acceptodds
Under review as a conference paper at ICLR 2027

Can Correctness Readouts Choose the Right Answer? A Question-Level Audit of Compact KV-Cache Signals

Abstract

Can a correctness readout that ranks responses well in aggregate identify which answer to choose within a question? We study this question with KV24, a compact 24-feature summary of late-layer key–value cache statistics, paired with a lightweight correctness predictor; the language model remains frozen. We read these features before the dedicated final-answer field, although reasoning may already contain the answer. On held-out GSM8K, assigning every Qwen response its question's mean score removes all within-question ordering. The area under the receiver operating characteristic curve (AUC), computed over responses pooled across questions, changes from .9166 to .9104. Within-question ranking evidence is positive for Llama but inconclusive for Qwen. Hidden-state and likelihood controls match or outperform KV24 under several comparisons. Without target refitting, the joint Qwen-Llama ranking test passes narrowly on OpenBookQA but not on a synthetic cross-domain suite. Changing the source-validation selection criterion yields no consistent answer-choice benefit across models, and same-pool majority voting matches or exceeds the selected readouts. Together, these results separate boundary readability from transport and answer-choice utility.

Primary area:
interpretability and explainable AI
Keywords:
KV cache, llm, reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Z9f2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper audits whether correctness scores derived from internal model signals can support answer selection. Using KV24, a compact summary of KV-cache statistics, it distinguishes pooled correctness discrimination, within-question ranking, and actual answer-choice accuracy. The central finding is that high pooled AUC can largely survive the removal of within-question ordering, while downstream selection gains remain inconsistent.

Critical strengths
  1. The paper clearly demonstrates an important evaluation pitfall: strong pooled correctness discrimination does not necessarily translate into better answer selection.
  2. The empirical audit is careful and transparent, including representation controls, frozen transfer tests, same-pool voting, and explicit reporting of negative results.
Critical weaknesses
  1. Limited incremental contribution. Question-level confounding is already recognized in prior work. The paper offers a detailed diagnostic study, but its additional methodological insight remains modest, and no effective corrective method is demonstrated.
  2. Insufficient evidence of generality. The investigation centers on KV24 and restricted readout families. Small mixed-question samples and model-dependent response eligibility further limit the evidence. These findings do not establish how broadly the observed failures apply to stronger answer-selection systems.
Questions for authors

I suggest evaluating a broader range of verifiers to establish whether the identified limitations are widespread. An analysis restricted to a few verifier families provides insufficient evidence for the broader significance needed for a strong contribution.

Reviewer_tB7MOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper examines whether correctness readouts that distinguish correct and incorrect language-model responses in pooled evaluations can also select a correct answer from multiple responses to the same question. Its central test case, KV24, extracts 24 late-layer key/value-cache magnitude statistics (mean absolute value and RMS across three layers, two token suffixes, and keys/values) at the generated final-answer marker. A ridge-logistic classifier is trained on response-level correctness labels while the base language model remains frozen. The paper differentiates pooled ROC AUC, within-question correct–incorrect concordance, and Best-of-8 answer-selection accuracy. On Qwen GSM8K, replacing response scores with their question-level means changes pooled AUC only from 0.9166 to 0.9104 while removing all sibling ordering (Figure 1; Table 13). Hidden-state and likelihood readouts often match or outperform KV24 (Tables 2–3). Source-frozen KV24 passes the prespecified dual-model one-sided test on OpenBookQA narrowly, but does not establish the corresponding result on a synthetic cross-domain suite (Tables 4–5). Compact readouts show no consistent, statistically resolved Best-of-8 advantage over same-pool voting (Tables 7–8 and Appendix E). The principal contribution is an evaluation audit, rather than a superior verifier architecture.

Critical strengths
  1. A decision-relevant evaluation distinction. The question-mean replacement experiment clearly illustrates why high pooled ROC AUC does not by itself imply the ability to rank candidate responses to a given question. The separation of within-question concordance from actual top-ranked answer accuracy is also important: pairwise ranking quality and Best-of-N selection are not interchangeable (Figure 1; Table 6).

  2. Careful statistical design. Question-level splits and resampling are appropriate for repeated generations. The authors account for invalid answers and full question-level selection denominators. The analysis that repeats source-validation readout choice during bootstrap resampling is particularly valuable; it avoids treating a source-selected verifier as though it were prespecified independently of the validation data (Appendix Table 31).

  3. Multiple controls and transparent reporting. The study includes equal-width and operator-matched hidden-state controls, likelihood features, alternative extraction boundaries, transfer tests, and voting. It reports negative findings openly and generally restricts representation-specific claims to the tested readouts (Sections 4–8). The distinction between selective filtering and answer selection is useful.

Critical weaknesses
  1. The headline answer-selection evidence covers a limited readout family. The main source-selected Best-of-8 comparison evaluates KV24, Hidden24, LM5, and KV24+LM5, whereas higher-dimensional Hidden4096 and the LiLaVe-style XGBoost verifier are assessed elsewhere but not in the same decision-level protocol (Appendix Tables 15, 26, 34; Section 7). This does not invalidate the negative conclusions for the tested compact readouts, but it prevents us from determining whether the outcome persists for more expressive verifiers. I recognize that those stronger controls sometimes use historical response pools, so a perfectly matched rerun may entail additional work. The same-pool vote also observes completed answer strings, whereas boundary readouts do not; matching generation budgets does not imply identical information budgets. The evidence supports no reliable advantage of the evaluated compact selectors over voting, not a universal failure of correctness readouts. Indeed, Qwen OpenBookQA has a negligible +0.20-point selected-readout difference whose interval includes zero (Appendix Table 28).

  2. Training and downstream objectives are not directly matched. The primary readouts minimize response-level binary logistic loss, while deployment chooses the highest-scoring sibling for each question. Logistic scoring is not inherently inappropriate: well-estimated correctness probabilities can support optimal selection. Nevertheless, with finite data and heterogeneous question difficulty, response-wise training may favor between-question discrimination over within-question ranking. A matched question-centered, pairwise-ranking, or top-1-oriented fitting ablation would test whether the loss function contributes to the limited selection performance. Such an experiment must disclose the number of mixed training questions, since only 17/160 source-test questions are mixed for Qwen (Figure 1).

  3. The source and transfer evaluations limit the strength of the generalization claim. With 17/160 mixed Qwen questions, the source-test within-question statistic is based on relatively sparse support; Best-of-8 source accuracies around 94% also leave little room to distinguish selectors (Appendix Table 27). The question-mean AUC preservation is still an empirical finding, not guaranteed by the mixed-question count, and Appendix Table 14 already includes prompt-conditioned controls. On the synthetic cross-domain suite, task design, model capability, checker validity, and response format vary considerably. For example, on scorable responses Llama's instruction/code accuracies are 10.72%/5.99% (Appendix Table 20), and only 2,240/4,000 Granite responses meet the original format criterion (Table 5). These factors complicate causal interpretation of transfer failures. Conversely, the OpenBookQA dual-model test does pass its prespecified one-sided criterion (Llama's lower bound is 0.5041 and the combined test reports p=0.02859), but the effect is small and the Llama complete-case subset is limited and inconclusive (Table 4; Appendix D.1, Table 18).

  4. The choice of magnitude-only KV24 features remains insufficiently investigated for conclusions about KV representation capacity. KV24 aggregates across heads and coordinates and discards sign, relative vector orientation, and most token/head identity. The main manuscript does not report a matched capacity or geometry ablation varying KV-layer count, head-specific statistics, or richer low-dimensional projections. In particular, the label-free KV-CoE trajectory score is not a test of adding supervised per-head cosine-similarity statistics to KV24: its representation, information support, and scoring rule differ (Appendix Table 11). A magnitude-only / angular-only / combined supervised probe comparison could clarify whether geometric information improves within-question discrimination, without assuming that it must. Care is needed with post-RoPE keys and with token support (a one-token window contains no token pairs). This limitation qualifies inference about richer KV-based probes; it does not overturn the paper's evaluation principle.

  5. Novelty and evidence-status tracking can be sharper. Prior research already examines question-level confounding and hidden-state verifier selection. The paper's added value appears to lie in the combined evaluation design and uncertainty treatment, which should be foregrounded more explicitly. Some comparisons are retrospective or use historical pools, while only some primary transfer decisions were outcome-isolated; the distinctions are difficult to follow. Appendix G provides statistical replays but states that the release does not reproduce all raw generation, extraction, fitting, and calibration stages end-to-end. These limitations are disclosed, but clearer documentation would improve confidence in the scope of the conclusions.

Questions for authors
  1. Decision-level comparison: Can the authors provide Best-of-N results for a stronger verifier (e.g., Hidden4096 or XGBoost) on a matched pool, or otherwise explain what evidence would be needed to extend the present negative result beyond compact readouts?

  2. Objective alignment: What happens if the same KV24 features and candidate pools are used with a question-centered or within-question ranking objective rather than response-wise binary classification? How many mixed-outcome questions are available in training?

  3. Transfer interpretation: Can the authors offer a clearer capability- and validity-stratified analysis of the cross-domain results, or identify whether the principal failure persists in comparably scorable, mixed-outcome candidate pools?

  4. Feature design: What motivated the three-layer, 1/4-token, magnitude-only KV24 construction? Can the authors summarize the head/trajectory explorations cited in Appendix G and clarify whether they tested supervised angular or per-head KV features? This is a secondary request to delimit the representation-specific conclusion, not a demand for a new architecture.

  5. Novelty and cohorts: Which components of the audit are methodologically new relative to the cited question-confounding and hidden-verifier papers? Could a single table separate the response pools and prospective versus retrospective analyses?

Reviewer_rg1FOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes KV24, a method for predicting the correctness of model responses using 24 magnitude statistics of keys and values from the model's late-layer KV caches. Experiments are conducted with Qwen3-8B, Llama3.1-8B, and Granite-3.3-8B on GSM8K and MathQA. Comparisons are provided against several variants such as Hidden24 and Random24.

Critical strengths

None

Critical weaknesses
  • The positioning of the paper is unclear. Its target scenario is essentially process reward modeling (PRM). Yet PRM is never mentioned in the paper.

  • Unlike classical PRMs that only require access to text strings, KV24 requires access to a model's internal states, which makes it inapplicable in evaluating the responses generated by API-based models.

  • The experiments do not provide comparison with any baseline. If I understand correctly, all the comparisons (KV24, Hidden24, PCA24, LM5, Random24, etc.) are variants of the proposed methods.

  • The datasets used---GSM8K and MathQA---are too easy for modern LLMs. Please switch to more commonly used ones such as MATH and AIME.

Questions for authors
  • How does KV24 compare with commonly used PRMs?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.