{"id":"2f68b94b-cb36-40d4-9c2d-024391fab833","arxiv_id":"2506.17748","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A training-free, single-generation hallucination detector based on HSIC dependence between hidden states of input and output tokens outperforms single-pass baselines and approaches multi-pass accuracy.","lead":"Researchers propose HIDE, a single-pass way to catch language model hallucinations by measuring how statistically independent the model's internal representations of the question and its answer are. Across four question answering datasets and six open models, HIDE beats other single-pass detectors by about 29 percent AUC on average and matches slower multi-pass methods while using roughly half the compute.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The core weakness is that the HIDE score is computed from independently ranked input/output tokens paired only by rank, so HSIC is not estimating the input-output statistical dependence that the hypothesis requires.","rationale":"The Pith reader's weakest_assumption is the same pairing/measurement issue, and I agree it is the load-bearing point. The empirical evaluation is extensive: six models, four datasets, multiple correctness metrics, and public code. That supports the method as a detection heuristic, and I would not reject the paper. But the central claim is not merely 'HIDE works'; it is that HIDE works because hallucinations correspond to statistical decoupling between input and output representations. Without a defined joint distribution or a principled pairing, \\HSICHIDE has no proven connection to HSIC, so Lemma 1 cannot license the decoupling interpretation. The proposed permutation/control experiment would distinguish several possibilities: (1) the score is sensitive to pairing/input, validating or at least requiring justification of the construction; (2) the score is insensitive to input, breaking the mechanism but leaving an output-only heuristic; (3) the score is sensitive to input but insensitive to pairing, showing that rank alignment is irrelevant while input identity matters. Any of these outcomes would sharpen the paper. The threshold-fitting issue in Section 6.4 is also real but does not affect the headline AUC comparisons, so it is secondary. Therefore, the reader's CONDITIONAL verdict remains apt; no verdict change is needed.","tokens_in":32501,"tokens_out":7616,"duration_ms":83188,"concrete_test":"On SQuAD and NQ with Llama-3-8B, rerun the HIDE pipeline from Algorithm 1 under two perturbations and compare AUCs to Table 2: (i) randomly permute the order of the KeyBERT-selected output tokens before pairing with the input ranks (50 seeds); (ii) replace the input keyword set with keywords extracted from a randomly chosen different prompt, keeping the output fixed. If AUCs under either perturbation stays within a small margin (e.g., 1-2 AUC points) of the reported values, the score is not measuring the claimed input-output dependence, and the mechanism must be revised; if AUCs drops substantially under (ii), the input is load-bearing, but the rank-pairing in (i) still needs a principled justification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.4.1 and Algorithm 1 construct X and Y by separately selecting the top-neff KeyBERT keywords from the input and output and pairing them by rank. HSIC (Section 3.1, Lemma 1) is a valid measure of statistical dependence only when (x_i, y_i) are samples from a joint distribution P_XY. No such joint distribution is defined over input token types and output token types for a single generation: the two sets are selected marginally and the rank pairing imposes an arbitrary correspondence. Consequently, \\HSICHIDE in Equation (4) is not an estimator of the HSIC between the LM's input and output representations, and the paper's central explanation - that low scores indicate 'statistical decoupling' - does not follow from Lemma 1. The authors call the score a 'heuristic' in Section 4.2, but the abstract and Section 6 still present the decoupling hypothesis as the mechanism. This does not by itself refute the empirical AUC results, but it means the method's theoretical foundation and the interpretability claim are unsupported as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Chatterjee, Goel, and Chakraborty propose HIDE, a single-pass and training-free hallucination detector. For an input prompt and one LM-generated output, HIDE selects a budget of salient tokens from the input and output via KeyBERT, extracts their hidden states from a chosen decoder layer, and computes an adapted HSIC-like score between the two representation sets. The score is thresholded to flag hallucinations. The paper's stated mechanism is that hallucinations correspond to a 'statistical decoupling' between the LM's internal representations of input context and generated output. The evaluation covers SQuAD, RACE, NQ, and TriviaQA with six Llama and Gemma models and compares HIDE against perplexity, energy, LN-entropy, lexical similarity, and Eigenscore under sentence-similarity, ROUGE-L, and exact-match correctness measures. The main reported results are an average ~29% relative AUC-ROC improvement over the best single-pass baseline, an average ~3% relative improvement over multi-pass methods, and ~51% lower computation time than multi-pass methods. Ablations examine token budget, layer choice, kernel choice, HSIC estimator, and token-selection strategy.","tokens_in":32708,"tokens_out":14316,"duration_ms":134526,"significance":"The empirical evaluation is broad and the AUC improvements over single-pass baselines are consistent, which makes HIDE a potentially useful practical contribution if the findings hold. The paper also ships open-source code and reports threshold-independent ranking metrics (AUC-ROC, PCC), which are not affected by threshold tuning. The main weakness is that the theoretical foundation does not support the stated mechanism: the HIDE score is computed from independently selected input/output tokens paired by rank, without a joint distribution, so Eq. (4) is not an HSIC estimate of input-output dependence in the sense of Lemma 1. The empirical score may still discriminate hallucinations as a heuristic, but the 'statistical decoupling' claim and the interpretability story need to be either made rigorous or explicitly downgraded. The threshold selection in Section 6.4 also uses evaluation labels, and some per-setting comparisons with multi-pass methods are overstated.","major_comments":[{"comment":"Definition 2 and Section 4.4.1 construct the samples X and Y by independently selecting the top-neff KeyBERT tokens from the input and output and pairing them by rank order. Lemma 1 guarantees HSIC=0 iff independence only when (x_i,y_i) are drawn from a joint distribution P_XY; for a single generation no such joint distribution over input token types and output token types is defined, and the rank pairing is an arbitrary coupling. Definition 2 is also not well-defined for repeated token types, since the same token type can occur with different hidden states (Appendix C shows duplicated selected tokens). Consequently the score in Eq. (4) is not an estimator of HSIC between the LM's input and output representations, and the central claim that low scores indicate 'statistical decoupling' does not follow from the theory presented. The paper itself calls the score a 'heuristic' in Section 4.2, yet the abstract and Section 6 present the decoupling hypothesis as the mechanism. The same objection applies to the SVD alignment strategy in Section 4.4.1, where rows of independently projected matrices are paired by singular-direction order. Please either define a genuine input-output coupling (for example, attention-based token pairing) with a corresponding joint distribution, or explicitly reframe HIDE as an empirical heuristic and remove the HSIC/independence interpretation from the abstract and conclusions.","section":"Section 4.1, Section 4.4.1, Algorithm 1"},{"comment":"The binary decision threshold is tuned on the evaluation labels. In Section 6.4, for each (model, dataset) pair the threshold tau is chosen by maximizing G-Mean using exact-match ground-truth labels, and the average value tau=0.12 is then recommended as an operating point. This is test-label leakage for the binary decision procedure: any accuracy, G-Mean, or deployment-oriented statement built on this threshold is not an unbiased estimate of performance. The AUC-ROC and PCC results are threshold-independent and therefore retain their validity, but the threshold analysis should be performed on a held-out split (or the threshold should be derived from a separate calibration set), and the practical binary-decision claims should be limited accordingly.","section":"Section 6.4, Table 3"},{"comment":"Section 6.2 says that on TriviaQA Eigenscore achieves 'marginally higher AUC-ROC and PCC values,' but Table 2 shows much larger gaps in four of the six models: for Llama-3.2-3B the AUCs values are 76.13 (Eigenscore) versus 58.36 (HIDE); for Llama-3.2-3B-Instruct, 78.35 versus 58.43; for Llama-3-8B, 81.06 versus 65.65; and for Llama-3-8B-Instruct, 82.83 versus 61.70. The aggregate ~3% relative improvement over multi-pass methods is therefore driven by gains on other datasets, and the abstract's 'competitive with multi-pass methods' claim needs qualification by dataset and by output length, especially for short-answer factuality tasks.","section":"Section 6.2, Table 2"}],"minor_comments":[{"comment":"The proof relies on the unquantified approximation HSICHIDE roughly equal to (1 - 3/n) times HSICu and does not bound the difference; the asymptotic bias and consistency statements may be true, but the presented argument is not yet a proof. Please replace it with a precise rate-of-convergence argument.","section":"Appendix A, Lemma 4"},{"comment":"There are several typos and formatting issues: 'the the' in Section 1, 'actuality hallucinations' in Section 6.2, 'repeatation' in Section 8.2, and 'LLama' in the Figure 6 caption.","section":"Throughout"},{"comment":"Under exact-match labels, Energy outperforms HIDE in some settings (for example, Gemma-2-9B-Instruct on SQuAD: 81.68 versus 72.62); the main text's 'outperforms other single-pass methods in almost all settings' claim should state which correctness measure it refers to.","section":"Table B.1"},{"comment":"The score is described as an 'adapted variant of unbiased HSIC,' but Definition 3 is explicitly biased; consider calling it an HSIC-based score throughout to avoid confusing readers.","section":"Figure 1, Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The empirical core is worth preserving, and the AUC-based results are threshold-independent, so I do not see a reason to reject outright. However, the main theoretical flaw is substantial: the HIDE score is not an HSIC estimate of input-output dependence under the construction in Algorithm 1, and the paper's central mechanism claim therefore needs either a genuine coupling definition or an explicit downgrade to a heuristic. The threshold leakage and the TriviaQA overstatement are also fixable, but together they require a careful revision of Sections 4, 6, and the abstract."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, HIDE is a genuinely useful empirical contribution: a single-pass, training-free detector that applies an HSIC-like score to hidden states of selected input and output tokens, and it beats perplexity/energy baselines consistently across six LMs and four QA datasets while staying competitive with multi-pass methods at about half the compute. Second, the paper's theoretical framing is overreaching. The HSIC machinery from Lemma 1 assumes paired samples from a joint distribution, but the method constructs X and Y by independently selecting top-neff keywords from input and output and pairing them by rank. That does not give you samples from any defined joint distribution, so the HIDE score is not actually estimating input-output statistical dependence, and the 'statistical decoupling' hypothesis as stated is not supported by the math. The authors do call the score a 'heuristic' in Section 4.2, but the abstract and introduction lean on the mechanism as if it were established.\n\nWhat the paper does well: the evaluation is broad and careful. Six open models, four datasets covering both faithfulness and factuality, multiple correctness metrics (exact match, ROUGE-L, sentence similarity), and two token-selection strategies (keyword and SVD) that behave nearly identically. The ablations on layer, kernel, and token budget are thorough, and the layer-agnostic result is a useful practical finding. The code is open, and the compute time comparison is concrete. The empirical AUC improvements are consistent and threshold-independent, which is the right kind of evidence for a detection method.\n\nThe soft spots, in order of weight. First, the pairing problem I mentioned. This is not merely cosmetic: it means the central interpretability claim does not follow from the cited theory. It does not refute the empirical AUC results, but the paper should either construct a proper paired sampling scheme or reframe itself as presenting an empirically motivated score, not an HSIC estimator. Second, the binary decision threshold tau = 0.12 is tuned on the test set (Section 6.4 maximizes G-Mean on exact-match labels). The AUC rankings are fine, but the binary-decision results are optimistic as a deployment estimate. A held-out threshold or per-dataset cross-validation would fix this. Third, no uncertainty estimates or variance across runs. That matters less given the scale, but it would strengthen the comparison.\n\nThe paper is honest about its limitations—Section 8 openly discusses failure on single-token outputs and prompt copying—which I appreciate. This is a paper for people actively working on hallucination detection, not for a general theory audience. It deserves a serious referee. I would recommend major revision, with the main ask being to fix or de-emphasize the HSIC-dependence claim and to report threshold selection honestly. The empirical core is worth keeping.","headline":"Solid empirical method for single-pass hallucination detection, but the HSIC decoupling story rests on a pairing assumption that doesn't hold; the AUC results stand anyway.","tokens_in":792,"tokens_out":1102,"would_cite":false,"duration_ms":30000,"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":"This paper claims that hallucinated LLM outputs can be caught in a single pass by measuring when a model's internal representation of its input and of its generated answer drift apart, and introduces HIDE, a training-free score that does…","keywords":["hallucination detection","language models","hidden states","HSIC","statistical dependence","single-pass detection","faithfulness hallucination","factuality hallucination"],"falsifier":"Shuffle the order of the selected output tokens relative to the selected input tokens and recompute the HIDE score on the same examples; if the score still separates hallucinated from faithful outputs, then it is not measuring input-output dependence between paired tokens, and the decoupling mechanism is not what carries the result.","tokens_in":32281,"feed_emoji":"🧠","tokens_out":7048,"duration_ms":66863,"temperature":0.7,"pith_summary":"The paper aims to show that a language model's hallucinated answers can be identified in a single generation by measuring how much its internal representation of the output has come loose from its representation of the input. It introduces HIDE, which computes a kernel-based statistical-dependence score, HSIC, between hidden-state vectors of salient prompt tokens and salient generated tokens; a low score signals that the output is decoupled from the context and therefore likely hallucinated. Across four question-answering benchmarks and six open-source language models, HIDE beats other single-pass detectors by roughly 29 percent average relative AUC-ROC and matches or slightly exceeds multi-pass methods while using about 51 percent less computation time. The practical appeal is that hallucination detection becomes cheap enough for real-time use without extra training or repeated sampling.","feed_headline":"Hallucinated LLM answers show a measurable decoupling inside the model","feed_subtitle":"A single-pass score gains ~29% relative AUC-ROC and runs at ~51% less compute than multi-pass methods.","key_machinery":"The load-bearing object is HSIC, the Hilbert-Schmidt Independence Criterion: a kernel-based measure that is zero exactly when two random variables are independent, estimated here from the hidden states of about twenty semantically salient input tokens and twenty salient output tokens taken from a middle decoder layer. The paper rescales the standard unbiased estimator's denominators to produce a bounded, numerically stable score $\\widehat{\\text{HSIC}}_{\\text{HIDE}}$ that remains well-defined even for one or two tokens, is asymptotically unbiased, and separates faithful from hallucinated outputs near a threshold of about 0.12. The RBF kernel supplies the characteristic-kernel guarantee that lets a near-zero score be interpreted as genuine independence.","core_discovery":"The central claim is that hallucination is visible as a measurable statistical decoupling between the model's internal representations of its input and the output it generates. Faithful answers keep the hidden states of selected input tokens and selected output tokens statistically dependent, while hallucinated answers let that dependence drop toward zero. HIDE quantifies this with an adapted HSIC estimator built from RBF-kernel Gram matrices of the selected hidden states, and flags an output as hallucinated when the resulting score falls below a threshold. The evidence offered is that HIDE outperforms the best single-pass baseline by an average relative improvement of about 29 percent in AUC-ROC across four datasets and six models, is competitive with multi-pass methods at about 3 percent relative improvement, and costs roughly 51 percent less computation time.","pith_inferences":["Editorial inference: the paper never tests the pairing assumption directly; the selected input and output tokens are chosen independently and matched only by rank order, so the score's empirical power may partly come from confounds such as lexical overlap or output length rather than from statistical dependence between paired representations.","A testable next step is to shuffle the order of the selected output tokens before computing HSIC; if the score still separates hallucinated from faithful outputs, then the decoupling story is not what carries the result.","The stability of the optimal threshold around 0.12 across datasets suggests a calibration-free deployment, but those thresholds were fit using ground-truth labels, so a fair test is to fix the threshold on one benchmark and measure transfer to unseen domains.","The decoupling hypothesis predicts a causal signature: interventions that re-ground generation, such as context-aware decoding, should raise the HIDE score on the same prompts; measuring that would connect the detector to the mechanism it claims to exploit."],"forward_implications":["If HIDE works as claimed, hallucination detection no longer requires multiple generations per input, removing the main latency barrier to real-time detection.","With roughly 29 percent average relative AUC-ROC gain over the best single-pass baseline, a HIDE-style score could replace or augment perplexity- and energy-based checks in production question-answering systems.","Because performance stays nearly flat across decoder layers and across most kernel choices, deployment needs little per-model tuning; a token budget near 15 to 20 and a threshold near 0.12 act as near-optimal defaults.","The method's known failure modes are short single-token answers, which produce an uninformative zero score, and outputs that copy the prompt verbatim, which inflate the score and can mask a hallucination.","The method is white-box and training-free, so any model with accessible hidden states can use it without additional data or fine-tuning."],"supporting_citations":[{"why":"Supplies the definition of HSIC and the lemma that HSIC is zero if and only if the variables are independent under characteristic kernels.","marker":"Gretton et al. (2005)"},{"why":"Supplies the unbiased V-statistic estimator for HSIC whose denominators HIDE rescales to form its adapted score.","marker":"Song et al. (2012)"},{"why":"Supplies the characteristic-kernel theory, via injective mean embeddings, that makes the RBF-based zero score meaningful.","marker":"Sriperumbudur et al. (2010)"},{"why":"Establishes that the RBF kernel is characteristic on R^d, grounding Lemma 2.","marker":"Sriperumbudur et al. (2008)"},{"why":"Provides the Eigenscore multi-pass baseline that HIDE must match, and notes that perplexity works well for short TriviaQA-style answers.","marker":"Chen et al. (2024a)"},{"why":"Provides the lexical-similarity consistency baseline used for comparison.","marker":"Lin, Liu, and Shang (2022)"},{"why":"Provides the length-normalized entropy multi-pass baseline used in the evaluation.","marker":"Malinin and Gales (2021)"},{"why":"Provides the energy-based single-pass baseline that HIDE is compared against.","marker":"Liu et al. (2020)"}],"fun_headline_variants":["Single-pass detection of LLM hallucinations via representation decoupling","LLM hallucinations betray themselves in hidden-state stats","HIDE detects LLM hallucinations in one pass, 51% less compute","Hallucination detection via internal decoupling: 29% AUC-ROC gain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method stands on the assumption that the tokens picked from the prompt and the tokens picked from the answer can be treated as paired samples of the same underlying relationship, but the selection procedure chooses the two sets independently and pairs them only by rank order.","fun_headline_variants_meta":{"raw":{"variants":["Single-pass detection of LLM hallucinations via representation decoupling","LLM hallucinations betray themselves in hidden-state stats","HIDE detects LLM hallucinations in one pass, 51% less compute","Hallucination detection via internal decoupling: 29% AUC-ROC gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000423,"raw_usage":{"total_tokens":2198,"prompt_tokens":1000,"completion_tokens":1198,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":1124}},"tokens_in":616,"tokens_out":1198,"duration_ms":8839,"temperature":1.0,"reasoning_tokens":1124,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:01:43.740128+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Shuffle the order of the selected output tokens relative to the selected input tokens and recompute the HIDE score on the same examples; if the score still separates hallucinated from faithful outputs, then it is not measuring input-output dependence between paired tokens, and the decoupling mechanism is not what carries the result.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the definition of HSIC and the lemma that HSIC is zero if and only if the variables are independent under characteristic kernels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the unbiased V-statistic estimator for HSIC whose denominators HIDE rescales to form its adapted score."},{"cited_title":"Lanckriet","cited_arxiv_id":null,"evidence_quote":"Supplies the characteristic-kernel theory, via injective mean embeddings, that makes the RBF-based zero score meaningful."},{"cited_title":"Gretton, K","cited_arxiv_id":null,"evidence_quote":"Establishes that the RBF kernel is characteristic on R^d, grounding Lemma 2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the length-normalized entropy multi-pass baseline used in the evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the energy-based single-pass baseline that HIDE is compared against."}],"review_version":2}