{"id":"2fa9e427-9679-4a90-8c31-19965faee98e","arxiv_id":"2502.07184","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An adaptive contrastive learning strategy that uses a model's own sampled response accuracy to create per-region positive and negative training pairs improves LLM truthful rate by up to 6.9% over IDK-SFT.","lead":"This paper proposes an adaptive contrastive learning method that fine-tunes large language models to answer known questions correctly and say \"I don't know\" to unknown ones. The method splits questions into knowledge regions based on the model's own repeated-response accuracy and trains with different positive and negative examples for each region.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Truthful Rate gains may be an artifact of self-referential labels: test known/unknown labels are generated by the same base-model repeated-sampling proxy the method is trained to match, so confirmation on externally labeled data is needed.","rationale":"Reading in good faith: the method is coherent and has a plausible mechanism; the contrastive losses are defined for each quadrant; the experiments are consistent; ALCUNA and RAG integration add some robustness. The load-bearing weakness is identification. The strongest claim is causal: adaptive contrastive learning improves Truthful Rate over SFT by 5–7 points. For that claim to hold, Truthful Rate must measure truthful behavior. It is measured against known/unknown labels that, as far as the paper describes, are generated from the same base-model repeated-sampling accuracy and thresholds used to construct the training data. Training on those labels and then evaluating on labels generated the same way rewards a model that simply matches the base model's confidence pattern; it does not establish that answers are correct or refusals appropriate with respect to actual knowledge. This is not a circularity in the loss but an evaluation-circularity in the headline metric. The concern would be fully resolved by an external label source; ALCUNA is a partial step but only tests artificial unknowns and shows a small gain. Because the central claim's quantitative support depends on this evaluation choice, the appropriate disposition is conditional acceptance pending an externally labeled evaluation. No ad hominem; the authors are transparent that thresholds are hyperparameters and that correctness is lexical matching.","tokens_in":19760,"tokens_out":6434,"duration_ms":60765,"concrete_test":"Take a random subsample (e.g., 500 questions per dataset) from the TriviaQA and Natural Questions test sets. Have independent annotators label each question as 'model should know' or 'model should not know' using gold Wikipedia passages (or human knowledge of the base model's training data). Recompute IK-IK, IK-IDK, and Truthful Rate for IDK-Prompting, IDK-SFT, and IDK-SFT-Adpt-Ctr against these external labels. If the adaptive method's advantage over IDK-SFT drops below the reported 5–7 pts or is confined to questions the base model already labels unknown, the central claim is not supported. Run the same check with IDK chosen on a validation split to remove threshold fitting.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central causal claim is that the adaptive contrastive loss, not the SFT signal, drives the 5.0–6.9 pt Truthful Rate gains (Section 4.4). The Truthful Rate is computed against per-question known/unknown labels (Appendix B.1), but the paper never states how the test-set labels are obtained; the only label-generation procedure described is repeated sampling of the base LLM and thresholding accuracy by IK/IDK (Section 3.1). Those same labels are the training targets for IDK-SFT and for the adaptive method, so the evaluation measures how well a trained model reproduces the base model's own accuracy pattern. A model that learns to answer exactly where the base model sampled correctly and to emit 'I don't know' exactly where it sampled incorrectly will raise Truthful Rate even if every label is wrong about the model's true knowledge. The ALCUNA experiment (Appendix C.1) is the right independent direction, but it shows only a 0.5 pt gain over IDK-SFT on artificially unknown questions, and the main OOD claim is still judged on self-generated labels. The IDK threshold that maximizes Truthful Rate is selected from the test set (Table 2), and no error bars are reported, so the headline improvement is also optimistic.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Adaptive Contrastive Learning (IDK-SFT-Adpt-Ctr) to improve the honesty of LLMs by refining their knowledge boundaries. The method defines upper and lower accuracy thresholds (IK and IDK) from repeated sampling of the base model, classifies questions into known, uncertain, and unknown regions, and constructs quadrant-specific positive and negative instances for contrastive instruction tuning. An adaptive loss, Eq. (3), combines generation loss with a contrastive loss whose weight is max(lambda, detach(L_gen/L_ctr)). Experiments on TriviaQA and Natural Questions with LLaMA-2-7B-chat and Mistral-7B-Instruct-v0.1 report Truthful Rate improvements over IDK-Prompting and IDK-SFT, with additional ablations on loss components, IDK threshold values, model size, ALCUNA, and RAG integration.","tokens_in":20115,"tokens_out":4897,"duration_ms":44602,"significance":"If the reported gains are real, the method would be a practical contribution to selective answering and hallucination mitigation, and the quadrant-specific contrastive data construction is a plausible mechanism. The paper has useful strengths: it evaluates on two base models, includes a loss-combination ablation (Table 3), a threshold analysis (Table 2), and an independent ALCUNA experiment (Table 4). However, the evaluation is built on labels derived from the same base-model repeated-sampling procedure used to construct training data, and the reported Truthful Rate may largely measure how well the model reproduces the base model's own confidence pattern. The absence of error bars, the absence of strong baselines, and the very small independent gain on ALCUNA mean the central claim is not yet established.","major_comments":[{"comment":"The test-set known/unknown labels are never specified. Section 3.1 describes generating labels only for training via repeated sampling of the base LLM and thresholding by IK/IDK, while Appendix B.1 defines IK-IK and IK-IDK rates relative to questions \"labeled as model knows\" or \"model-don't-know\" in the test set. If the same self-derived labels are used for the 11,313 TriviaQA and 3,610 Natural Questions test instances, then a model that learns to answer exactly where the base model sampled correctly and to refuse exactly where it sampled incorrectly will attain a high Truthful Rate regardless of whether those labels reflect true knowledge. The authors must state how test labels were obtained, and should re-evaluate on externally labeled data (human annotation, ALCUNA-style artificial unknowns, or labels from a different model) to support the hallucination-reduction claim.","section":"Section 3.1 and Appendix B.1"},{"comment":"The IDK threshold is selected on the test set without a validation protocol, and no error bars or significance tests are reported. Table 2 shows IDK=0.7 maximizing Truthful Rate on the test set, which makes the headline numbers optimistic. In Table 1, the Mistral TriviaQA gain of IDK-SFT-Adpt-Ctr over IDK-SFT is 1.3 points (73.5 vs. 72.2); with 11,313 test questions, this difference may be within sampling noise. Multi-seed runs or bootstrap confidence intervals are needed before the 5.0-6.9 point claims can be taken at face value.","section":"Table 2 and Section 4.4"},{"comment":"The contribution of the \"adaptive\" weighting mechanism itself is not ablated. Table 3 ablates the quadrant-specific loss components, but every configuration uses the same max(lambda, detach(L_gen/L_ctr)) weighting; a fixed-weight contrastive loss, or a variant with lambda only, is needed to determine whether the adaptive term causes the improvement or whether the improvement comes from the combination of quadrant-specific contrastive data alone. Without this control, the central causal claim that the adaptive loss drives the gains is unsupported.","section":"Eq. (3) and Section 5.2"},{"comment":"The comparison set is too narrow. The baselines are only IDK-Prompting and IDK-SFT, both constructed by the authors; established methods for \"I don't know\" calibration, such as R-tuning (cited as Zhang et al. 2024c), and standard factuality benchmarks (e.g., TruthfulQA) are absent. The only independent evaluation, ALCUNA in Table 4, shows a 0.5 point Truthful Rate gain over IDK-SFT (98.3 vs. 97.8) on questions that are unknown by construction, and no significance test is given. This is a weak basis for the paper's strong conclusions about reducing hallucinations.","section":"Section 4.4 and Appendix C.1"}],"minor_comments":[{"comment":"The text says lambda is the \"upper bound of the weight,\" but max(lambda, detach(L_gen/L_ctr)) actually makes lambda a lower bound on the contrastive weight; the terminology should be corrected.","section":"Eq. (3) and Eq. (6)"},{"comment":"The sentence \"We select LLaMA-2-7B-chat and Mistral-7B-Instruct-v0.1 as base models for testing on TriviaQA and Natural Questions, respectively\" is ambiguous, since Table 1 reports both models on both datasets; clarify which model is evaluated on which dataset.","section":"Section 4.1.2"},{"comment":"Figure 4 contains four panels whose labels are inconsistent (\"Unknown Question\" vs. \"UnknownQuestions\", \"Known Question\" vs. \"KnownQuestions\"); clarify how the repeated-sampling accuracy categories are computed and why there are two versions of each plot.","section":"Section 5.3 and Figure 4"},{"comment":"The column header \"Knowing Rate\" is unclear; it should be \"IDK Rate\" or \"I Don't Know Rate\" to match the description in the text.","section":"Table 2"},{"comment":"The definition of IK-IK rate contains a typo: \"We calculate the number of correct answers given by the model as a percentage of the total number of questions in the dataset as a fraction of the IK-IDK rate\" should read \"as the IK-IK rate.\"","section":"Appendix B.1"},{"comment":"No code or data release link is provided, which hinders reproducibility of the data construction and training pipeline; please include the implementation details or a public repository.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The self-referential evaluation is the most serious concern. If the test labels are generated by the same repeated-sampling procedure used for training, the headline Truthful Rate improvements are not convincing evidence of reduced hallucination. I would ask for an external-label evaluation and a proper validation protocol before considering this work acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the method is a sensible extension of prior knowledge-boundary SFT: it divides questions into known, uncertain, and unknown via repeated-sampling accuracy, then builds quadrant-specific contrastive pairs with an adaptive loss weight. The ablations show each loss component contributes something, and the results are consistent across model sizes and datasets. Second, the evaluation has a circularity problem serious enough that I do not trust the headline 5–7 point Truthful Rate gains.\n\nWhat is actually new: prior work did knowledge-boundary SFT (Cheng et al., R-tuning) and contrastive instruction tuning (Yan et al.) separately. Combining them with four-quadrant data construction and a gradient-weighted contrastive loss is a real, if incremental, contribution. The experimental effort is also real: two base models, 13B scaling, ALCUNA, RAG integration, and loss ablations. The paper is transparent about its thresholds and hyperparameters.\n\nThe load-bearing issue is that the known/unknown labels for training come from the base model's own sampling accuracy (Section 3.1), and the paper never states how the test-set labels are obtained. If they come from the same procedure, then the \"Truthful Rate\" rewards the model for matching the base model's accuracy pattern, not for being correct against an external standard. A model trained to answer exactly where the base model sampled correctly and to refuse exactly where it sampled incorrectly would look good even if its actual knowledge were unchanged. The paper needs to state how test labels were generated, or better, evaluate on externally labeled sets where \"unknown\" is grounded. ALCUNA is the right direction, but it shows only 0.5 points over IDK-SFT. The RAG-Bench experiment is suggestive but still uses the same IK-IK/IK-IDK framing.\n\nSecond-order issues: the IDK threshold that maximizes Truthful Rate is chosen from the test set (Table 2), no error bars or multiple seeds are reported, and two key baselines the paper cites—Cheng et al. and R-tuning—are missing from the comparisons. Refusal is detected by string matching, which can be gamed. These are fixable, and with independent evaluation the method may well stand.\n\nWho this is for: people working on refusal behavior and hallucination mitigation via SFT will want to know this quadrant-aware contrastive setup. It is not a breakthrough, but it is a reasonable entry in a crowded space. My recommendation: I would send this to serious review, but only with a required revision that the known/unknown test labels be independent of the model's repeated-sampling labels and with external hallucination benchmarks added. As it stands, the headline numbers are optimistic.","headline":"Adaptive quadrant-aware contrastive SFT is a plausible and well-ablated method, but the reported Truthful Rate gains are not trustworthy until the known/unknown labels used for evaluation are independent of the model's own sampling.","tokens_in":20617,"tokens_out":2465,"would_cite":false,"duration_ms":24082,"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":"A training strategy that sorts an LLM's knowledge into known, uncertain, and unknown reduces hallucination by teaching each region differently.","keywords":["hallucination mitigation","knowledge boundaries","contrastive learning","instruction tuning","truthful rate","large language models","I don't know refusal","knowledge quadrants"],"falsifier":"Build a test set of questions where the base model repeatedly answers correctly with a confident but factually wrong answer (accuracy comfortably above $IK$), then train with IDK-SFT-Adpt-Ctr and check whether it still emits those wrong answers on a gold-labeled test. If Truthful Rate stays high while factual accuracy on those questions does not improve, the method is learning to lock in its own confidence pattern rather than refining knowledge.","tokens_in":1943,"feed_emoji":"🤖","tokens_out":2485,"duration_ms":75775,"temperature":0.7,"pith_summary":"Large language models hallucinate when they answer questions beyond their knowledge. This paper argues that the remedy is to help the model act on an accurate picture of what it knows: consolidate correct knowledge, strengthen half-learned knowledge, forget wrong knowledge, and answer \"I don't know\" when knowledge is absent. To do this it proposes Adaptive Contrastive Learning, which first sorts questions into known, uncertain, and unknown using the model's own repeated-sampling accuracy against two thresholds, $IK$ and $IDK$, and then builds different positive and negative training pairs for each sort. The reported effect is a Truthful Rate—correct answers plus correct refusals—of up to 78.2% versus 73.2% for supervised fine-tuning on TriviaQA with LLaMA-2-7B-chat, and up to 6.9 points over IDK-SFT on the out-of-distribution Natural Questions set.","feed_headline":"Contrastive training lifts honest LLM answers by 6.9 points","feed_subtitle":"Self-sampled confidence thresholds decide when to answer, consolidate, or say \"I don't know\".","key_machinery":"The central object is a two-threshold knowledge partition ($IK$, $IDK$) computed from the model's repeated-sampling accuracy, which splits questions into three regions: accuracy $\\ge IK$ means the model knows it knows, accuracy between $IDK$ and $IK$ means the model is uncertain, and accuracy $\\le IDK$ means the model does not know. On each region the paper builds contrastive instruction triplets (anchor, positive, negative) with quadrant-specific semantics, then optimizes an adaptive loss $L^{Adap} = L^{gen} + \\max(\\lambda, \\text{detach}(L^{gen}/L^{ctr})) L^{ctr}$, where $L^{gen}$ is standard cross-entropy generation loss and $L^{ctr}$ is a cosine-similarity contrastive loss that pulls positive hidden states together and pushes negative hidden states apart.","core_discovery":"On the paper's own terms, the discovery is that a single fine-tuning strategy can simultaneously improve both sides of honesty—giving correct answers when the model knows and refusing when it does not—by adapting contrastive learning to the model's current mastery. The authors construct knowledge quadrants from repeated-sampling accuracy with thresholds $IK$ and $IDK$, and show that treating each quadrant with a purpose-built contrastive loss raises Truthful Rate consistently across two base models and out-of-distribution datasets. For known questions, the correct answer is pulled toward the anchor while \"I don't know\" and wrong answers are pushed away; for uncertain questions, the golden answer is pulled in and wrong answers are pushed away; for unknown questions, \"I don't know\" is pulled in and wrong answers are pushed away. They also report that the full multi-quadrant loss beats any single-quadrant or pair combination, and that the method remains effective when combined with retrieval-augmented generation and when scaled to a 13B model.","pith_inferences":["If the model's repeated-sampling accuracy is an imperfect mirror of true knowledge—for example, if a model confidently repeats a plausible falsehood—then the training targets inherit that error, so part of the reported gain may be the model learning to reproduce its own confidence pattern rather than correcting facts; a direct test would measure factual precision on a set of consistently-wrong-but","Because Truthful Rate counts correct answers and correct refusals as interchangeable, a model that over-refuses can score well on unknown-heavy test sets; the choice of $IK$ and $IDK$ thresholds therefore does real work in the metric, not just in the training loss.","The adaptive loss weight $\\max(\\lambda, \\text{detach}(L^{gen}/L^{ctr}))$ implicitly up-weights the contrastive term when it is small relative to the generation loss, acting as a stabilizer against the contrastive signal vanishing—a ratio design that could transfer to other contrastive fine-tuning objectives.","Applying the same quadrant construction with an external knowledge oracle (such as retrieval or a trusted fact base) instead of self-sampling would separate \"knowing\" from \"thinking it knows,\" which is a natural next experiment to isolate where the gains actually come from."],"forward_implications":["On TriviaQA with LLaMA-2-7B-chat, the method lifts Truthful Rate from 73.2% with IDK-SFT to 78.2%, with the IK-IK rate (correct answers on known questions) rising 9.3 points even as correct refusals remain close to the SFT level.","On out-of-distribution Natural Questions, Truthful Rate rises 6.1 points over IDK-Prompting and 6.9 points over IDK-SFT for LLaMA, and 15.9 and 6.8 points for Mistral, indicating the learned answer-or-refuse boundary transfers beyond the training data.","Loss ablations show that consolidating uncertain knowledge contributes the most to correct answering, the known-knowledge loss makes the model more conservative, and the unknown-knowledge loss encourages forgetting wrong answers; only the full combination reaches the maximum Truthful Rate.","Combining with retrieval-augmented generation does not break the method: on RAG-Bench, retrieved context raises the method's Truthful Rate from 66.9% to 71.2%, while shifting some refusals into answers as expected.","Scaling from 7B to 13B parameters improves all metrics slightly for every method, and the proposed method keeps its advantage, suggesting the strategy composes with model scale."],"supporting_citations":[{"why":"Prior work on making AI assistants know what they don't know; the paper follows its setup for sampling responses and constructing known/unknown question data.","marker":"Cheng et al. (2024)"},{"why":"R-tuning work that instructs LLMs to say \"I don't know\"; supplies the knowledge-boundary and honest-refusal framing the paper extends.","marker":"Zhang et al. (2024c)"},{"why":"Contrastive instruction tuning; the hidden-state contrastive loss and instruction-pair construction the paper adapts for its quadrant losses.","marker":"Yan et al. (2024)"},{"why":"Source of the standard cross-entropy generation loss used alongside the contrastive term to preserve generative capability.","marker":"Liu et al. (2022b)"},{"why":"TriviaQA, the dataset from which the in-distribution training and test sets are built.","marker":"Joshi et al. (2017)"},{"why":"Natural Questions, used as the out-of-distribution test set.","marker":"Kwiatkowski et al. (2019)"},{"why":"Source of the knowledge-quadrant framing (known knowns, known unknowns, unknown knowns, unknown unknowns) used to describe model knowledge.","marker":"Yin et al. (2023b)"}],"fun_headline_variants":["Adaptive contrastive learning refines LLM knowledge and honesty","LLMs learn to say 'I don't know' via adaptive contrastive learning","Self-paced contrastive learning makes LLMs more truthful","Adaptive contrastive learning: know it or say 'I don't know' for LLMs"],"cache_read_input_tokens":22784,"weakest_assumption_plain":"The load-bearing premise is that a question's knowledge status—known, uncertain, or unknown—is faithfully measured by how often the model itself answers it correctly across repeated samples; if that self-accuracy proxy is wrong, the positive and negative training pairs encode the model's own confidence rather than true knowledge.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive contrastive learning refines LLM knowledge and honesty","LLMs learn to say 'I don't know' via adaptive contrastive learning","Self-paced contrastive learning makes LLMs more truthful","Adaptive contrastive learning: know it or say 'I don't know' for LLMs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000561,"raw_usage":{"total_tokens":2654,"prompt_tokens":928,"completion_tokens":1726,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":1646}},"tokens_in":544,"tokens_out":1726,"duration_ms":12722,"temperature":1.0,"reasoning_tokens":1646,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T13:31:35.431475+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a test set of questions where the base model repeatedly answers correctly with a confident but factually wrong answer (accuracy comfortably above $IK$), then train with IDK-SFT-Adpt-Ctr and check whether it still emits those wrong answers on a gold-labeled test. If Truthful Rate stays high while factual accuracy on those questions does not improve, the method is learning to lock in its own confidence pattern rather than refining knowledge.","supporting_citations":[{"cited_title":"Weld, and Luke Zettlemoyer","cited_arxiv_id":null,"evidence_quote":"TriviaQA, the dataset from which the in-distribution training and test sets are built."},{"cited_title":"Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming - Wei Chang, Andrew M","cited_arxiv_id":null,"evidence_quote":"Natural Questions, used as the out-of-distribution test set."}],"review_version":1}