{"id":"e9775ae6-c6bc-47f5-8544-d25d2f030fdf","arxiv_id":"2505.07528","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"SEReDeEP replaces ReDeEP's lexical scores with semantic entropy probe scores, but its claimed 3-10% accuracy gains are contradicted by its own tables and its probes are trained on the evaluation datasets.","lead":"SEReDeEP combines semantic entropy probes with ReDeEP's decoupling of external context and internal knowledge to score hallucinations in retrieval-augmented language models. The paper reports 3-10% accuracy gains over ReDeEP and SEP, but the evaluation is contaminated by training the probes on the same datasets used for testing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed >3% gain over ReDeEP is contradicted by the paper's own Table 1, indicating evaluation leakage or overfitting.","rationale":"The reader's weakest assumption was that training probes on RAGTruth/HalluRAG and then testing on the same datasets leads to inflated gains; my analysis confirms that concern is real, and the paper's own Table 1 shows the headline >3% claim is false even on the authors' numbers. This is internally inconsistent, not a matter of consensus. I also found additional support for the reader's concern in Section 4.2, where head/layer/coefficient choices are optimized per dataset, making the comparison to ReDeEP non-rigorous. Given the claim fails on its own reported numbers and the evaluation setup leaks training information, the verdict should remain REJECT.","tokens_in":18967,"tokens_out":1437,"duration_ms":10462,"concrete_test":"Rerun the SEReDeEP evaluation with the semantic entropy probes trained on the original SEP probe training data (or on held-out corpora), and use a single fixed head/layer/coefficient configuration per model across both datasets, then recompute Table 1. If the accuracy deltas versus ReDeEP remain below 3% in absolute terms, or if the configuration change drops accuracy below the ReDeEP baseline, the paper's headline claim and the superiority of the proposed method are both falsified.","verdict_should_be":"REJECT","load_bearing_attack":"The central quantitative claim (Section 4.3: 'accuracy improvements exceeding 3% compared to ReDeEP') is not supported by Table 1. Computing exact deltas: LLaMA3-7B RAGTruth ACC 0.8601 vs 0.8323 = +2.78%; LLaMA3-7B HalluRAG ACC 0.8478 vs 0.8596 = -1.18%; Mistral-7B RAGTruth +2.00%; Mistral-7B HalluRAG +2.40%; Qwen2.5-7B RAGTruth +1.51%; Qwen2.5-7B HalluRAG +2.73%. All deltas are below +3%, and one is negative, so the headline claim fails on the paper's own numbers. The evaluation protocol compounds this: Section 4.1 states that 'These datasets' original outputs served as critical training material for our semantic entropy probes', so the probes are trained on RAGTruth and HalluRAG responses and then evaluated on the same datasets. Section 4.2 further selects per-dataset copy heads and FFN layers (LLaMA3-7B: top 1 head/top 2 layers on RAGTruth vs top 1 head/top 10 layers on HalluRAG; Mistral-7B: top 3 vs top 5 heads; Qwen2.5-7B: top 7 vs top 10 FFN layers) and per-dataset regression coefficients (beta 0.2 vs 0.4 vs 1.0 vs 1.6 vs 0.8 vs 1.2), i.e., the evaluation is not independent. With the central claim unsupported and the evaluation protocol unable to rule out train/test leakage, the reported gains are not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SEReDeEP, a modification of the ReDeEP hallucination detector for retrieval-augmented generation. It replaces ReDeEP's External Context Score (ECS) and Parametric Knowledge Score (PKS) with semantic-entropy-probe-derived External Context Entropy (ECE) and Parametric Knowledge Entropy (PKE), then combines them through linear regression. The authors evaluate on RAGTruth and HalluRAG across LLaMA3-7B, Mistral-7B, and Qwen2.5-7B, reporting accuracy, AUC, F1, and recall against ten baselines, and claim hallucination-prediction accuracy improvements exceeding 3% over ReDeEP and surpassing 10% over SEP.","tokens_in":19305,"tokens_out":4966,"duration_ms":41868,"significance":"The conceptual direction is reasonable and potentially useful: semantic entropy probes are cheap, and making ReDeEP's mechanistic scores semantically robust is a sensible goal. If the gains were validated out-of-sample, this would be a practical contribution to RAG hallucination detection. However, the paper's own numbers contradict the headline improvement claim, and the evaluation protocol is in-sample: the probes are trained on the same datasets used for evaluation, and coefficients, copy heads, and FFN layers are selected separately per dataset. The reported results therefore do not establish the claimed advantage.","major_comments":[{"comment":"The central quantitative claim is not supported by Table 1. Direct ACC deltas of SEReDeEP over ReDeEP are: LLaMA3-7B RAGTruth +2.78% (0.8601 vs 0.8323); LLaMA3-7B HalluRAG -1.18% (0.8478 vs 0.8596); Mistral-7B RAGTruth +2.00% (0.8257 vs 0.8057); Mistral-7B HalluRAG +2.40% (0.8697 vs 0.8457); Qwen2.5-7B RAGTruth +1.51% (0.8135 vs 0.7984); Qwen2.5-7B HalluRAG +2.73% (0.8598 vs 0.8325). None of these reaches the claimed 'exceeding 3%', and one is negative; on LLaMA3-7B HalluRAG the AUC is also essentially unchanged (0.7631 vs 0.7623). The abstract's '3%-10%' and Section 4.3's 'exceeding 3%' are therefore contradicted by the paper's own reported results.","section":"Abstract and Section 4.3, Table 1"},{"comment":"The evaluation is in-sample. Section 4.1 states that RAGTruth and HalluRAG 'original outputs served as critical training material for our semantic entropy probes', and Section 4.2 selects regression coefficients (alpha, beta) and copy-head/FFN-layer configurations separately for each model-dataset pair, e.g., LLaMA3-7B top 2 vs top 10 FFN layers, Mistral-7B beta 1.0 vs 1.6, and Qwen2.5-7B beta 0.8 vs 1.2. Because the same dataset labels inform probe training, coefficient fitting, and architectural-component selection, Table 1 measures fitted in-sample behavior rather than predictive accuracy. An out-of-sample evaluation with cross-validation or a held-out test split and fixed hyperparameters is required before any improvement claim can be supported.","section":"Sections 4.1 and 4.2"},{"comment":"The text asserts 'statistically significant performance improvements' but reports no confidence intervals, standard errors, significance tests, or repeated-run variability. Given that several accuracy deltas are small (some below 0.005 in absolute terms, e.g., LLaMA3-7B RAGTruth AUC is 0.7824 for SEReDeEP versus 0.7829 for ReDeEP), this assertion is unverifiable from the data presented and should be removed or replaced with proper uncertainty quantification.","section":"Section 4.3"}],"minor_comments":[{"comment":"There is a typo: 'SEDeDeEP' should be 'SEReDeEP'.","section":"Section 4.6"},{"comment":"The figure captions contain raw unicode artifacts (e.g., '/uni00000014/...'), making part of the text unreadable; these should be repaired.","section":"Figure 4 captions"},{"comment":"The ablation rows use unexplained symbols '!' and '%' without a legend; please state explicitly which symbol corresponds to PKE and which to ECE.","section":"Table 2"},{"comment":"The text contains a sentence fragment: 'Beyond their primary evaluation function.Neither RAGtruth nor HalluRAG...' should be two complete sentences.","section":"Section 4.1"},{"comment":"The notation 'ECE11' and 'PKE12' is used without defining the numeric subscripts; clarify whether these refer to layer indices.","section":"Section 4.4"},{"comment":"Reference [52] attributes Qwen2.5 to 'Mistral AI team'; the correct attribution should be to the Qwen team, and the reference format should be made consistent.","section":"References"}],"recommendation":"reject","confidential_remarks":"I concur with the reader's assessment. The discrepancy between the abstract's >3% claim and Table 1, combined with in-sample probe training and per-dataset hyperparameter fitting, leaves the central result unsupported. A revision would require a substantially new out-of-sample evaluation, which is beyond the scope of the current manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the method is a sensible incremental combination—semantic entropy probes feeding ReDeEP's E/P decoupling—but the paper's central claim is not supported by its own Table 1, and the evaluation setup can't rule out leakage. I wouldn't trust the numbers as written.\n\nThe actual new bit is replacing ReDeEP's lexical similarity scores (ECS, PKS) with probe-based semantic entropy scores (ECE, PKE). That's a sensible direction: semantic entropy should be more robust to paraphrase, and the probes are cheap at inference time. The paper does a reasonable job explaining the mechanics, and the intervention experiments are a decent sanity check that ECE and PKE track external context and parametric knowledge rather than just correlation.\n\nThe soft spots are serious, though. The abstract and Section 4.3 claim 'improvements exceeding 3% compared to ReDeEP.' Table 1 doesn't show that. For LLaMA3-7B on RAGTruth the gain is +2.78%; on HalluRAG it's -1.18%. Mistral is +2.00% and +2.40%; Qwen is +1.51% and +2.73%. Every delta is below 3%, and one is negative. So the headline claim fails on the paper's own numbers. Worse, the evaluation is in-sample: Section 4.1 says the probes were trained on the same RAGTruth and HalluRAG responses that are used for testing, and Section 4.2 shows per-dataset selection of regression coefficients and copy-head/FFN layers. That means the reported accuracies are fitted behavior, not independent prediction. Combined with no released code or data, there's no way to tell what holds up out-of-distribution.\n\nThere are also some minor rough edges—unusual unicode artifacts in Section 4.4, a missing reference in the PKE caption, and the ablation table is hard to parse. Those are fixable, but the evaluation problem is not.\n\nWho should read this? Someone working on hallucination detection in RAG might find the ECE/PKE construction worth trying, but they'd need to redo the experiments with a proper held-out split. As is, the paper doesn't deserve serious peer-review time. I'd desk-reject it with encouragement to redo the evaluation.","headline":"A plausible incremental idea whose central claim is contradicted by the paper's own Table 1, and whose evaluation cannot rule out train/test leakage.","tokens_in":19867,"tokens_out":3287,"would_cite":false,"duration_ms":27908,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Trading token-level scores for semantic-entropy probes makes RAG hallucination detection 3-10% more accurate than the ReDeEP and SEP baselines, on the paper's own evaluation.","keywords":["hallucination detection","retrieval-augmented generation","semantic entropy","semantic entropy probes","mechanistic interpretability","RAG hallucination","external context entropy","parametric knowledge entropy"],"falsifier":"Train the probes and fix $\\alpha$, $\\beta$, and the head/layer choices using only RAGTruth, then evaluate frozen on HalluRAG (and vice versa), or on a third RAG hallucination corpus never used for training; if the reported improvement of over 3% over ReDeEP shrinks or reverses, the gains are specific to the per-dataset tuning rather than to semantic-entropy measurement.","tokens_in":18691,"feed_emoji":"🧠","tokens_out":6865,"duration_ms":56947,"temperature":0.7,"pith_summary":"This paper claims that the two token-level metrics at the heart of the ReDeEP hallucination detector for retrieval-augmented generation can be replaced by semantic-entropy-probe outputs, and that this substitution improves detection accuracy by more than 3% over ReDeEP and more than 10% over the semantic-entropy-probe baseline SEP on the RAGTruth and HalluRAG corpora. The motivation is that RAG makes models paraphrase heavily while the underlying meaning stays the same, so token-level uncertainty and self-consistency scores diverge by more than 50% for semantically identical answers. If the claim holds, hallucination assessment in RAG becomes a single cheap forward pass that agrees with human annotation at near-annotator level, without the sampling cost of semantic entropy proper. The paper further argues the two probes isolate the two known hallucination mechanisms: insufficient use of retrieved context and excessive reliance on parametric knowledge stored in feed-forward layers.","feed_headline":"Entropy-probe swap lifts RAG hallucination detection by 3-10%","feed_subtitle":"Meaning-aware scores replace token-level uncertainty, so paraphrases no longer flip a response's hallucination verdict.","key_machinery":"The load-bearing object is the semantic entropy probe: a logistic-regression classifier, trained on hidden states from a single forward pass, that maps a representation to the probability that the eventual response has high semantic entropy, with the training threshold chosen by a one-dimensional clustering objective. Semantic entropy itself clusters sampled responses by bidirectional entailment, so the probe inherits invariance to paraphrase. The paper uses these probes in place of ReDeEP's cosine-similarity ECS and LogitLens-based PKS: ECE is the Z-Score of the current token's probe output against the probe outputs of highly attended context tokens, and PKE is the absolute difference of probe outputs across the FFN. The regression coefficients $\\alpha$ and $\\beta$, and the choice of copy heads and FFN layers, are selected per model and per dataset, and probes are only applied at layers after the 9th, where semantic representations are rich enough for probe accuracy.","core_discovery":"SEReDeEP keeps ReDeEP's regression-decoupling insight — that RAG hallucinations emerge from a disequilibrium between external context E and internal parametric knowledge P — but rebuilds how each is measured. External Context Entropy (ECE) runs a trained semantic-entropy probe over the hidden states of the current token and of the attention-attended context tokens, then z-scores the current token's entropy against the context distribution; Parametric Knowledge Entropy (PKE) takes the absolute difference of probe predictions on the residual stream just before and just after the feed-forward network. A linear regression $R(r) = \\sum_{l \\in F} \\alpha \\cdot P_n^l - \\sum_{l,h \\in A} \\beta \\cdot E_{n}^{l,h}$ then combines the two into the final hallucination score. The paper reports that on RAGTruth and HalluRAG across LLaMA3-7B, Mistral-7B, and Qwen2.5-7B, this version outperforms ReDeEP by over 3% and SEP by over 10% in prediction accuracy, and its correlation analysis shows ECE scores negatively correlated and PKE scores positively correlated with hallucination labels, as expected.","pith_inferences":["The gains are measured with per-dataset regression coefficients and per-dataset head/layer selections; a sterner test would fix these once on a training split and evaluate on unseen datasets, since the reported numbers leave open how much of the 3-10% reflects the semantic probes versus fitting to dataset idiosyncrasies.","Probe training and evaluation share the same two corpora (RAGTruth and HalluRAG); comparing against a corpus not used for probe training would separate genuine semantic-entropy transfer from corpus-specific calibration.","The Z-score ECE construction suggests a natural extension: replacing the scalar entropy of each context token with its full entropy distribution, or weighting context tokens by attention mass, might sharpen context-usage quantification further.","If the qualitative correlation results (ECE negative, PKE positive with hallucination) hold, the same probes could be inverted as an intervention signal — layer- and head-resolved entropy readouts could drive targeted suppression, not just a final score."],"forward_implications":["Hallucination scores in RAG become insensitive to paraphrase: answers that mean the same thing but differ in wording receive similar scores, closing the reported more-than-50% score divergence for semantically identical outputs.","Detection runs in a single forward pass: probes replace multi-sample generation, so hallucination monitoring costs are orders of magnitude below sampling-based methods like SelfCheckGPT or INSIDE.","The decoupled ECE/PKE view keeps the causal story of ReDeEP: one can still attribute a detected hallucination to under-used retrieved context or over-relied parametric knowledge, and the same intervention machinery (attention boosting, FFN suppression) applies.","The method transfers across decoder-only architectures of the 7B class, with the reported gains consistent across three different model families.","Because probe training needs only answer texts and entailment labels, the approach can be retargeted to new domains or languages without architectural changes."],"supporting_citations":[{"why":"Supplies the external-context/internal-knowledge decoupling framework, the regression scheme, and the baseline whose ECS and PKS metrics ECE and PKE replace.","marker":"[50]"},{"why":"Provides the semantic entropy probe architecture and training procedure embedded in the decoder layers.","marker":"[28]"},{"why":"Defines the semantic cluster entropy that the probes approximate, including the bidirectional entailment clustering of sampled responses.","marker":"[12]"},{"why":"Provides the RAGTruth evaluation corpus with word-level hallucination annotations used for training probes and measuring accuracy.","marker":"[42]"},{"why":"Provides the HalluRAG evaluation corpus with sentence-level hallucination annotations used for training probes and measuring accuracy.","marker":"[45]"},{"why":"Cited as the origin of the ECS/PKS metric lineage and as a sampling-based semantic baseline whose covariance-eigenvalue approach the method's single-pass design contrasts with.","marker":"[5]"}],"fun_headline_variants":["Semantic probe beats token-level uncertainty for RAG hallucination checks","RAG hallucination detector gains 3-10% with meaning-aware entropy","Entropy-based fusion sharpens RAG truthfulness scores","Context-parameter fusion reduces RAG hallucination mislabels","Semantic entropy decouples RAG knowledge clash for better detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that semantic entropy probes trained on the RAGTruth and HalluRAG response corpora, together with regression coefficients and head/layer selections tuned per dataset, still deliver an unbiased measure of hallucination-detection accuracy when evaluated on those same two datasets.","fun_headline_variants_meta":{"raw":{"variants":["Semantic probe beats token-level uncertainty for RAG hallucination checks","RAG hallucination detector gains 3-10% with meaning-aware entropy","Entropy-based fusion sharpens RAG truthfulness scores","Context-parameter fusion reduces RAG hallucination mislabels","Semantic entropy decouples RAG knowledge clash for better detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000437,"raw_usage":{"total_tokens":2261,"prompt_tokens":1027,"completion_tokens":1234,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":1145}},"tokens_in":643,"tokens_out":1234,"duration_ms":8222,"temperature":1.0,"reasoning_tokens":1145,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:15:47.564372+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the probes and fix $\\alpha$, $\\beta$, and the head/layer choices using only RAGTruth, then evaluate frozen on HalluRAG (and vice versa), or on a third RAG hallucination corpus never used for training; if the reported improvement of over 3% over ReDeEP shrinks or reverses, the gains are specific to the per-dataset tuning rather than to semantic-entropy measurement.","supporting_citations":[],"review_version":1}