{"id":"069040d8-9699-4556-95b9-d2518fadedc1","arxiv_id":"2505.20487","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"InFACT trains LLMs with hierarchical informativeness rewards plus abstention, improving factual precision on QA benchmarks while largely preserving recall.","lead":"This paper introduces InFACT, a training method that rewards language models for giving more detailed correct answers and for saying \"I don't know\" instead of guessing. The authors report that models trained this way become both more informative and more factually precise on several question-answering benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central factuality claim rests on an unvalidated GPT-4 abstention detector used both to set training rewards and to compute precision; if that detector treats wrong answers as abstentions, the reported F1 gains become artifacts of the detector rather than genuine factuality improvements.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing point: the GPT-4 abstention detector is unvalidated and is used both for training rewards and for the precision/F1 evaluation. My read of the paper confirms this is the single most fragile link in the central claim. The paper's only error analysis covers missed abstentions (Section 5.5, 3%), not false abstention classifications, so the direction of bias that would inflate the headline is entirely unexamined. The training objective in Eq. (2) gives 0 for abstain versus -1 for wrong, and the precision metric excludes abstained outputs from the denominator, so the detector's classification directly determines both the learned behavior and the reported improvement. This is a concrete, testable measurement threat, not a disagreement with consensus. The rest of the paper—consistent gains in informativeness, ablations, and the fact that precision gains are accompanied by only minor recall drops—is plausible and reproducible in principle, and the authors release code and data. Conditional acceptance is the right stance because the method may be sound, but the headline factuality result should not be fully accepted until the detector is validated against human labels and the tables are recomputed. I therefore keep the reader's CONDITIONAL verdict unchanged and agree that the detector is the weakest assumption.","tokens_in":16343,"tokens_out":2941,"duration_ms":29142,"concrete_test":"Construct a gold abstention set from the released code/data: sample about 200 outputs per model and benchmark (or at least the Llama-3.1-8B rows in Tables 3–5), have two independent annotators label each output as abstain versus answer, and also judge the correctness of each answer. Measure the GPT-4 detector's precision, recall, and F1 against these labels, stratifying by whether the model output is correct, wrong, or abstaining. Then recompute the precision, recall, and F1 values in Tables 3–7 using human abstention labels instead of GPT-4 labels. If the recomputed values no longer place InFACT first in F1 on most benchmarks, or if the precision gains shrink to below roughly 5 points, the central factuality claim is an artifact of the detector rather than genuine model behavior.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim—highest F1 across all models and factuality benchmarks, with gains driven by precision—depends on the abstention detector of Section 3.2. That detector is used twice: as the training reward in Eq. (2) (0 for abstain, -1 for wrong) and as the filter that defines the precision denominator in the Section 4 precision metric (correct answers among non-abstaining outputs). If GPT-4 systematically labels wrong generations as abstentions, the model is rewarded for saying \"I don't know\" instead of a wrong answer, and the same bias then removes those outputs from the precision denominator, inflating F1. The paper reports no accuracy, precision, recall, or human validation for this detector. Its own error analysis in Section 5.5 checks only undetected abstentions (3%), not the opposite error—answers incorrectly classified as abstain—which is the direction that would inflate the headline result. Because the training objective explicitly assigns higher reward to abstaining than to being wrong, the model has an incentive to produce outputs that the detector will label as abstain, and the evaluation then reuses that same label as ground truth. This creates a reward-hacking loop: improved precision may reflect alignment with the detector's idiosyncrasies rather than improved factuality. The concern is therefore not that abstention is an invalid goal, but that the claimed factuality improvement is measured with the same unvalidated instrument used to train it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces InFact, a two-stage alignment method aimed at improving the factual informativeness of LLM answers. A hierarchy of increasingly informative correct answers is constructed for questions from GRANOLA QA, QAMPARI, and RoMEQA; the model is first tuned to move from less informative to more informative answers and to abstain when its answer is wrong, then trained with a reward or preference objective that assigns larger rewards to more informative correct answers, zero reward to abstentions, and negative reward to wrong answers. Factuality is measured on TriviaQA, PopQA, TruthfulQA, Natural Questions, and PIQA by computing precision, recall, and F1 over outputs that a GPT-4-based abstention detector classifies as non-abstaining. The paper reports substantial gains in informativeness on the training-source benchmarks and, in most cases, higher F1 on the factuality benchmarks driven by precision, alongside ablations and an error analysis.","tokens_in":16660,"tokens_out":4129,"duration_ms":44587,"significance":"If the reported factuality gains are genuine, the paper would make a useful contribution: a lightweight alignment objective that jointly improves answer informativeness and reduces wrong answers would be valuable for LLM deployment. The breadth of the experiments—five base models, multiple training variants, and five factuality benchmarks—is a strength, as is the public release of code and data. However, two load-bearing issues currently limit the significance of the results. First, the factuality evaluation depends on an unvalidated GPT-4 abstention detector that is also used to assign training rewards, creating a potential reward-hacking loop. Second, the informativeness evaluation largely measures performance on the same hierarchy construction used to define the training reward, so the informativeness improvements are partly by construction. The paper should be credited for making the method and data available and for including ablations, but the central empirical claims need additional validation before the contribution can be accepted as stated.","major_comments":[{"comment":"The GPT-4 abstention detector is load-bearing but never validated. It is used in two places: in Eq. (2) it assigns reward 0 to abstentions versus -1 to wrong answers, and in the Section 4 precision metric it removes detected abstentions from the denominator. If the detector misclassifies wrong answers as abstentions, the same bias both rewards the model for producing detector-idiosyncratic abstentions and inflates the reported precision. The error analysis in Section 5.5 checks only undetected abstentions (the 3% case) and never reports false abstention errors, which are exactly the errors that would inflate the headline result. Please provide human-validated accuracy/precision/recall for the detector on held-out outputs from each model, and re-report factuality with abstentions treated as incorrect rather than excluded.","section":"§3.2, §4 (Evaluation), Eq. (2)"},{"comment":"The data split used for informativeness evaluation is not precisely specified. The text says the model is evaluated on the \"test split\" of GRANOLA QA, QAMPARI, and RoMEQA, but it does not state whether questions in the training subset were excluded from the test set, how the hierarchy for each test question was constructed, or whether any of the test examples were seen during in-context-learning baselines. Ambiguity here matters because the same hierarchy construction is used in training and evaluation. Please specify the exact split procedure and ensure no overlap between training and evaluation examples.","section":"§4 (Evaluation Data)"},{"comment":"The informativeness evaluation is partly circular. For QAMPARI and RoMEQA, the hierarchy of all pairs, triplets, and larger subsets is used both to construct the training reward in Eq. (2) and to score the test predictions through the original benchmark metrics. For GRANOLA QA, the same granularity hierarchy is used as the training label and as the evaluation metric. Improvements on these benchmarks can therefore reflect the model learning the specific hierarchy format rather than a general notion of informativeness. The paper should add an external or human evaluation of informativeness, or evaluate on hierarchies that were not used to define the reward.","section":"§4, §5.1, Table 2"},{"comment":"The claim that \"across all model sizes and all benchmarks, the overall F1 performance of our models is the highest\" is not supported by the tables. In Table 7, Qwen2.5-7B on PIQA has base model F1 78.6 and FT baseline F1 77.8, both higher than the informativeness-alignment F1 of 76.4. The claim should be qualified to the benchmarks where it is actually observed, or the contradicted cases should be discussed.","section":"§5.2, Tables 3–7"},{"comment":"The paper states that the PPO/DPO gap is \"not compelling enough to be statistically significant,\" but no statistical significance tests or error bars are reported anywhere in the empirical sections. Given that Section 1 also claims \"significant improvement in factual precision,\" the absence of variance estimates or significance testing makes the strength of the claims difficult to assess. Please report results across multiple seeds or provide confidence intervals and appropriate significance tests for the main comparisons.","section":"§5.3, Tables 2–3"}],"minor_comments":[{"comment":"The prompting baseline contains a typo: \"refrain form answering\" should be \"refrain from answering.\"","section":"§4 (Baselines)"},{"comment":"The table formatting is inconsistent: the Mistral row in Table 6 is typeset as one continuous line with missing separators, and several rows in Table 7 have uneven spacing. The tables should be reformatted for readability.","section":"§5.2, Tables 6 and 7"},{"comment":"For QAMPARI and RoMEQA, the phrase \"all possible pairs, triplets, and so on\" is potentially combinatorial; please clarify how the hierarchies were generated in practice, including any truncation or filtering, since the average of 8.9 levels per example suggests some selection rule.","section":"§2.2, §4 (Training Data)"},{"comment":"The ablation description says the results appear in Tables 2, 3, 4, and 5, but comparable ablation rows are not shown for Mistral and Qwen in Tables 6 and 7; please add those rows or adjust the cross-reference.","section":"§5.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal, but the unvalidated abstention detector is a serious issue that affects the central claim. I would not recommend rejecting the paper outright, because the concern is addressable with additional validation and a re-analysis that treats abstentions conservatively. The overclaim about 'highest F1 across all benchmarks' should also be corrected. Given the currently missing validation and the partly circular informativeness evaluation, major revision is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a solid, practical paper. The idea is simple and useful: train LLMs to prefer the most informative correct answer in a hierarchical QA setup, and to abstain rather than hallucinate. The two-stage recipe — structure tuning followed by DPO/PPO on a level-based reward — is a new combination, and the results are consistent across five model families.\n\nWhat it does well: the hierarchy-based informativeness objective is clean and the authors release code and data. They show large gains in informativeness on GRANOLA and better recall on QAMPARI and RoMEQA. The factuality results on six unrelated QA benchmarks are promising: precision goes up substantially while recall stays roughly flat. The ablations show both stages matter, which strengthens the internal story.\n\nNow the soft spots, in proportion. The main one is the GPT-4 abstention detector (Section 3.2). It is used twice: as a training reward (abstain = 0, wrong = -1) and to define the precision denominator in evaluation. If GPT-4 tends to label wrong answers as abstentions, the apparent precision gains are partly an artifact. The error analysis only checks missed abstentions, not the opposite error — answers wrongly classified as abstain — which is exactly the direction that would inflate the headline result. The paper reports no validation of this detector, and that is a real gap.\n\nSecond, the informativeness evaluation uses the same hierarchy structure used in training, so improvement there is partly by construction. The factuality benchmarks are genuinely unrelated, so the core claim is not circular overall. Third, there are no error bars or significance tests across the five models, which limits confidence in the consistency claim. Fourth, the training-data sampling is ambiguous: they say they randomly sample 1k examples from each dataset but do not state whether that is from the train split, even though evaluation is on the test splits of the same benchmarks.\n\nThe abstention-detector issue is fixable: a human-validated confusion matrix would settle whether the detector is reliable. The split ambiguity also needs a one-sentence clarification. The method itself is plausible and the experimental breadth is real.\n\nWho this is for: anyone working on alignment for factuality, selective prediction, or QA. It deserves a serious referee. Send it to review, but ask the authors to validate the abstention detector and clarify the splits.","headline":"A solid, practical two-stage alignment method for more informative and more precise LLM answers, but the headline factuality gains rest on an unvalidated GPT-4 abstention detector used both to train and to evaluate.","tokens_in":17143,"tokens_out":2824,"would_cite":true,"duration_ms":27917,"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":"LLMs trained to prefer the most informative correct answer they know become more factual as a side effect, abstaining rather than guessing when unsure.","keywords":["factual completeness","informativeness","factuality","hallucination","abstention","preference optimization","reinforcement learning","question answering"],"falsifier":"Sample model outputs from the informativeness-aligned models, have human annotators label each as correct, wrong, or abstaining, and recompute precision and F1 using the human labels instead of the GPT-4 detector. If the precision advantage over the base model mostly disappears once mislabeled abstentions are counted as errors, the reported factuality improvement is an artifact of the detector rather than a genuine behavioral change.","tokens_in":16146,"feed_emoji":"🎯","tokens_out":12391,"duration_ms":94699,"temperature":0.7,"pith_summary":"Large language models often answer with facts that are true but vague: asked where Barack Obama was born, they may say 'the United States' when they also know 'Honolulu, Hawaii.' The paper argues this is a training-signal problem, not just a knowledge problem, and proposes an alignment objective that rewards answers that are both correct and as informative as possible, with a specific reward for each level of answer granularity and a zero reward for abstaining instead of guessing. Trained this way with PPO or DPO, models of several sizes and architectures produce more informative answers on granularity and many-answer QA benchmarks and reach the highest F1 on five factuality benchmarks, with gains coming mostly from precision. The central finding is that optimizing for informativeness also improves factuality: the model learns to give the better answer it already knows, and to abstain when it does not know.","feed_headline":"Rewarding detailed correct answers lifts LLM factuality","feed_subtitle":"Training LLMs to prefer specific answers and to abstain when unsure raises precision while keeping recall.","key_machinery":"The load-bearing object is the answer hierarchy $H_i = (A_1, \\ldots, A_{L_i})$ attached to each question, where $A_1$ is the most informative set of correct answers and lower levels are progressively less specific. Training runs on the reward function $R(M, \\hat{y}) = 1/\\sqrt{j}$ when the model's answer falls in level $A_j$, $0$ when it abstains, and $-1$ when it is wrong; the square-root scaling gives diminishing returns for coarser levels while still preferring any correct answer over abstention. Structure tuning supplies the initial behavior the reward assumes, by teaching the model to climb the hierarchy and to abstain instead of hallucinating. The GPT-4-based abstention detector is the mechanism that turns 'I don't know' into a distinguishable outcome for both reward assignment and precision scoring.","core_discovery":"The paper's central claim is that factual correctness and informativeness are separable axes of answer quality, and that aligning a model to prefer the most informative correct answer also makes it more factual. The authors define an informativeness-evaluation dataset in which each question is paired with a hierarchy of correct answers, from most specific (Level 1) to most general, and use that hierarchy in a two-stage training framework. Structure tuning first teaches the model to move up one level when its answer is correct but coarse, and to say 'I don't know' when its answer is wrong. Informativeness alignment then rewards the model with $1/\\sqrt{j}$ for an answer in level $j$, $0$ for abstention, and $-1$ for a wrong answer, optimized with PPO or DPO. Across Llama, Mistral, and Qwen models, the trained models show the highest overall F1 on TriviaQA, PopQA, TruthfulQA, Natural Questions, and PIQA, with precision gains from more effective abstention and no major recall drop; the same models also score higher on informativeness metrics for GRANOLA QA, QAMPARI, and RoMEQA.","pith_inferences":["If the effect is causal, informativeness-aligned training could serve as a behavioral alternative to confidence thresholds and semantic-entropy methods, teaching abstention rather than estimating it at inference time, which may be cheaper to deploy.","The error analysis notes that 13% of the model's mistakes are conspicuously long answers, suggesting the model latches onto length as a proxy for informativeness; a controlled experiment that decouples answer length from granularity would test whether the reward is being gamed.","Because the reward places abstention (0) above a wrong answer (-1) but below any correct answer, the method implicitly assumes users prefer a refusal over a guess; in settings where partial answers are useful, the hierarchy and reward would need reweighting.","The approach depends on answer hierarchies, so extending it to domains without such structure would require constructing hierarchies from knowledge graphs or annotations; the paper's transformation of flat multi-answer datasets into pair and triplet levels is one such recipe."],"forward_implications":["On all five factuality benchmarks tested, the informativeness-aligned models achieve the highest F1 of any method, including confidence-threshold, prompting, in-context, P(True), and semantic-entropy baselines.","Factual precision rises because the model abstains on questions it would previously have answered wrongly, while recall stays close to the base model's level, indicating that the training does not erase stored knowledge.","Improvements hold across model families and sizes (Llama-3.2-1B/3B, Llama-3.1-8B, Mistral-7B, Qwen2.5-7B), so the effect is not tied to one architecture.","Both training stages matter: removing structure tuning degrades precision because the base models do not abstain effectively without it.","The reward design makes PPO and DPO roughly interchangeable, with no statistically significant gap between them on the 1B model."],"supporting_citations":[{"why":"Supplies GRANOLA QA, the multi-granularity benchmark whose answer hierarchies define the informativeness task and metric.","marker":"(Yona et al., 2024)"},{"why":"Supplies QAMPARI, the many-answer QA dataset converted into hierarchical training and evaluation examples.","marker":"(Amouyal et al., 2023)"},{"why":"Supplies RoMQA, the multi-answer benchmark used for training and informativeness evaluation.","marker":"(Zhong et al., 2022)"},{"why":"Provides TriviaQA, a general-knowledge factuality benchmark used to measure F1 gains.","marker":"(Joshi et al., 2017)"},{"why":"Provides PopQA, a subject-relation-object QA benchmark used in factuality evaluation.","marker":"(Mallen et al., 2022)"},{"why":"Provides TruthfulQA, the falsehood-mimicry benchmark used to measure factuality.","marker":"(Lin et al., 2021)"},{"why":"Provides Natural Questions, a real-user query benchmark used in factuality evaluation.","marker":"(Kwiatkowski et al., 2019)"},{"why":"Provides PIQA, the physical commonsense benchmark used in factuality evaluation.","marker":"(Bisk et al., 2019)"},{"why":"Defines the P(True) calibration baseline, a representative uncertainty-based abstention competitor.","marker":"(Kadavath et al., 2022)"},{"why":"Defines the semantic-entropy uncertainty baseline that the method compares against.","marker":"(Kuhn et al., 2023; Aichberger et al., 2024)"}],"fun_headline_variants":["Informativeness alignment lifts LLM factuality","Rewarding specificity improves LLM truthfulness","Detail-focused training makes LLMs more factual","Teaching LLMs to be specific also boosts accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes the GPT-4-based abstention detector correctly recognizes when a model has refused to answer; it sets training rewards and also defines the precision metric, and the paper reports no validation of the detector, so any systematic confusion between wrong answers and abstentions would inflate the factuality gains.","fun_headline_variants_meta":{"raw":{"variants":["Informativeness alignment lifts LLM factuality","Rewarding specificity improves LLM truthfulness","Detail-focused training makes LLMs more factual","Teaching LLMs to be specific also boosts accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000186,"raw_usage":{"total_tokens":1320,"prompt_tokens":932,"completion_tokens":388,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":331}},"tokens_in":548,"tokens_out":388,"duration_ms":3925,"temperature":1.0,"reasoning_tokens":331,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:53:27.846620+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sample model outputs from the informativeness-aligned models, have human annotators label each as correct, wrong, or abstaining, and recompute precision and F1 using the human labels instead of the GPT-4 detector. If the precision advantage over the base model mostly disappears once mislabeled abstentions are counted as errors, the reported factuality improvement is an artifact of the detector rather than a genuine behavioral change.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies QAMPARI, the many-answer QA dataset converted into hierarchical training and evaluation examples."}],"review_version":1}