{"id":"a3553bb8-1a33-478e-91c9-a2fd1cf8a047","arxiv_id":"2505.12265","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"RATE-FT, a fine-tuning method that adds rationale generation and an auxiliary question-answering task, improves reference-free hallucination detection in long-form generation by about 3 balanced-accuracy points over standard fine-tuning.","lead":"This paper finds that a model's own confidence scores cannot tell factual from hallucinated claims in open-ended long-form answers, and proposes a fine-tuning method with rationales and question answering that improves detection by about three balanced-accuracy points. The result suggests a practical path to reference-free hallucination detection, but the method depends on an automated search-based labeling process and no code or data were released.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Automated Google-Search labels are used for both training and evaluation; without a manual audit, RATE-FT's gain may reflect learning the annotator's bias rather than true hallucination detection.","rationale":"The reader's weakest assumption is essentially the same, and I agree: the automated annotation pipeline is the single most load-bearing premise. The concern is load-bearing because all headline numbers depend on that pipeline for both training and evaluation. The paper has good internal controls (RATE-FT_half, Fine-Tuning_para, ablations, multiple backbones), but those controls cannot correct a shared label bias. I am not claiming the annotations are wrong; I am claiming that the central empirical result is not yet independently verified. The proposed human audit is the minimal check that would settle whether the gain is genuine. Given the absence of code, data, and manual label verification, the appropriate verdict remains CONDITIONAL.","tokens_in":15283,"tokens_out":5192,"duration_ms":52225,"concrete_test":"Manually audit a stratified random sample of at least 200 claims from the LongFact test split (and 100 from Biography) with two independent human annotators using the original source-based verification protocol; compute BAcc for the retrained Fine-Tuning and RATE-FT detectors on the human-labeled subset. If RATE-FT's advantage over Fine-Tuning is not preserved on human labels, the central claim fails; if it is preserved, the annotation-bias concern is resolved for those test sets.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 3 constructs factual/hallucinated labels by having an LLM decompose responses, generate multi-step Google Search queries, and reason about whether search results support each claim; Section 5.1.6 confirms this pipeline is the sole source of supervision. The same pipeline labels the training, validation, and test sets for every model in Tables 3 and 7. The paper reports no manual audit or inter-annotator agreement for these labels, and it does not release the constructed data or code, so label fidelity cannot be checked externally. This matters especially for RATE-FT: the added rationales (Section 5) and the auxiliary QA examples for hallucinated claims are generated from the very same model-produced rationales that accompany the noisy labels. The detector is therefore trained to reproduce the annotator's judgment, including any systematic tendency to accept fluent but wrong claims or to reject correct but awkwardly phrased ones. The +3% BAcc gain over Fine-Tuning, and the consistent gains across model families, are computed against this same noisy test set. A significant p-value (Table 3) shows only that the gap is unlikely under sampling noise conditional on those labels; it does not address label bias. Thus the central claim that auxiliary-task training improves real hallucination detection is not yet fully established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies reference-free hallucination detection for open-domain long-form generation. It first reports that internal states (token probabilities and entropies) are not better than random at separating factual from hallucinated atomized claims, then compares prompting, probing, and fine-tuning, finding fine-tuning most effective. The main contribution is RATE-FT, which augments fine-tuning with (a) rationales in a 'label-rationale' format and (b) an auxiliary question-answering task built from the same claims. Experiments on LongFact and biography generation across several LLMs show consistent BAcc gains, e.g., +3 points over fine-tuning on LongFact, with ablations suggesting the gain comes from the auxiliary QA task rather than extra data.","tokens_in":15490,"tokens_out":5257,"duration_ms":52264,"significance":"If the results hold, RATE-FT is a practical, training-only enhancement that improves reference-free hallucination detectors without raising inference cost, because the model emits the label token first and stops after the first token. The paper also offers a systematic comparison of internal-state, prompting, probing, and fine-tuning baselines, which is useful for the community. The ablation design is a strength: Fine-Tuning_para and RATE-FT_half (Tables 4 and 5) directly address the 'more data vs. auxiliary task' confound, and the gains replicate across four model families and two datasets. The main weaknesses are the lack of any manual audit of the automatically constructed labels and the lack of clarity about the train/test split granularity, both of which are load-bearing for the central claim. Code and data are not released, limiting reproducibility.","major_comments":[{"comment":"The same automated pipeline (LLM-based claim decomposition, multi-step Google Search queries, and LLM reasoning over search results) is used to create the training, validation, and test labels, and Section 5.1.6 confirms that external search is used for all supervision. No manual audit, inter-annotator agreement, or human-verified subset is reported. Because RATE-FT additionally trains on rationales and QA examples generated from these same labels, the reported improvement over Fine-Tuning may reflect learning the annotator's systematic biases (e.g., accepting fluent but wrong claims) rather than improved true hallucination detection. The paper should provide a manual quality audit of a random sample of labels, or evaluate RATE-FT on a human-verified benchmark, to establish that the gains are not an artifact of the shared annotation source. This is essential for the central claim that RATE-FT improves hallucination detection.","section":"Section 3 and Section 5.1.6"},{"comment":"The 2,711 claims are split into training (70%), validation (20%), and test (10%) sets, but the paper does not state whether the split is at the claim level or grouped by the original response/prompt. If claims from the same long-form response appear in both training and test sets, the reported BAcc numbers, including the +3% gain of RATE-FT, may be inflated by response-level leakage, because the model could memorize response-specific patterns. The authors should clarify the split granularity; if it is claim-level, they should re-run the main comparisons with a response-level split and report whether the gains persist.","section":"Section 4"},{"comment":"The statement 'RATE-FT is significantly better than Fine-Tuning with p-value<0.01' is not supported by any described statistical test. A single BAcc number per method on one test set cannot by itself yield a p-value; the authors must specify the test (e.g., paired bootstrap over claims, multiple random seeds, or a significance test across models/datasets) and report the corresponding test statistic and confidence interval. This is directly relevant to the central claim of a significant improvement.","section":"Table 3"}],"minor_comments":[{"comment":"The conclusion that internal states are 'not better than random guessing' is based only on visual inspection of overlapping histograms in Figures 2--5. Please also report quantitative metrics such as AUC or BAcc with confidence intervals to support this claim.","section":"Section 3"},{"comment":"The BAcc-unknown metric counts 'unknown' predictions as correct, assuming the external tool is perfect. This makes the metric sensitive to the hyperparameters α_low and α_high and can favor methods that output more unknowns. The paper should report the fraction of 'unknown' predictions and consider a metric that accounts for the cost of tool invocation.","section":"Section 5.1.4"},{"comment":"The comparison with F2 (Hu et al., 2024) is only one sentence and does not include any experimental comparison. Please either add a quantitative comparison or clarify in what sense the methods differ beyond the stated goal (detection vs. mitigation).","section":"Section 5.1.3"},{"comment":"The paper does not release data or code, which limits reproducibility and makes external verification of label quality impossible. Releasing the constructed dataset and training scripts would strengthen the paper.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The shared-source issue between training and evaluation labels is the most serious concern. Even if a manual audit is added, the paper should clarify whether the reported gains are robust to response-level splitting. I do not see grounds for rejection because the ablations and cross-model consistency suggest a real effect, but the current evidence is not sufficient for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is RATE-FT: take the existing fine-tuning baseline for claim-level hallucination detection and add two cheap ingredients - rationale-labeled examples in 'label-rationale' format and an auxiliary QA task on the same claims. The paper does a good job isolating that the QA task, not extra data, drives the gain. The ablations with Fine-Tuningpara and RATE-FThalf are exactly the right controls, and the gain replicates across four model families and on both LongFact and Biography, including an OOD transfer from LongFact to Biography. That is a solid empirical result, even if the absolute numbers are modest.\n\nThe internal-states finding - that token probabilities and entropies essentially don't separate factual from hallucinated claims in long-form generation - matches Kapoor et al. and is a useful sanity check, though not a surprise at this point.\n\nThe soft spot is the annotation pipeline. The same model-generated, Google-Search-verified labels are used for training, validation, and test. There is no manual audit, no inter-annotator agreement, and no release of code or constructed data. That means the test set is the annotator's judgment, and a detector that learns that annotator's quirks will look better on it. The relative comparison between fine-tuning and RATE-FT is still meaningful in the sense that both are measured against the same labels, but the claim that RATE-FT improves 'real' hallucination detection in the wild is not fully established by the paper as written. This is a real limitation, but not a fatal one: it is the standard limitation of benchmark construction in this area, and the paper honestly states its reliance on external search for data construction.\n\nMinor issues: no error bars on most tables (only the p-value for the main comparison), and the abstract's 'new paradigm' framing oversells what is an incremental but useful training recipe. The paper would be stronger with a small manual audit of a few hundred labels, released data, and error bars from multiple seeds.\n\nOverall, this is a competent, honest empirical study. It deserves serious peer review. I would send it out, and I would expect the reviewers to push on the label-fidelity point and on reproducibility, but the central comparison is well enough controlled to support the paper's main claim as stated for the constructed benchmark.","headline":"A well-controlled empirical recipe that buys a consistent ~3-point BAcc gain over fine-tuning for reference-free hallucination detection, though the shared search-based annotation pipeline for training and test labels means part of the gain may reflect learning the annotator's biases.","tokens_in":727,"tokens_out":800,"would_cite":true,"duration_ms":32820,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A reference-free hallucination detector for long-form answers improves by jointly training on an auxiliary question-answering task, gaining about 3 balanced-accuracy points over fine-tuning on LongFact.","keywords":["hallucination detection","long-form generation","reference-free detection","fine-tuning","auxiliary task learning","rationale augmentation","internal states","balanced accuracy"],"falsifier":"Take a random sample of claims labeled by the paper's search-and-reason pipeline and have independent human fact-checkers label them, then retrain plain fine-tuning and RATE-FT on human-verified labels and compare on a human-verified test set; if RATE-FT's advantage over fine-tuning shrinks or disappears, the central claim is not established.","tokens_in":15074,"feed_emoji":"🎯","tokens_out":6966,"duration_ms":64948,"temperature":0.7,"pith_summary":"This paper asks whether a language model can detect its own factual errors in open-domain long-form answers without relying on external fact-checking tools. The authors first show that internal confidence signals—output probabilities and entropies, including entity-focused variants—cannot separate factual from hallucinated claims better than random chance. After comparing prompting, probing, and fine-tuning, they find fine-tuning to be the strongest existing approach. Their central claim is that fine-tuning is not the ceiling: augmenting it with rationale-augmented labels and an auxiliary question-answering task (a method they call RATE-FT) raises balanced accuracy from 76.1 to 79.6 on LongFact and from 78.2 to 80.9 on biography generation, with consistent gains across four model families. If this is right, reference-free hallucination detection in long-form generation can be improved substantially by multi-task supervision alone.","feed_headline":"Auxiliary QA training lifts hallucination detection 3 points","feed_subtitle":"Fine-tuning a detector to also answer questions about claims beats plain fine-tuning across models and datasets.","key_machinery":"The load-bearing mechanism is RATE-FT (Rationale and Auxiliary Task Enhanced Fine-Tuning), a training protocol that pairs each (claim, label) example with a 'label-rationale' format—the model learns to emit True or False first and then explain—and with a generated question-answering example in which the model must answer a question about the key information in the claim. For hallucinated claims, the rationale is used to construct the correct answer, so the QA task always teaches a factual response. This second task is what carries the gain: ablations show that dropping either the rationales or the QA task noticeably lowers balanced accuracy, and controls with paraphrased claims or halved training data show the gain is not just a data-volume effect. The label-first ordering keeps inference cost identical to plain fine-tuning, since the label is read from the first output token.","core_discovery":"The paper's discovery is that a detector's judgment about a claim gets more reliable when, during fine-tuning, it is asked not only to label the claim True or False but also to reconstruct the factual core of the claim through a derived question-and-answer task. The authors argue that internal states such as token probability and entropy are not informative for long claims because they reflect confidence in surface wording rather than correctness of content. They then show that fine-tuning outperforms prompting and probing, and that their RATE-FT protocol—training jointly on claim-label pairs, label-first rationales, and QA examples built from each claim—outperforms ordinary fine-tuning by about three balanced-accuracy points on LongFact. The improvement is attributed to the auxiliary QA task rather than to having more training data, because a paraphrase-augmented control with the same data volume does not match it, and halving RATE-FT's data still exceeds fine-tuning. The same pattern holds on a biography benchmark and across models of different families and sizes.","pith_inferences":["Beyond the paper's experiments, the QA-auxiliary training can be read as teaching the model to 'repair' hallucinated claims into correct answers; a testable extension is whether RATE-FT also improves the factuality of the model's own generation, not just its detection accuracy.","If the search-and-reason labeling pipeline carries a systematic bias—for instance, favoring fluent claims—then part of the measured gain might be adaptation to that annotation style; rerunning on human-verified labels would determine how much of the +3 points is genuine detection skill.","The same multi-task recipe could apply to other fine-grained classification problems where labels are expensive but a correct-answer reconstruction is cheap, such as faithfulness hallucination detection or factual consistency scoring.","The uncertainty-routing result suggests a natural deployment pattern: run RATE-FT as a cheap pre-filter and reserve external verification for the uncertain minority of claims, which the paper did not explicitly propose as a system."],"forward_implications":["Reference-free hallucination detectors for long-form answers can be trained without any external fact-checking tool at inference time, with accuracy near that of search-based verification.","The RATE-FT recipe transfers across model families and scales, so the gain is not tied to one backbone.","The auxiliary QA task, not extra training examples, drives the improvement; future detector training can focus on task design rather than data volume.","Adding an 'unknown' output and routing uncertain claims to external tools yields further gains, enabling hybrid pipelines that spend search budget only where the model is unsure.","Because RATE-FT performs better than fine-tuning across short, medium, and very long responses, the benefit persists even as response length grows."],"supporting_citations":[{"why":"Supplies the LongFact benchmark and the search-based claim verification procedure used to build factual and hallucinated labels.","marker":"Wei et al. (2024)"},{"why":"Provides SelfCheckGPT and the internal-state confidence measures (probability and entropy) that the paper tests and extends.","marker":"Manakul et al. (2023)"},{"why":"Gives the fine-tuning approach the paper treats as the strongest baseline and supports the view that models must be taught what they do not know.","marker":"Kapoor et al. (2024)"},{"why":"Provides the biography-generation benchmark and the fine-grained atomic evaluation style used for the second dataset.","marker":"Min et al. (2023)"},{"why":"Supplies the probing method based on a multilayer perceptron over contextualized embeddings, used as a baseline.","marker":"Su et al. (2024)"},{"why":"Gives F2, the closest prior method that integrates rationales and auxiliary tasks, serving as a comparison target.","marker":"Hu et al. (2024)"},{"why":"Supplies the fine-tuning implementation used to run LoRA experiments in the paper.","marker":"Zheng et al. (2024)"},{"why":"Provides the balanced-accuracy evaluation metric used for all detection results.","marker":"Tang et al. (2024)"}],"fun_headline_variants":["QA sidetask lifts hallucination detection 3% over fine-tuning","Training detector to answer claims beats plain fine-tuning","RATE-FT: auxiliary QA task outperforms fine-tuning for fact-check","Internal states fail; QA-augmented fine-tuning detects lies better"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the automated pipeline which decomposes long responses into claims, searches the web, and reasons about whether search results support each claim produces correct factual-versus-hallucinated labels for both training and evaluation; if that labeling is biased, the reported gains may reflect fitting the annotation style rather than genuine hallucination detection ability.","fun_headline_variants_meta":{"raw":{"variants":["QA sidetask lifts hallucination detection 3% over fine-tuning","Training detector to answer claims beats plain fine-tuning","RATE-FT: auxiliary QA task outperforms fine-tuning for fact-check","Internal states fail; QA-augmented fine-tuning detects lies better"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000661,"raw_usage":{"total_tokens":3026,"prompt_tokens":956,"completion_tokens":2070,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":1996}},"tokens_in":572,"tokens_out":2070,"duration_ms":14309,"temperature":1.0,"reasoning_tokens":1996,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:37:26.220303+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of claims labeled by the paper's search-and-reason pipeline and have independent human fact-checkers label them, then retrain plain fine-tuning and RATE-FT on human-verified labels and compare on a human-verified test set; if RATE-FT's advantage over fine-tuning shrinks or disappears, the central claim is not established.","supporting_citations":[],"review_version":1}