{"id":"e3851501-a1a6-427c-8896-a6140d72b014","arxiv_id":"2507.03998","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Adding data-agnostic probability and entropy features to hidden-state probes improves cross-task generalization in most but not all evaluated transfer pairs.","lead":"This paper tests whether adding simple probability and entropy features to hidden-state probes helps LLM uncertainty estimators generalize across tasks and datasets. The results are mixed: the added features help in most cases but hurt in others, and feature selection only sometimes amplifies the gains.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No error bars or repeated-seed runs anywhere; the central 'generally improves in most cases' claim rests on single point estimates with deltas as small as 0.0015, so the empirical map may be noise.","rationale":"The reader's weakest assumption concerns unspecified Rouge-L thresholds and unvalidated proxy labels for short-form datasets; that is a genuine validity threat, and it should be fixed. However, the multiple-choice transfer pairs use exact-match labels and already show 16 of 18 improvements, so the Rouge-L issue alone cannot sink the abstract's broader 'most cases' claim. The absence of any uncertainty quantification is more load-bearing because it threatens every pair, including the exact-label MC results: the count of 'improved' pairs is based on point estimates that may be within run-to-run noise, and the conclusion's word 'significantly' is unsupported. The paper is otherwise an honest exploratory study with candid limitations, and the proposed re-run with seeds would directly test whether the observed pattern is real. Therefore the correct verdict remains CONDITIONAL—unchanged from the reader—but the stated condition should prominently include statistical reliability, not only label validation.","tokens_in":16517,"tokens_out":6019,"duration_ms":72125,"concrete_test":"Re-run the complete pipeline (all 36 transfer pairs, both LLMs, all three feature configurations, with and without data-agnostic features) using 20 independent random seeds that control both the train/test split and RandomForestRegressor(random_state=seed). For each transfer pair, compute the paired delta ΔAcc_i = Acc(+Δ)_i - Acc(-Δ)_i across seeds, then report the mean, 95% bootstrap CI, and a paired Wilcoxon signed-rank test against zero. The central claim survives only if the majority of pairs show a mean ΔAcc significantly greater than 0, and the 'certain scenarios degrade' claim only if at least one pair shows significantly less than 0. Also report the same analysis for ROC AUC.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—that adding data-agnostic features improves cross-dataset probe accuracy in most transfer pairs—is supported only by single point estimates in Tables 5–8, with no error bars, confidence intervals, repeated-seed runs, or significance tests. Many deltas are tiny: e.g., SWAG–MMLU (Llama2) Acc 0.5255→0.5270 (+0.0015) in Table 5, and MMLU–SWAG 0.5170→0.5170 (0.0000). The paper counts such pairs as 'improved' or 'not degraded' and concludes that probability/entropy metrics 'significantly enhance performance,' but the word 'significantly' has no statistical backing. The RandomForestRegressor is stochastic unless random_state is fixed (not stated), and the train/test split ratio and seed are not specified, so even larger deltas could arise from one unlucky or lucky draw. This concern applies to all 36 transfer pairs, including the exact-label multiple-choice datasets, so it is more load-bearing than the Rouge-L threshold issue: if the observed deltas are within run-to-run noise, the 'generally enhances in most cases' claim is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies whether supplementing hidden-state features with probability- and entropy-based \"data-agnostic\" features improves the cross-dataset generalization of supervised LLM uncertainty probes. It trains random forest regressors on hidden states from Llama2-7B and Mistral-7B across three tasks (multiple-choice QA and short-form QA), and compares hidden-state-only probes against hybrid feature sets, including a top-k feature-selection variant and a 5-layer variant. Across Tables 5-8 the paper reports that the hybrid features improve accuracy on most, but not all, transfer pairs, and a SHAP analysis suggests that data-agnostic features receive high feature importance in successful cases. The authors openly acknowledge limitations such as baseline drift, high misclassification rates, and the arbitrary choice of data-agnostic features, and they release code and complementary materials.","tokens_in":16733,"tokens_out":4875,"duration_ms":47904,"significance":"If the central empirical claim held, the paper would provide a low-cost, practical recipe for improving the cross-task generalization of hidden-state uncertainty probes, which is a recognized limitation in the literature. The paper has clear strengths: it releases code, reports both accuracy and ROC AUC, provides a feature-contribution analysis with SHAP, and explicitly acknowledges several limitations that other papers often omit. However, the headline claim rests on single point estimates without any measure of variability, and the labeling pipeline for short-form datasets is underspecified, so the significance of the reported effects cannot currently be assessed.","major_comments":[{"comment":"The central claim that adding data-agnostic features 'significantly enhances' generalization is based on single point estimates with no error bars, confidence intervals, repeated seeds, or significance tests. The RandomForestRegressor is stochastic unless a random state is fixed, and the manuscript does not state whether random_state was set or how the train/test split was chosen. Many reported deltas are tiny, e.g., SWAG–MMLU with Llama2-7B in Table 5 changes from 0.5255 to 0.5270 (+0.0015), and MMLU–SWAG changes by 0.0000; these are counted in the 'improved' or 'not degraded' totals. The word 'significantly' in the Conclusion therefore has no statistical backing. I would require repeated runs over multiple seeds, reporting of mean and variance (or confidence intervals), and ideally paired significance tests over transfer pairs before the qualitative 'generally improves in most cases' conclusion can be accepted.","section":"§3.6, §3.7, Tables 5-8, §6"},{"comment":"The binary labels for the short-form datasets (TriviaQA, SQuAD, Winogrande) are derived from Rouge-L, but the threshold used to convert a Rouge-L similarity score into a 0/1 label is never stated. Section 3.4 says only 'the results of Rouge-L will be the labels,' and Section 3.8 refers to a 'predefined threshold' without distinguishing the 0.5 accuracy threshold from the Rouge-L binarization threshold. Since these labels are used both to train the probes and to compute all reported metrics, an arbitrary or ill-documented threshold could materially change every number in Tables 7-8 and the conclusions drawn from them. The authors should specify the exact binarization procedure and, ideally, validate the proxy against human judgments or at least a labeled subset.","section":"§3.4, §3.8"},{"comment":"The summary statistics in the text are inconsistent with the reported tables. The text says '7 improved out of 9 for Llama2 – 7B on multiple choice question answering datasets; 8 improved out of 9 for Mistral – 7B on short-form question answering datasets; 9 improved out of 9 for Llama2 – 7B on multiple choice question answering datasets; and 9 improved out of 9 for Mistral – 7B on short-form question answering datasets.' However, Table 8 (Mistral-7B short-form) shows all 9 transfer pairs improving, not 8, and Table 6 (Mistral-7B multiple-choice) shows 8 improvements, which appears to be the intended '8 out of 9' entry. As written, the reader cannot verify the counts that support the abstract's 'generally enhances in most cases' claim; this needs correction and verification.","section":"§3.9"}],"minor_comments":[{"comment":"The notation paragraph says a probe trained on dataset A and tested on dataset B is denoted 'B-A', but the table headers use the reverse form (e.g., 'MMLU–RACE' in Table 1), which is confusing. The direction convention should be stated unambiguously and used consistently.","section":"§3.9"},{"comment":"The rows in Table 16 appear not to be sorted in descending order as in the other SHAP tables: feature 1924 has mean SHAP 0.000260, feature 1233 has 0.000092, and feature 2172 has 0.000894, which breaks the monotonic ordering. Please check the table construction.","section":"Table 16"},{"comment":"The first bullet in the Limitation section says the authors 'randomly select the data-agnostic features,' which is misleading because the features are actually predetermined by the method (probabilities and entropies). I suggest rewording to 'arbitrarily selected' or 'chosen without an ablation.'","section":"§7"},{"comment":"The citation 'Liu et al. (2024)' in the sentence about using the last token should be 'Liu et al. (2024b)' to disambiguate from Liu et al. (2024a) in the reference list.","section":"§3.3"},{"comment":"There is a typo in the subsection heading: 'Compreshension' should be 'Comprehension.'","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is an honest empirical study with openly acknowledged limitations, but the central quantitative claim is not currently supported by the evidence because of the absence of repeated-seed experiments and significance testing. The missing Rouge-L threshold is also a correctness issue for the short-form datasets. Both are fixable within the manuscript's scope, so I do not recommend rejection, but the authors should be required to supply the missing statistical details and to rerun or re-analyze the experiments with multiple seeds before the claim 'significantly enhances' can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is an honest, exploratory empirical study with a real but modest novelty. The idea of adding data-agnostic features (token probabilities, entropies) to hidden-state probes for cross-task generalization is a sensible extension of Liu et al. 2024b and He et al. 2024, and the paper gives a systematic map: two LLMs, six datasets, three task types, and multiple feature configurations. The SHAP-based explanation for why some transfer pairs fail is a nice touch, and the limitations section is unusually candid, acknowledging baseline drift, random feature selection, and the underweighting of data-agnostic features.\n\nThe soft spot is load-bearing: there is no variance information anywhere. Every accuracy in Tables 5–8 is a single point estimate from one run of a stochastic random forest, with no stated seed or train/test split ratio. Some deltas are tiny (SWAG–MMLU at +0.0015, MMLU–SWAG at 0.0000), so the paper's central claim that data-agnostic features 'generally enhance' performance could easily be run-to-run noise. The conclusion's word 'significantly' has no statistical backing, and the stress-test concern about this is fair. A second issue: the Rouge-L threshold for converting scores into binary labels for short-form datasets is never specified, so those labels are not auditable. Both problems are addressable in revision, but until then the empirical map is a set of hypotheses rather than a stable result.\n\nThere is also a minor editorial slip in the results section where the counts of improved pairs mix up model/dataset combinations; the tables themselves are clear, but the text should be fixed.\n\nWho is this for? Researchers working on LLM uncertainty estimation via internal states. They will find the experimental map a useful starting point, but they should treat the point estimates as preliminary and should not rely on the specific deltas. I would not cite the directional claims in my own work until the variance question is resolved.\n\nRecommendation: send it to peer review, but with the clear expectation that the authors add repeated seeds, error bars, and a significance test, and that they specify the Rouge-L threshold and the data split. The paper is worth refereeing because the question is timely and the authors have been transparent about limitations; the missing statistical rigor is fixable without changing the experimental design.","headline":"Honest empirical map, but the central claim is not yet established because every number is a single point estimate with no variance.","tokens_in":17273,"tokens_out":4162,"would_cite":false,"duration_ms":39593,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding probability and entropy features improves LLM uncertainty probes' cross-task generalization in most settings, but gains depend on the probe giving those features enough weight.","keywords":["uncertainty quantification","hidden-state probes","data-agnostic features","out-of-domain generalisation","large language models","feature selection","SHAP values","hallucination detection"],"falsifier":"Re-run the short-form transfer experiments with a stated, validated Rouge-L threshold (or with human-annotated correctness labels) and compare the accuracy gaps between hidden-state-only and hybrid probes; if the gaps disappear or reverse, the central claim is an artifact of the label proxy. A second, mechanism-specific test: on a failing commonsense pair like SWAG-to-RACE under Llama2, train a probe on the data-agnostic features alone; if underweighting is the cause, this probe should generalize at least as well as the hybrid probe on that pair.","tokens_in":16284,"feed_emoji":"🎯","tokens_out":7911,"duration_ms":68577,"temperature":0.7,"pith_summary":"Large language models can be confidently wrong, so a reliable estimate of whether an answer is factually correct matters. The paper asks whether a cheap, task-independent feature augmentation can make uncertainty probes—classifiers trained on the model's internal hidden states—generalize to new tasks and datasets. It trains random-forest probes on hidden states, with and without data-agnostic features (token probabilities and entropy), and measures cross-dataset accuracy, ROC AUC, and calibration. The central finding is that adding data-agnostic features improves out-of-domain generalization in most transfer settings, particularly on short-form question answering, but degrades performance in a minority of cases. The paper attributes those failures to the probe underweighting the data-agnostic features relative to hidden-state features, and shows that pruning hidden states to the most informative dimensions tends to amplify the contribution of the data-agnostic signals.","feed_headline":"Probability features boost LLM uncertainty probes' cross-task accuracy","feed_subtitle":"Adding cheap token-probability and entropy signals helps hidden-state probes generalize to new tasks—but only when the probe weights them…","key_machinery":"The machinery is the hybrid feature set fed to a random forest regressor: the 1-layer (or 5-layer) hidden-state vector of the last answer token, plus a small set of data-agnostic features that do not depend on the dataset—for multiple-choice questions, the four softmax probabilities over A/B/C/D sorted descending and their entropy; for short-form answers, average and maximum per-token negative log-probability and average and maximum per-token entropy. The paper also uses top-300 hidden-feature selection by absolute Pearson correlation with the correctness label, and Tree SHAP values to measure how much each feature contributes to the probe's predictions. The data-agnostic features are the lever: the paper's hypothesis is that they carry task-independent signal, so when the probe assigns them high weight they improve out-of-domain transfer.","core_discovery":"The paper's claim is that the generalization failure of hidden-state uncertainty probes is partly a feature-competition problem, not an information problem. When a random forest probe is trained on 4,096 hidden-unit features plus a handful of data-agnostic statistics (sorted token probabilities and entropy for multiple-choice; average and maximum negative log-probability and entropy for short-form answers), the hybrid probe usually transfers better to an unseen dataset than the hidden-state-only probe: in the reported tables, 7/9 to 9/9 of transfer pairs improve, with the largest accuracy gains on short-form benchmarks. The improvement is not universal—some commonsense-reasoning pairs such as SWAG-to-RACE get worse—and the paper identifies the cause in SHAP feature-importance values: in failing cases the data-agnostic features are not ranked at the top, or are intermixed with lower-ranked hidden features, so the probe underweights them. Selecting the top-300 hidden features by absolute Pearson correlation with the labels generally amplifies the data-agnostic features' importance and yields the expected ordering of gains (selected greater than one-layer greater than five-layer) in most, but not all, transfer pairs. The paper concludes that probability- and entropy-based features significantly enhance generalization, especially on short-form QA, while noting that the probe still underperforms on commonsense transfer pairs where data-agnostic features are not prioritized.","pith_inferences":["An implicit, testable recipe follows: before deploying a hybrid probe on a new target domain, one could inspect SHAP rankings of the data-agnostic features; if they are underweighted, reweighting them or increasing their number might recover the generalization gain the paper reports.","The paper leaves the Rouge-L binarization threshold unspecified; a natural extension is to sweep that threshold or replace the automatic labels with human judgments to see whether the reported generalization differences persist under cleaner targets.","The asymmetric transfer pattern (e.g., RACE-to-SWAG improves while SWAG-to-RACE degrades) suggests that data-agnostic features interact with the training distribution; an extension would be to test whether the direction of the asymmetry flips when the source and target datasets are swapped systematically across all three tasks.","Because the mechanism is feature weighting rather than new data, one could try to make the probe explicitly regularize hidden-state features (e.g., by penalizing their variance) to force the data-agnostic features to carry more weight, which the paper does not attempt."],"forward_implications":["If the finding holds, uncertainty probes can be made more portable without collecting additional training data from multiple tasks, since data-agnostic features are computed from the model's own outputs.","On short-form question answering benchmarks, the hybrid feature set yields the largest generalization gains, so this is a promising direction for open-ended generation where labels are harder to obtain.","Feature selection that keeps only the most label-correlated hidden dimensions amplifies the relative importance of data-agnostic features and usually, though not always, produces larger gains than using the full hidden-state vector.","Commonsense-reasoning transfer pairs remain the failure mode: when data-agnostic features do not rank at the top of SHAP importance, adding them can hurt, so the benefit is conditional on feature weighting rather than guaranteed.","The reported accuracy improvements are not calibration improvements; ECE values stay in a narrow range, so the probe's reliability ordering changes without materially changing calibration."],"supporting_citations":[{"why":"Backbone probe design: a random forest regressor trained on middle-layer hidden states of the last answer token, plus the five grey-box features used for multiple-choice datasets.","marker":"Liu et al. (2024b)"},{"why":"Source of the four short-form data-agnostic features: average and maximum negative log-probability and entropy over generated tokens.","marker":"Manakul et al. (2023)"},{"why":"Introduces the claim that data-agnostic features improve generalization because they are task-independent, which this paper extends to cross-task transfer.","marker":"He et al. (2024)"},{"why":"Provides evidence that hidden states are task-specific and that generalization happens within similar skills; also reports the same asymmetric transfer pattern.","marker":"Orgad et al. (2024)"},{"why":"Baseline alternative that trains on multiple datasets to find a universal truthfulness hyperplane; this paper contrasts its single-dataset hybrid approach against it.","marker":"Liu et al. (2024a)"},{"why":"Establishes the trained-probe approach on hidden states and the 0.5 confidence threshold used for accuracy evaluation.","marker":"Azaria and Mitchell (2023)"},{"why":"Defines the Rouge-L metric used to binarize correctness labels for short-form question answering.","marker":"Lin (2004)"}],"fun_headline_variants":["Cheap probability signals help LLM probes transfer—not always","Data-agnostic features boost uncertainty probes—when weighted right","Probability features improve LLM probe transfer—most of the time","Feature weighting decides if probability signals boost probe generalization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the correctness labels for short-form answers, obtained by binarizing Rouge-L similarity scores, are accurate enough to train on; the paper never states the cutoff used to turn a Rouge-L score into a 0/1 label or checks those labels against human judgment, so noisy targets could drive the measured generalization differences.","fun_headline_variants_meta":{"raw":{"variants":["Cheap probability signals help LLM probes transfer—not always","Data-agnostic features boost uncertainty probes—when weighted right","Probability features improve LLM probe transfer—most of the time","Feature weighting decides if probability signals boost probe generalization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001568,"raw_usage":{"total_tokens":6352,"prompt_tokens":1128,"completion_tokens":5224,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":744,"completion_tokens_details":{"reasoning_tokens":5156}},"tokens_in":744,"tokens_out":5224,"duration_ms":39644,"temperature":1.0,"reasoning_tokens":5156,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:57:07.182002+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the short-form transfer experiments with a stated, validated Rouge-L threshold (or with human-annotated correctness labels) and compare the accuracy gaps between hidden-state-only and hybrid probes; if the gaps disappear or reverse, the central claim is an artifact of the label proxy. A second, mechanism-specific test: on a failing commonsense pair like SWAG-to-RACE under Llama2, train a probe on the data-agnostic features alone; if underweighting is the cause, this probe should generalize at least as well as the hybrid probe on that pair.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Rouge-L metric used to binarize correctness labels for short-form question answering."}],"review_version":1}