acceptodds
Under review as a conference paper at ICLR 2027

Answer-Symmetrized Representations Form an Error Field in Language Models

Abstract

Large language models may make errors that mislead users and cause harm. Monitoring internal states is promising because model activations encode rich information about generation, but a detector may learn which answer token was produced rather than whether the answer is wrong. We define the Error Field (E-FIELD) as information about whether a response is wrong that becomes readable after suppressing information about the answer token. We introduce Answer-Frame Symmetrization (AFS) to expose this signal. AFS suppresses information about what the model answered and highlights information about whether that answer is correct. On 6,400 visual questions generated by rules across 16 reasoning types, 11 detectors, four Qwen3.5 model sizes, and both in-domain (ID) and out-of-domain (OOD) settings, AFS + Variational Information Bottleneck Probe (VIB-Probe) achieves a mean macro-averaged F1 score (Macro-F1) of 71.38%, compared with 63.52% for the strongest baseline selected separately in each setting (+7.86%↑), leading in seven of eight model–domain settings. These gains recur across four model families and extend to real-world benchmarks. Code is available in the [anonymous repository](https://anonymous.4open.science/r/afs-efield-artifact-2027-3600/) for reproducibility and further analysis.

Primary area:
interpretability and explainable AI
Keywords:
large language models, AI interpretability, error detection

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_RthPOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

The paper studies whether internal error detectors rely on answer-label information rather than correctness. It proposes Answer-Frame Symmetrization (AFS): after generating an answer once, the method permutes the four answer options, teacher-forces the label corresponding to the original semantic answer, and averages internal representations across all 24 permutations. The authors interpret the resulting correctness-related signal as an “Error Field.”

Experiments use a synthetic visual-reasoning benchmark containing 6,400 questions across 16 reasoning types, with additional model-family comparisons and zero-shot transfer to GPQA Diamond. AFS combined with VIB-Probe achieves 71.38% mean Macro-F1 across eight model–domain settings, compared with 63.52% for the strongest baseline selected separately in each setting. The paper also analyzes answer-label decoding, cross-label transfer, linear subspace removal, and representation geometry.

Critical strengths
  • The paper identifies a meaningful potential confound in internal error detection. Distinguishing correctness information from answer-label identity is important for evaluating whether probes learn transferable signals. The cross-answer transfer experiments and permutation-based analysis provide useful diagnostics beyond reporting aggregate detection accuracy.
  • AFS produces meaningful empirical improvements, particularly in-domain. The approach improves both linear probes and VIB-Probe, and its gains recur across several model families.
Critical weaknesses
  1. The formulation and evaluation are restricted to multiple-choice questions. AFS requires a finite set of answer options whose labels can be permuted while preserving the selected semantic answer. This assumption does not directly extend to open-ended generation, summarization, or other common error-detection settings. The paper therefore does not establish that its separation of answer identity and correctness generalizes beyond this narrow setting.

  2. The novelty relative to existing work on selection and answer-order bias needs clearer justification. The influence of option order and answer labels on multiple-choice predictions is already extensively studied. Although this paper focuses on internal error detectors, it should more clearly distinguish its contribution from established permutation-based debiasing and aggregation approaches. Naming the resulting signal an “Error Field” does not by itself establish a distinct phenomenon.

  3. The significance of the issue at larger model scales remains unclear. The main scaling experiments cover Qwen3.5 models from 0.8B to 9B, whose correctness on the benchmark is only around 30–39%. It is uncertain whether answer-label interference remains comparably consequential in substantially larger, more capable models. Evaluating such models would help establish whether the findings reflect a broad limitation or primarily the behavior of smaller models on difficult tasks.

Questions for authors
  • At which exact token position are representations extracted for each detector—before predicting the mapped label or after consuming it?
  • Please provide confidence intervals or paired significance analyses for the main comparisons and GPQA transfer.
  • Option permutations can change model preferences and confidence. What evidence supports Equation (7)’s assumption that the error component remains unchanged while only an additive answer-token component varies?
8 comments

Dear reviewer:

Thank you for recognizing the meaningful potential confound and meaningful empirical improvements in our work. Our main contribution is precisely this: identify an overlooked source of answer-frame dependence → address it → improve error detection.

“The formulation of a problem is often more essential than its solution, which may be merely a matter of mathematical or experimental skill.” — Albert Einstein and Leopold Infeld, The Evolution of Physics (1938) The first step toward progress is often the hardest: recognizing a problem that others have learned to overlook. Once a problem is made visible, even a simple solution can represent a meaningful advance. What ultimately matters is not how impressive a method appears, but whether it withstands the test of evidence.

Token prediction ≠ correctness detection. Activations contain information about how an answer is expressed, alongside information useful for assessing correctness. We study whether the detector's judgment changes when the answer's meaning and correctness stay fixed. The original VIB detector's mean agreement with its per-item majority decision across equivalent frames is 77.96% on ID / 86.82% on OOD, demonstrating a concrete monitoring problem.

The table compares the papers' research focus: interference from answer-label tokens (A/B/C/D) in correctness prediction when the semantic answer is held fixed. ✓ = a central research question; ✗ = not a central research question.

Paper / methodVenue / publicationFocus on answer-token identity interference?Mean Macro-F1 (%)
Semantic Entropy Probes: Robust and Cheap Hallucination Detection in LLMsarXiv 2024✗52.50
Real-Time Detection of Hallucinated Entities in Long-Form GenerationarXiv 2025✗56.96
Lookback Lens: Detecting and Mitigating Contextual Hallucinations in Large Language Models Using Only Attention MapsEMNLP 2024✗52.83
HALP: Detecting Hallucinations in Vision-Language Models without Generating a Single TokenEACL 2026✗58.67
VIB-Probe: Detecting and Mitigating Hallucinations in Vision-Language Models via Variational Information BottleneckACL 2026✗62.60
Hallucination Detection in LLMs with Topological Divergence on Attention GraphsACL 2026✗51.23
Spilled Energy in Large Language ModelsICLR 2026✗46.57
INSIDE: LLMs' Internal States Retain the Power of Hallucination DetectionICLR 2024✗52.23
Detecting Hallucinations in Large Language Models Using Semantic EntropyNature 2024✗52.60
SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language ModelsEMNLP 2023✗52.00
Language Models (Mostly) Know What They KnowarXiv 2022✗51.47
AFS + VIB-Probe (ours)This submission✓71.38

AFS: fixed semantic answer → equivalent answer frames → averaged activations → frame-invariant representation. Selection-bias research examines which answer is generated; our work examines whether the fixed answer is correctly judged.

Evidence already providedResult
Strongest baseline selected separately in each setting → AFS63.52% → 71.38%; ↑7.85 pp
Best detection performance7/8 model–domain settings
Improvement with a linear detector8/8 settings
Model scales evaluated0.8B · 2B · 4B · 9B
Broader evaluationFour model families + GPQA Diamond transfer

These results show the value of examining the representations used by established detectors. Is identifying a practical weakness and addressing it with a simple, effective method not a substantive scientific contribution?

We are a university research team working with limited GPU resources. Evaluating substantially larger models is constrained by the computing capacity available to us. Within these resources, we have made a sustained effort to evaluate four model scales, multiple families, and transfer beyond the original benchmark. We hope these completed experiments receive appropriate weight in the assessment.

We chose multiple-choice questions because they allow precise control: change the answer frame while preserving meaning and correctness. This makes the problem directly testable. Extending the approach to open-ended natural-language generation is a direction for our future work. A carefully controlled first step can establish a useful contribution and a foundation for broader study.

**Given the meaningful problem and empirical gains already acknowledged in your review, we sincerely hope you could reconsider the rating in light of the clarified research question, the completed experiments, and the demonstrated improvement over established detectors.

We have revised the PDF under the title “Answer-Frame Symmetrization Improves Internal Error Detection in Language Models.” The revised paper removes the E-Field definition and additive decomposition, and centers its contributions on fixed-response error detection. Below, we clarify the technical distinctions and provide statistical analyses of the reported results.

W1. Multiple-choice formulation and experimental design

The multiple-choice formulation provides an exact control for studying correctness detection. Relabeling changes the answer frame while preserving the question, option meanings, selected semantic response, and its correctness. This lets us examine whether a detector changes its judgment when the answer being assessed remains the same.

The revised abstract and Section 3 explicitly define this four-option problem. The evaluations cover two complementary settings:

EvaluationExperimental purposeResult
Controlled visual benchmark: 6,400 questions, 16 reasoning typesProgrammatically verified correctness; held-out reasoning types; paired detector comparisonsAFS + VIB achieves 71.38% mean Macro-F1, versus 63.52% for the strongest per-setting baselines.
GPQA Diamond: 198 independent text questionsTransfer from visual-source training to a non-synthetic text benchmark, with the source detectors and thresholds frozenMacro-F1 increases from 31.83% to 57.44% with AFS-4 and to 52.82% with AFS-24.

The scientific question is how response-preserving transformations affect an error detector, and the finite answer set makes this intervention precisely testable. Extending this construction to open-ended responses requires defining an appropriate family of response-preserving transformations. The results above establish its value for the specified multiple-choice detection problem, including transfer across datasets and modalities.

W2. Novelty relative to answer-selection debiasing

AFS studies the correctness of an already-generated answer. Answer-selection debiasing studies how the model chooses that answer. These objectives lead to different interventions and evaluation targets.

For a concrete comparison, PriDe estimates an option-ID prior and corrects the answer-selection distribution. PA-GRPO, already cited in our Introduction, trains the answering model for more consistent decisions across option permutations. PriDe, ICLR 2024; PA-GRPO, ACL 2026.

AspectAnswer-selection debiasing, illustrated by PriDe and PA-GRPOAFS
Prediction targetWhich semantic answer should the model select?Is the model's originally selected semantic answer incorrect?
Role of permutationsEstimate selection bias or encourage consistent answer choicesConstruct equivalent frames for the same fixed response
InterventionCorrect answer probabilities or optimize the answering policyAverage internal representations before correctness prediction
Assessed responseThe method operates on answer selectionThe original semantic response is fixed and its mapped label is teacher-forced
Output and evaluationAnswer choices, selection bias, and answering performanceError scores, correctness Macro-F1, and detector transfer

The contribution also includes detector-specific findings. With the response and its correctness fixed, the original VIB detector's mean agreement across equivalent frames is 77.96% on ID and 86.82% on OOD. AFS improves the linear detector in all eight model–domain settings, and AFS + VIB exceeds the strongest per-setting baselines in seven of eight, by 7.85 pp on average. The new training controls further show that item-specific frame variation can improve transfer even when all evaluation features remain the same complete AFS average.

These are findings about the reliability and transfer of correctness readouts. The review names a related research area but does not identify a prior method that already supplies this fixed-response detection construction or these diagnostic findings. Sharing the permutation operation alone does not establish equivalent contributions. We ask that the novelty assessment consider the prediction target, the fixed-response control, and the measured detector behavior together.

W3. Model scale and task difficulty

The detection gains persist at the largest model in the Qwen scaling study and recur across independently evaluated model families. At Qwen3.5-9B, AFS + VIB improves ID Macro-F1 from 74.69% to 82.34% and OOD Macro-F1 from 60.83% to 65.37% relative to the strongest baselines in those settings.

The additional family comparisons hold each model's generated responses, correctness labels, and test examples fixed between Raw VIB and AFS + VIB:

BackboneID Macro-F1: Raw → AFS (%)OOD Macro-F1: Raw → AFS (%)OOD gain, 95% paired CI (pp)
Gemma-4-E4B-IT71.04 → 76.5159.32 → 67.51+8.19 [6.24, 10.12]
Phi-4-Multimodal70.29 → 76.9058.42 → 67.14+8.72 [6.97, 10.45]
GLM-4.6V-Flash74.11 → 76.1061.07 → 66.01+4.94 [2.93, 6.94]

The reported 30–39% describes the answering model's correctness. Within each detector comparison, that correctness is identical because the original responses are unchanged. Thus, the measured improvement comes from predicting errors more effectively on the same answers. Macro-F1 gives equal weight to the correct and incorrect classes.

The scale evidence currently covers Qwen3.5-0.8B through 9B; the additional experiments test recurrence across architectures. Establishing the magnitude of the effect in substantially larger models remains an empirical question. These comparisons establish recurrence across the evaluated sizes and architectures.

Q1. Exact token positions used by the detectors

AFS + VIB and AFS + Linear use the mapped answer-label position after that label has been consumed. Let tt be the position of that label in the supplied sequence. Its representation is computed from a prefix that includes the label; the state used to predict the label itself is at t−1t-1.

Detector or analysisRepresentation or signal used in our implementation
AFS + VIB / Raw VIBInputs to the full-attention layers' oprojo_{\mathrm{proj}} modules at tt, after consuming the selected or mapped answer label
Linear probe; answer-label decoding; INLP diagnosticsDecoder hidden states at tt, at the specified candidate layers
Real-Time Entity ProbeFinal decoder-layer hidden state at tt
Semantic Entropy ProbesHidden states from the final five decoder layers at the final answer token, immediately before EOS in the full sequence
HALPVision-encoder features, decoder states at the final visual token, and states at the final question token; these positions precede the generated response
Lookback LensAttention ratios at each supplied response-token position, including tt, averaged over response positions
TOHAAttention graphs over the prompt and supplied response tokens, through tt
Spilled EnergyLogits immediately before and after each supplied response token, including t−1t-1 and tt for the answer label; minimum pooling over the response
INSIDEHidden state after consuming each sampled answer label, at the selected middle layer; covariance across sampled answers
Semantic Entropy / SelfCheckGPTStatistics of sampled answers and their agreement
P(True)True/False likelihoods under a separate correctness-assessment prompt

For AFS, the supplied label denotes the model's own original answer, mapped to the new frame. Section 3.3 now explicitly states the post-consumption extraction position for AFS + VIB. Section 4.3 also treats near-perfect answer-label decoding at this position as an expected reference for the INLP intervention.

Reviewer_gWUkOpenReview ↗
Rating
4(+2)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

General Comment: I like the idea behind this paper, and I think the proposed method is interesting. However, I feel the paper is more focused on presenting the results as a new discovery than on explaining what is actually happening. The visual presentation is attractive, but several figures are difficult to understand, and some observations are presented as important findings when they are already expected or known in the literature.

My main concern is the definition of the Error Field (E-Field). From my understanding, the method is essentially averaging internal representations across different answer permutations and using the resulting representation for error detection. The experiments show that this improves performance, which is valuable, but I am not convinced that this demonstrates the existence of a new internal Error Field. I feel the paper makes stronger claims than what the mathematical formulation and experimental results actually support.

Overall, this is a well-written and polished paper, especially regarding its figures and visuals. However, certain claims appear overstated. I am willing to increase my score if the authors address the concerns raised above.

Mathematical Formulation: There are some problems with the notation. For example, (m) is introduced in Equation (2) as the method index, but the notation is confusing because it could also be interpreted as the layer from which the internal state is extracted. Later, the authors introduce different notation for layers and reuse (m) in Appendix C for the number of permutations. I think this should be clearer. The same happens with (S_4) and (\sigma). Although these are standard mathematical notations for permutations, a brief explanation would help the reader understand the formulation without having to infer their meaning.

My main problem is with Equations (3) to (8). Equations (3) and (7) introduce essentially the same decomposition, but Equation (7) makes an additional assumption that the error-related component remains constant across answer permutations. Why should this component remain invariant if the input and internal representations change?. Equation (8) only shows that the answer-token component becomes independent of the original answer label under the assumed decomposition. This does not demonstrate that the error-related component is isolated or even that the decomposition exists in the actual model representations.

From my understanding, the method is simply averaging internal states and training a detector over these representations. The paper does not independently identify the proposed E-Field. I think this is an important distinction because improving error detection does not necessarily mean that a separate error-related representation has been discovered.

About the Figure 2: I find this figure difficult to follow. For example, the blue dots are not clearly explained, and the reader has to infer what they represent. I initially assumed that they indicated the correct answer, but they actually represent the selected semantic answer under different permutations. This is explained in very small text that is difficult to read. Another problem is the visual question. I think using a real example from the proposed dataset, including its question and possible answers, would make the procedure much easier to understand. The fonts are generally too small, and the sequence of operations is difficult to follow. The figure looks attractive, but I do not think it explains the method clearly enough.

Experimental Observations: In Section 4.2, is Observation 2 really a surprising finding? The degradation of error-detection performance under distribution shifts is already a known problem in the literature. Although this experiment is useful for evaluating the baselines, I do not see what new understanding this observation provides. The same happens with Observation 3. The authors show that answer-token identity can be decoded with 100% accuracy from the answer-token representations. Is this surprising when the representations are extracted at the position of the answer token? I feel this result is overanalysed, particularly because it does not directly establish that answer-token information interferes with correctness detection.

Observations 6 and 7 are also very similar in their interpretation. I understand that they evaluate different detectors, but both essentially show that AFS improves error-detection performance. This is a useful result, but why is it repeatedly interpreted as evidence that an E-Field has been revealed? The results demonstrate that the transformed representations are more useful for error prediction, not necessarily that the proposed internal mechanism exists.

Another result that calls my attention is Appendix C. Why does using only four permutations produce better OOD performance than using all 24? If the main argument is that averaging over equivalent answer frames suppresses irrelevant information, I think this behavior deserves a much deeper explanation.

Section 4.6: E-Field Visualization: I have concerns about the conclusions drawn from this section. Is the qualitative visualization sufficient to support such a strong claim about the geometry of the E-Field?

The analysis uses only 120 examples from one model, and the authors select a local region to illustrate the separation between correct and incorrect answers. However, I can see overlapping correct and incorrect points in other regions of the figure. The selected region looks like a favorable example rather than evidence of a general geometric structure.

Why is this not evaluated quantitatively? For example, silhouette scores, nearest-neighbor purity, or other measures of class separability could help determine whether these representations actually form correctness-related structures. This could also be evaluated across different models and datasets instead of relying on a single PCA visualization. I think the figure is visually attractive, but the conclusions are much stronger than the evidence provided.

Novelty and Related Work: Another concern is the absence of relevant prior work, particularly Truth as a Trajectory (Damirchi et al., ACL 2026). That paper investigates how lexical and token-related information in hidden states can interfere with learning generalizable signals about reasoning validity. It proposes using representation changes across layers to reduce these confounds and improve correctness prediction.

I understand that AFS uses a different transformation based on permutations of answer labels, and I consider this an interesting methodological difference. However, the central motivation is very similar: both approaches attempt to reduce irrelevant information in internal representations to improve the detection of correct and incorrect answers. I think the authors should explain how AFS differs from this previous research direction and what new scientific understanding the proposed E-Field provides beyond another way of transforming internal representations.

Critical strengths
  1. Interesting motivation: The paper addresses an important problem regarding how internal representations encode answer identity and correctness, and how this information can affect error detection.
  2. Simple and interesting method: Answer-Frame Symmetrization is a relatively simple approach that can be applied to different internal error detectors without modifying the underlying model weights.
  3. Extensive experimental evaluation: The authors evaluate several model sizes and families, different error-detection methods, and both ID and OOD settings. The improvements obtained with AFS are promising, particularly when compared with the original detectors.
  4. Useful analysis: The cross-answer transfer experiments and the INLP analysis provide interesting evidence about the limitations of directly probing answer-token representations.
Critical weaknesses

Many of my concerns are explained in the summary:

  1. Mathematical formulation and notation
  2. Figure 2 and presentation
  3. Experimental observations and interpretation
  4. Section 4.6: E-Field visualization
  5. Novelty and related work
  6. Computational cost and practical application: The method requires extracting representations across multiple answer-frame permutations. Although some computation can be shared, using all 24 permutations introduces a considerable additional inference cost.
Questions for authors
  1. E-Field formulation: What evidence supports the assumption in Equation (7) that the error-related component remains invariant across answer-frame permutations? How can the authors distinguish the proposed E-Field from a representation that becomes more useful for supervised correctness prediction after averaging?
  2. Mechanism of AFS: How can the authors establish that the improvements come specifically from suppressing answer-token information rather than from averaging several representations? Could alternative aggregation strategies or additional controls help distinguish these explanations?
  3. Number of permutations: Why does AFS with four permutations outperform the complete 24-permutation version on the OOD evaluation in Appendix C? Does this suggest that complete symmetrization is not necessarily beneficial for error detection?
  4. Representation geometry: Why is the E-Field visualization evaluated only qualitatively using 120 examples from one model? Could the authors provide quantitative evidence using silhouette scores, nearest-neighbor purity, or other separability metrics across different models and datasets?
  5. Novelty: How does the proposed approach differ from Truth as a Trajectory (Damirchi et al., ACL 2026), particularly regarding the motivation of reducing token-related confounds in internal representations to improve correctness prediction? What additional scientific understanding does the E-Field formulation provide?
  6. Computational efficiency: What is the trade-off between the number of answer-frame permutations, error-detection performance, and computational cost? Could a smaller number of permutations provide a more practical alternative without sacrificing performance?
  7. Scientific interpretation: Several observations demonstrate improvements in error-detection performance or expected properties of answer-token representations. What additional evidence supports interpreting these results as the discovery of a distinct Error Field rather than the effectiveness of a representation-transformation method?

Comments on the paper

1 comment

3. Contributions of the learned detector, standardizer, and threshold

A training effect remains when the standardizer and threshold are both held fixed. We further analyzed the completed common-standardizer control: both detectors use the AFS-24 Train-ID mean and scale, and both receive identical AFS-24 validation/test features. The training features are the experimental variable. Checkpoints are selected by the same ID-validation criterion.

Training featuresOOD Macro-F1, ID-selected threshold (%)OOD Macro-F1, common threshold 0.5 (%)OOD AUROC
AFS-2463.92 ± 2.3062.99 ± 0.910.7436 ± 0.0165
AFS-467.26 ± 1.9466.63 ± 1.470.7648 ± 0.0158

Results use five paired training seeds. With both the standardizer and decision threshold fixed, the learned detectors differ by 3.63 pp Macro-F1. The 0.0212 AUROC improvement measures a change in score ranking independently of threshold choice. This isolates an effect of the learned scoring function under a common evaluation protocol.

The within-type/cross-type analysis further explains how score transfer differs:

DiagnosticAFS-4AFS-24
Within-type AUROC0.8113 ± 0.01120.8116 ± 0.0126
Cross-type AUROC0.7621 ± 0.01310.7304 ± 0.0159
OOD Macro-F1 before score-mean alignment, confirmation seeds (%)68.36 ± 0.7363.09 ± 3.13
OOD Macro-F1 after score-mean alignment, confirmation seeds (%)71.45 ± 1.5872.31 ± 2.16

An additive shift shared by a reasoning type preserves ordering within that type but changes comparisons with other types and decisions under a global threshold. This is visible for the OOD timeline task: its class-balanced logit midpoint shifts by +0.251 for AFS-4 and −2.208 for AFS-24 relative to ID validation, moving AFS-24 scores toward “correct.”

Aligning score means by visual form closes the aggregate performance gap while keeping the detector weights, standardizer, and original ID threshold fixed. The alignment uses unlabeled target scores from other folds. Together, these results identify a training effect with normalization and threshold controlled, and locate an important transfer difference in the relative score positions of different reasoning types.

4. Generality of four frames and selection of the budget

The four-frame Macro-F1 advantage also appears on GPQA Diamond. The visual-source detectors and thresholds are transferred directly to GPQA without retraining.

Evaluation, Qwen3.5-4B + VIBAFS-4 Macro-F1 (%)AFS-24 Macro-F1 (%)Difference (pp)
Visual benchmark, OOD68.1057.63+10.47
GPQA Diamond57.4452.82+4.62

The cross-size composition ablation adds evidence about how the benefit depends on the model and the choice of frames. This experiment uses linear probes and two four-frame constructions: label-position moves the selected answer to A/B/C/D while preserving the distractors' relative order; random balanced samples one frame per selected-answer label.

BackboneFour label-position framesFour random balanced framesAll 24 frames
Qwen3.5-0.8B62.7856.86 ± 3.3460.61
Qwen3.5-2B55.5255.20 ± 3.1854.88
Qwen3.5-4B63.4459.92 ± 5.2061.18
Qwen3.5-9B56.6958.31 ± 4.6760.02

Entries are OOD Macro-F1 (%); random-balanced results summarize ten subset seeds. These constructions differ from the fixed four-frame VIB configuration above. The label-position construction exceeds the full average at three model sizes, while the ordering reverses at 9B. Frame composition and the learned detector jointly determine the useful budget.

For the VIB experiments, the fixed set is ABCD, BADC, CDAB, DCBA, where each string specifies the new labels assigned to the original A/B/C/D options. Every original label maps to each possible label exactly once. Four is the smallest budget that achieves this balanced coverage.

ChoiceHow it was determined
Four permutationsA deterministic balanced block containing the identity permutation, constructed from the permutation set without using data or detection scores
Budgets evaluatedFixed grid 1, 4, 8, 12, 24, with a separately trained detector at each budget
Standardizer and detector fittingTrain-ID
Decision thresholdsValidation-ID, then frozen for testing
Practical preference for four framesAn empirical conclusion from the reported ID/OOD budget ablation and runtime measurements

The main comparison uses the complete 24-frame average. The four-frame results establish a useful lower-cost configuration and show that maximum averaging and maximum transfer performance can occur at different budgets.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.