{"id":"ac48c3f2-7766-45c3-b6b3-1c5ff1e50fa6","arxiv_id":"2412.02588","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ExpCTR jointly trains an LLM explanation generator and a CTR model via two reinforcement rewards, reporting higher AUC than ID-based baselines on BookCrossing, ML-20M, and Amazon Books.","lead":"ExpCTR is a new method that uses a large language model to write a one-sentence explanation of why a user would like a product, then feeds that sentence back into the click prediction model as an extra input. It reports better prediction accuracy and explanations on three real-world datasets without needing hand-written explanation data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Co-adaptation between the explanation generator and the Stage-2 CTR model makes the reported ExpCTR-Aug AUC gains—and with them the accuracy/interpretability claim—unverified; the LC reward also optimizes the exact metric reported for ExpCTR-LLM.","rationale":"The reader's weakest assumption is essentially the one I would attack: the rewards are not validated as measures of explanation quality, and the IC reward in particular is computed with a CTR model that is itself trained on the generator's output. I agree with that diagnosis and with the conditional verdict. My stress-test sharpens it in two ways. First, the co-adaptation is not merely a possible concern about spurious text patterns; it is built into the training loop because Stage 3 optimizes the generator against the exact Stage 2 model that will later be used for evaluation. Second, the LC reward is even more directly circular in evaluation: the ExpCTR-LLM metric is the same label-agreement function that PPO maximizes, so the reported improvement over ICL is partly a statement about optimization success, not about explanation fidelity. These are correctness risks, not internal inconsistencies, and they can be settled empirically. The paper also lacks code, error bars, and a comparison to the closest RL-based baseline [40], all of which the reader noted. Because the proposed tests are straightforward and the framework may still be salvageable, keeping the conditional verdict is appropriate rather than rejecting outright.","tokens_in":15626,"tokens_out":6302,"duration_ms":72248,"concrete_test":"Hold the final explanation generator fixed after Stage 3 and retrain the CTR model f from scratch on a training subset of the generated explanations only, with no further PPO updates to the generator. Evaluate this freshly trained, non-co-adapted f on the test split; compare its AUC with the reported ExpCTR-Aug and with DeepFM. If the non-co-adapted gain over DeepFM collapses (e.g., drops by more than one AUC point or becomes negative), the Stage-2/3 co-adaptation is responsible for the headline gains. To also check the LC side, score the frozen generator's explanations with a different LLM judge (e.g., LLaMA-3-70B or GPT-4) and recompute AUC; if the ICL-relative gain does not transfer, the explanations are specialized to the base judge rather than faithful to user reasoning.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that ExpCTR 'significantly enhances both recommendation accuracy and interpretability' depends on the reward signals in Eqs. (9) and (13) measuring genuine explanation quality. The most load-bearing failure mode is the IC loop in Section 4.3.2. R_IC(Z) = 1 - |y - s^r| + |s^r - s_tilde^r| is evaluated with the CTR model f trained in Stage 2 on the explanations produced by the generator being optimized in Stage 3. Maximizing R_IC therefore rewards any textual feature—including spurious n-gram or embedding patterns—that helps f predict the training label. The reported ExpCTR-Aug results are computed with that same co-adapted f (plus the final generator), so the AUC gains over DeepFM (18.2%, 11.9%, 17.8%) conflate explanation value with mutual adaptation. The LC component has an analogous circularity: ExpCTR-LLM's reported AUC is the AUC of s in Eq. 8, which is exactly the quantity whose label agreement R_LC in Eq. 9 maximizes; hence the improvement over ICL is partly by construction. Interpretability is never independently measured: no human evaluation, no faithfulness or grounding metric, only two case studies in Table 2. Therefore the paper's dual claim is currently unsupported as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ExpCTR, a framework that integrates LLM-generated textual explanations into CTR prediction. A generator LLM is fine-tuned with LoRA and PPO using two rewards: LC alignment (Eq. 9), which rewards explanations that make a frozen LLM's binary like/dislike prediction agree with the ground-truth label, and IC alignment (Eq. 13), which rewards explanations that change the prediction of an ID-based CTR model (DeepFM) toward the label. The training proceeds in three stages: LC-alignment fine-tuning of the generator, training the CTR model with generated explanations as extra text features, and IC-alignment fine-tuning of the generator; the last two stages are repeated iteratively. The paper reports AUC/LogLoss/MAE/RMSE on BookCrossing, ML-20M, and Amazon Books, plus two qualitative case studies, and claims that ExpCTR significantly improves both recommendation accuracy and interpretability without manually constructed explanation datasets.","tokens_in":15919,"tokens_out":5394,"duration_ms":52842,"significance":"If the claims were fully supported, ExpCTR would be a notable step toward joint training of explanation generation and CTR prediction, and the proposed two-reward RL scheme is interesting. The paper is also to be credited for making the method concrete, using public datasets, and reporting multiple metrics. However, the central evidence is currently undermined by a circular reward loop in the IC alignment stage and by the absence of any quantitative validation of interpretability. The reported accuracy gains are therefore not yet established as gains in explanation quality, and the interpretability claim rests entirely on two case studies. With additional ablations and an independent evaluation of explanation faithfulness, the method could be made convincing, but as it stands the paper's dual claim is unverified.","major_comments":[{"comment":"The IC alignment reward is circular. In Stage 2, the CTR model f is trained on explanations produced by the generator πRL_φ; in Stage 3, R_IC = 1 − |y − s^r| + |s^r − s̃^r| is computed with that same f and used to update the very generator that produced the training explanations. Because f is trained to exploit any textual pattern that predicts the label, the second term |s^r − s̃^r| rewards any text feature that changes f's prediction, including spurious n-gram or embedding correlations. The ExpCTR-Aug gains over DeepFM reported in Table 1 (18.2%, 11.9%, 17.8% AUC) are thus computed with the co-adapted f and conflate explanation value with mutual adaptation. To support the accuracy claim, the authors should compare ExpCTR-Aug against DeepFM augmented with non-reward text features (e.g., random item descriptions, template sentences, or frozen ICL explanations) using the same concatenation and training recipe, and should also report the Stage-2 CTR model's performance before any Stage-3 RL update.","section":"§4.3.2, Eq. (13), and Table 1"},{"comment":"The reported ExpCTR-LLM AUC is exactly the AUC of s^u_{u,i} in Eq. (8), which is the score whose label agreement R_LC in Eq. (9) maximizes. The PPO objective therefore directly optimizes the evaluation metric, so the improvement over ICL (e.g., 9.1% on ML-20M) is partly by construction. This does not by itself invalidate the method, but it means the result cannot be read as evidence that the explanations reflect user intentions. An independent evaluation is needed, such as a human rating of explanation helpfulness, a faithfulness metric computed against the user's actual review text, or a control experiment in which the same R_LC reward is applied to a non-explanatory text input to see whether the AUC gain is attributable to the explanation content.","section":"§4.2.2, Eq. (9), and §5.2, ExpCTR-LLM"},{"comment":"The interpretability claim is not quantitatively validated. The paper provides only two qualitative case studies in Table 2; there is no human evaluation, no agreement metric with user reviews, no explanation ranking metric, and no faithfulness/grounding measure. The abstract states that ExpCTR 'significantly enhances both recommendation accuracy and interpretability,' but interpretability is asserted rather than measured. The authors should add a human study or an automatic fidelity evaluation (e.g., entailment/overlap with the user review, or feature attribution agreement) to support the interpretability part of the central claim.","section":"§5.4, Table 2"},{"comment":"The paper cites Yang et al. (2024) [40], an RL-based explainable recommendation method, as related work but does not include it in the compared methods of Section 5.1.2 or in Table 1. Since ExpCTR is also an RL-based explainable method and the paper claims state-of-the-art performance, omitting [40] leaves the comparison incomplete. Please add this baseline or justify its exclusion.","section":"§2 and §5.1.2"}],"minor_comments":[{"comment":"References [44] and [45] are the same survey (Yongfeng Zhang et al., 'Explainable recommendation: A survey and new perspectives'); one should be removed or the citations should be merged.","section":"References"},{"comment":"The text lists DSSM among the ID-based methods, but DSSM does not appear in Table 1; conversely, xDeepFM appears in Table 1 but is not described in the compared-methods list. Please align the list of baselines with the table.","section":"§5.1.2 vs. Table 1"},{"comment":"The notation in Eq. (8) is typeset incorrectly: the subscript appears as \"𝑠𝑢𝑢,𝑖\" and should be something like s^u_{u,i}; also, the superscripts for s^r and s̃^r in Eq. (13) should be defined clearly before first use.","section":"Eq. (8)"},{"comment":"The prompt template in Section 4.1 instructs 'The answer should be within one sentence,' but the explanations shown in Table 2 are multiple sentences long. Please either update the template or the examples so they match.","section":"§4.1 vs. Table 2"},{"comment":"The term 'data-free' is misleading: the method still requires user-item interaction data and item titles; it avoids only manually constructed explanation datasets. Rewording to 'explanation-data-free' would be more precise.","section":"Abstract and §1"},{"comment":"No variance or statistical significance tests are reported. Given that several baseline AUC differences are small (e.g., 0.02–0.05), the claim of 'significantly enhances' should be supported by results over multiple seeds or significance tests.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the circularity between the IC reward and the CTR model used for evaluation; this makes the headline AUC gains uninterpretable without additional ablations. The LC-reward evaluation is also self-referential. I would encourage the editor to request the proposed control experiments and an independent interpretability evaluation before considering publication. The missing comparison with the cited RL-based explainable baseline is another point reviewers will likely raise."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look, but read Table 1 with suspicion. The core idea—train an LLM to generate explanations that are then used as text features in an ID-based CTR model, with two RL rewards to align them—is genuinely new as a combination. The three-stage LoRA/PPO loop is clearly described, and the authors are upfront that they want to escape the post-hoc, curated-explanation-dataset paradigm. That's a reasonable research direction and the paper does real engineering.\n\nWhat the paper does well: it attacks a real problem, the setup is sensible, and the experiments span three datasets with a lot of baselines. The LC reward, which uses a frozen LLM's Yes/No label probability as a quality signal, is a clever trick. The IC reward, which rewards the CTR model's prediction change when the explanation is added, is also a defensible signal in principle.\n\nThe soft spots are serious but mostly fixable. First, the IC loop has a co-adaptation problem: f is trained in Stage 2 on explanations from the generator, then the generator is optimized in Stage 3 against that same f, and the reward includes |s_r - s_tilde|. That can reward spurious text patterns. The reported ExpCTR-Aug gains over DeepFM (18%, 12%, 18% AUC) are therefore unverified. The paper needs an ablation with frozen or random text embeddings, and it needs error bars. Second, interpretability is never measured. There is no human evaluation, no faithfulness or grounding metric, only two cherry-picked case studies. That's not enough to support 'significantly enhances interpretability.' Third, the LC reward is not circular in the way the stress-test note claims—every supervised model is evaluated on the objective it was trained on—but it does mean ExpCTR-LLM is just a supervised LLM scorer, so beating zero-shot ICL is not evidence of explanation quality. Fourth, the closest prior work [40] is cited but not compared; that's a notable omission. No code is released either.\n\nWho's this for? Researchers working on explainable recommendation or LLM-based CTR. It deserves a serious referee, but with a request for major revision: add the ablations, report variance, run a real interpretability evaluation, and include [40]. As it stands, the central research idea is worth engaging with; the empirical claims are not yet supported.","headline":"ExpCTR is a plausible framework for joint CTR and explanation generation, but the paper's dual accuracy-plus-interpretability claim is only half-supported: the accuracy gains need ablations and variance, and interpretability is never actually measured.","tokens_in":16428,"tokens_out":3236,"would_cite":false,"duration_ms":33119,"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":"ExpCTR jointly trains LLM explanations with the CTR model and reports higher accuracy and interpretability on three real-world datasets.","keywords":["Large Language Models","Explainable Recommendation","CTR Prediction","Reinforcement Learning","Proximal Policy Optimization","LoRA","Data-free Training"],"falsifier":"Replace each generated explanation in the test set with a random sentence sampled from the same review corpus and recompute ExpCTR-Aug's AUC; if the AUC drops by less than about one point, the explanation features are not carrying the signal and the reward was optimizing for correlation rather than explanation content.","tokens_in":15420,"feed_emoji":"🎯","tokens_out":9064,"duration_ms":79471,"temperature":0.7,"pith_summary":"The paper argues that the standard post-hoc pipeline for explainable recommendation—train a black-box CTR model, then fit a separate explanation generator—can be replaced by a joint framework in which the explanations themselves are trained to serve the CTR task. ExpCTR uses reinforcement learning (PPO) with two rewards: one that rewards an explanation when a frozen LLM can predict the user's actual interaction from it, and one that rewards an explanation when adding it as a textual feature improves an ID-based CTR model's prediction. The authors claim that this joint training removes the need for curated explanation datasets while improving both classification accuracy (AUC) and explanation quality on BookCrossing, MovieLens-20M, and Amazon Books. If the claim holds, explanation generation stops being an afterthought and becomes a trainable component of the recommender itself, with direct implications for trustworthy recommendation.","feed_headline":"Training LLM explanations with the CTR model lifts AUC up to 18%","feed_subtitle":"A two-reward PPO loop aligns machine reasons with user intent and recommender logic; no explanation dataset needed.","key_machinery":"The load-bearing object is the two-part reward in the PPO loop, computed by two predictors that stay fixed while the generator is updated. The LC head (Eq. 9) scores an explanation by how much it moves a frozen LLM's binary Yes/No prediction toward the true label; the IC head (Eq. 13) scores it by how much adding its text embedding to an ID-based CTR model improves prediction relative to a zero-embedding counterfactual. These normalized rewards (Eq. 10) feed a three-stage iterative training process—LC alignment, CTR training with textual features, and IC alignment—that closes the feedback loop between explanation and recommender.","core_discovery":"ExpCTR's central claim is that a single LLM, fine-tuned with LoRA under a PPO objective, can generate explanations that are simultaneously faithful to user intent and useful to the recommender itself. The LC alignment reward (Eq. 9) checks whether the explanation, fed to a frozen LLM as the user's thought, makes that LLM's like/dislike prediction match the ground-truth label; the IC alignment reward (Eq. 13) checks whether appending the explanation's mean-pooled embedding to an ID-based CTR model moves its prediction toward the label, compared with zeroing the embedding. The paper reports that ExpCTR-Aug, the CTR model trained with the generated explanations as features, outperforms the DeepFM baseline by 18.2%, 11.9%, and 17.8% in AUC on BookCrossing, ML-20M, and Amazon Books, and that the case-study explanations align with user reviews more closely than an in-context-learning baseline. This constitutes, the authors claim, the first data-free joint training of explanations and CTR prediction.","pith_inferences":["A testable boundary of the claim: on datasets where review text is only weakly correlated with interactions, the IC reward should plateau and the reported AUC gains should shrink, which would delimit when joint training helps.","The LC reward makes the frozen LLM's own judgment the arbiter of 'user intent', so the method inherits that base model's biases; subgroups on which the base LLM is miscalibrated could receive confidently wrong explanations.","Nothing in the two rewards prevents the generator from producing verbose, template-like text that is predictive but not human-understandable; a readability or human-preference term would be a natural extension.","The paper does not isolate how much of the gain comes from the RL loop versus the LLM prior; comparing against a small review-trained generator with a similar parameter budget would separate the two."],"forward_implications":["CTR models can treat LLM-generated free text as a first-class prediction feature, with no curated explanation dataset.","LoRA keeps the tuning cheap: one base LLM serves as the generator, the LC reward model, and the frozen encoder for explanations.","The three-stage loop makes explanation quality and CTR quality co-dependent, so improving one can improve the other in later rounds.","The method's prompt template is domain-transferable to any setting where items have textual titles and explicit like/dislike signals, not only books."],"supporting_citations":[{"why":"Supplies the PPO-with-KL-penalty recipe (Eq. 7) and the InstructGPT-style reward-model training that ExpCTR adapts to explanation generation.","marker":"[26]"},{"why":"The proximal policy optimization algorithm used as the RL backbone for the three-stage training.","marker":"[30]"},{"why":"LoRA, the low-rank adapter method that makes lightweight fine-tuning of the LLM generator feasible.","marker":"[12]"},{"why":"DeepFM, the ID-based CTR model used as the backbone for the IC alignment reward and as the base for ExpCTR-Aug.","marker":"[22]"},{"why":"The frozen pre-trained language model used to produce dense embeddings of generated explanations (Eq. 11).","marker":"[39]"},{"why":"Justifies the normalization and clipping of reward values (Eq. 10) for stable PPO training.","marker":"[47]"},{"why":"TALLRec, a baseline method and the source of the stratified random sampling protocol used to construct the datasets.","marker":"[1]"}],"fun_headline_variants":["LLM explanations lift CTR prediction by 18% AUC","Data-free LLM reasoning improves CTR prediction up to 18%","Two-reward LLM training with CTR model: up to 18% AUC gain","ExpCTR: LLM reasonings aligned with recommender to boost AUC 18%","First data-free joint training of LLM explanations and CTR: 18% AUC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training assumes that 'predictive usefulness' is a faithful stand-in for 'explanation quality': the LC reward relies on the frozen LLM's Yes/No verdict reflecting the user's real reasoning, and the IC reward relies on the ID-based CTR model's score change measuring truthfulness rather than exploitability.","fun_headline_variants_meta":{"raw":{"variants":["LLM explanations lift CTR prediction by 18% AUC","Data-free LLM reasoning improves CTR prediction up to 18%","Two-reward LLM training with CTR model: up to 18% AUC gain","ExpCTR: LLM reasonings aligned with recommender to boost AUC 18%","First data-free joint training of LLM explanations and CTR: 18% AUC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001437,"raw_usage":{"total_tokens":5783,"prompt_tokens":926,"completion_tokens":4857,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":4756}},"tokens_in":542,"tokens_out":4857,"duration_ms":30565,"temperature":1.0,"reasoning_tokens":4756,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:17:36.280888+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace each generated explanation in the test set with a random sentence sampled from the same review corpus and recompute ExpCTR-Aug's AUC; if the AUC drops by less than about one point, the explanation features are not carrying the signal and the reward was optimizing for correlation rather than explanation content.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Justifies the normalization and clipping of reward values (Eq. 10) for stable PPO training."}],"review_version":1}