, confidence, and length would directly address this issue. The self-training results do not resolve this concern either. The improvements over outcome filtering are consistently positive, but amount to 0.8-1.0 percentage points in strict robust accuracy and do not survive the reported multiple-comparison correction over three seeds. Moreover, Table 1 does not include a selector based on the matched predictor without latent features. Thus, these results provide encouraging evidence for the selection procedure, but do not isolate the contribution of the latent construction.

Questions for authors
  1. Does the distilled advantage require the latent features in the teacher? Please compare the same student, with the same inputs, training data, and tuning budget, distilled from GAUGE-LR+ versus the matched predictor without latent features already evaluated in Appendix H.3. This comparison would substantially affect my assessment of the methodological contribution.

  2. Which results use candidate-derived controls throughout? In particular, were the unconditional evaluations, best-of-(N) experiments, and “no labels” self-training arm run with misleading-answer cues generated from candidate answers? If the reference-derived cues were used, how do the results change with the fully reference-free construction?

2 comments

We thank the reviewer for a precise review. We ran the matched-teacher comparison exactly as specified, and it gave the clearest result of the revision: the distilled advantage does not depend on the latent teacher. Section, table and appendix numbers are unchanged from the submission; the three new tables are Tables 35–37 in a new Appendix X.

Changes in the revision

  • Teacher control (Q1): §4.4 (new paragraph "Zero-generation verification"), Appendix H.3, Table 35.
  • Scope of the latent metric: §4.5, §5 and Appendix M now state that, with robustness labels, it adds no supervised accuracy.
  • Reference-free cues (Q2): every affected probe rollout regenerated and every evaluation rescored (Appendix R).
  • Best-of-8 with the zero-generation verifier: one sentence in §4.4; Appendix H.6.
  • Self-training with an answer-only selector: §4.6 and Table 37.

Q1. Does the distilled advantage require latent features in the teacher?

No. We used the same student, inputs, training data, splits and ridge selection, and changed only the teacher (Table 35). Values are conditional AUROC over 23 configurations, with two-level bootstrap intervals.

TeacherStudent: hidden states + confidence/lengthConfidence/length onlyHidden-state gain
GAUGE-LR+0.7300.710+0.021 [+0.005, +0.035], 19/23
Matched predictor without latent features (App. H.3)0.7340.709+0.025 [+0.009, +0.040], 19/23
Typed-probe answer agreement0.7330.709+0.023 [+0.001, +0.042], 18/23

The two students the reviewer asked about differ by −0.004 [−0.011, +0.002]. The advantage therefore belongs to the student: a candidate's own hidden states predict counterfactual robustness beyond confidence and length, whatever teacher supplies the target. The signal sits in the candidate's trajectory, not in the probes: with probe answers withheld, the probes' hidden states add nothing (0.676 vs. 0.677; Appendix H.3).

This makes the readout practical. An unweighted rank sum of the student with the self-consistency of the already sampled pool reaches 0.782 conditional AUROC with no probe generated:

  • +0.056 [+0.026, +0.087] over pool self-consistency (0.726), positive in 19/23 configurations;
  • on par with the probe-based predictor under matched tuning (0.783).

At best-of-8, the rank sum submits a strictly robust answer 69.7% of the time, against 68.7% for self-consistency (+1.0 pp [+0.1, +2.0]) and 70.0% for the probe-based GAUGE-LR+.

The latent metric. We accept the reviewer's reading of the matched ablation (0.783 vs. 0.783): with robustness labels, the learned metric adds no supervised accuracy. The revision says so (§4.5, §5, Appendix M), and the method's claims now rest on the label-free score and on the zero-generation readout above.

Self-training. We agree that Table 1 cannot isolate the latent construction. The revision presents it as evidence for the selection procedure, as the reviewer suggests, not for the latent metric.

To add a selector without latent features, we ran a new study (Qwen3-4B; the same 1,500-prompt mixed pool; 3 seeds × 3 rounds of LoRA rejection fine-tuning). All arms use the same 4 samples per prompt and keep one gold-correct candidate per eligible prompt, so they differ only in which candidate is kept; total compute is measured, not equalized. Two selectors are new. An answer-only selector, typed-probe agreement, has no latent features; it is the signal the matched predictor is built on, used directly rather than refitted. A probe-free selector is the student above, distilled once from typed-probe agreement, plus pool self-consistency.

Selector (3 seeds)Extra generations at selectionGPU-h / seedPaired Δ vs. outcome (Table 1 test)
Outcome-RFT03.5—
Probe agreement (answer-only, no latent features)7.6 / candidate17.8+0.2 pp [−2.6, +3.0]
Probe-free student + pool self-consistency04.4+0.4 pp [−2.2, +3.0]

The probe-free selector matches the answer-only selector at a quarter of its compute (4.4 vs. 17.8 GPU-h per seed) and with no probes, at 1.26× the cost of outcome filtering. As in Table 1, the downstream effect of either selector is small and not resolved at three seeds. The reason is leverage: on the base-policy pool, typed-probe agreement differs among the correct candidates of only 21% of eligible prompts (232 of 1,130), so a selector can change the kept candidate's robustness on few prompts. Per-suite results are in Table 37.

Q2. Which results use candidate-derived cues?

One transformation read the reference: the decoy in the misleading-answer cue. It was screened against the reference answer. We reran every affected evaluation with the decoy screened against the candidate's own answer; when that answer cannot be parsed, no option is excluded (Appendix R).

  • Conditional evaluations (correct candidates): the cue is identical for 17,171 of 17,204 candidates (99.8%; the rest are \frac/\dfrac variants). Conditional AUROC moves by ≤ 0.0001.
  • Unconditional evaluations and best-of-N: 684 of 20,920 cues differ (3.3%), almost all on incorrect candidates. We regenerated these 684 rollouts with the production seed rule, token budget and feature extractor; stored rollouts re-featurize at cosine ≥ 0.9998. We then rescored with the unchanged pipeline.
    • Unconditional AUROC moves by ≤ 0.0002 on average (0.0045 in the most affected configuration).
    • Best-of-N strict robust accuracy moves by −0.0004 [−0.0013, 0.0000].
  • "No labels" self-training arm: this arm fits no robustness labels, but its probes used the same decoy rule. On the base-policy training pool of our new run, we regenerated all 294 affected cue rollouts (82 prompts) and reran the production selection under both rules. The selected training sets overlap by 99.9%.

Every reported evaluation is therefore that of a fully reference-free pipeline, up to average changes of ≤ 0.0002 AUROC (0.0045 in the worst configuration). Fitting is a separate matter: GAUGE-LR+ and the frozen coefficients use robustness labels on the fit split, as Appendix R states.

We hope the teacher control resolves the reviewer's main concern. The distilled advantage does not depend on the latent teacher, and it yields a verifier that needs no extra generations and matches the probe-based one.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.