{"id":"fb41b3cf-66ac-48bf-a0cd-40f60fca5834","arxiv_id":"2502.09604","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"SelfCite uses context-ablation probability differences as a self-supervised reward to improve LLM sentence-level citations, raising LongBench-Cite citation F1 from 73.8 to 79.1.","lead":"This paper introduces SelfCite, a method that scores an LLM's candidate citations by measuring how the model's probability of generating a statement changes when cited evidence is removed or isolated. This self-supervised reward raises citation F1 on the LongBench-Cite benchmark by over 5 points, suggesting verifiable AI answers can be trained without human labels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unreported evaluation noise and an unexplained 1.8-F1 gap versus the official LongCite-8B baseline leave the headline 5.3-F1 gain unverified; variance estimates and a baseline reconciliation are needed.","rationale":"The reader's weakest_assumption focuses on the transfer from contributive attribution to corroborative citation quality. That is a real conceptual limitation, and the paper acknowledges it in Section 5. However, the empirical evidence on five datasets suggests the two objectives at least partially align, and the ablations (Table 3) show that the reward outperforms simple length-based selection, so the conceptual concern, while real, is not the most load-bearing threat to the central claim. The more immediate threat is the reliability of the numerical headline: a 5.3-F1 gain with no reported variance, measured against a self-reproduced baseline that is 1.8 points higher than the official published number. If evaluation noise is on the order of 1-3 points, the headline could be within noise; if the reproduction differs from the official pipeline, then the comparison to all baselines in Table 1 is on unequal footing. These issues are concrete, directly tied to the central claim, and fully testable. The reader's rationale does mention the baseline gap and the lack of error bars, but the reader identified a different weakest_assumption as primary. I therefore mark agreement as partial. My recommendation is UNCHANGED because the paper is solid enough to warrant conditional acceptance, but the authors should supply the missing variance estimates and baseline reconciliation before full acceptance.","tokens_in":24204,"tokens_out":12692,"duration_ms":136429,"concrete_test":"Run the three key configurations (LongCite-8B reproduced, +BoN, +SimPO then BoN) three times with different generation seeds, and evaluate each resulting output five times with independent GPT-4o judge calls using the official LongBench-Cite evaluation script. Report mean and standard deviation of citation F1 over the 15 evaluations for each configuration. Separately, rerun the released LongCite-8B checkpoint with the exact decoding hyperparameters from the paper (temperature 0.95, top-p 0.7) and with the official benchmark's evaluation pipeline; document all differences from the published 72.0 F1 (e.g., judge model version, prompt text, temperature, example count).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim is the 5.3-point F1 improvement of SelfCite + SimPO then BoN over the LongCite-8B baseline on LongBench-Cite (Table 1). Two unaddressed issues make this number insecure. First, no error bars are reported anywhere; the evaluation depends on stochastic generation (temperature 0.95, top-p 0.7, and even higher temperature for BoN candidates) and on a GPT-4o judge that is itself nondeterministic. LongBench-Cite datasets are small enough that a few judge disagreements could move F1 by several points. Second, the reproduced LongCite-8B baseline in Table 1 is 73.8 F1, while the published LongCite-8B row (Zhang et al., 2024, also in Table 1) is 72.0. The paper never explains this 1.8-point gap. If the authors' evaluation pipeline is systematically more lenient than the official one, then the absolute comparison to all previously published baselines is distorted, and the 5.3-point gain is measured against a reproduction whose provenance is unclear. Without a demonstration that the gain exceeds the evaluation noise and that the baseline gap is understood (e.g., different GPT-4o version, prompt, decoding, or judge temperature), the central claim is not fully verified. This is a load-bearing concern because the paper's contribution is specifically an empirical improvement; the conceptual contributive-to-corroborative tension is acknowledged in Section 5 and partially mitigated by the consistent gains across five datasets, but the numerical headline has no such mitigation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SelfCite, a self-supervised method for improving sentence-level citation quality in long-form question answering. A reward is computed from the language model's own probabilities under context ablation: removing the cited sentences should lower the probability of the response statement (necessity), and retaining only the cited sentences should keep it high (sufficiency); the two scores are summed into a single reward (Eq. 1). This reward is used in two ways: best-of-N reranking of citation candidates (Section 2.3) and SimPO preference optimization (Section 2.4). Experiments on LongBench-Cite across five datasets report that applying best-of-N to a SimPO-finetuned LongCite-8B raises average citation F1 from a reproduced baseline of 73.8 to 79.1, a gain of 5.3 points, and that the SimPO model alone reaches 77.9 F1. The paper also includes a fully self-supervised setting where an SFT model is trained on ContextCite-generated citations, ablations of the reward components, length balancing, training-size studies, and a comparison against an NLI-reward baseline.","tokens_in":24526,"tokens_out":3653,"duration_ms":40746,"significance":"If the empirical claims hold, this is a valuable result: it shows that citation quality can be improved without human annotation or proprietary API supervision at the alignment stage, using a simple and interpretable probability-based reward. The paper includes several commendable controls: the NLI-reward SimPO baseline uses the same pipeline with only the reward changed; the ablation in Table 3 fixes the candidate set and varies only the reward; the length-balancing ablation isolates a known shortcut; and the authors explicitly acknowledge the contributive-versus-corroborative mismatch in Section 5. The code is released, which aids reproducibility. The main weakness is that the headline numerical claim is supported by single-run point estimates with an unexplained gap between the reproduced baseline and the published LongCite-8B number, so the central quantitative claim is not yet fully verified.","major_comments":[{"comment":"The paper reports no variance or significance information for any of the main F1 numbers. Generation is stochastic (temperature 0.95, top-p 0.7, and temperature 1.2 for best-of-N sampling), and the citation F1 judge is a nondeterministic GPT-4o model. On the dataset sizes typical of LongBench-Cite, a small number of judge disagreements or one unlucky seed could plausibly move average F1 by several points. Please report variance across multiple seeds, a bootstrap over evaluated instances, or at least a judge-sensitivity analysis, and state whether the reported numbers are averages over repeats.","section":"Section 3.4, Table 1"},{"comment":"The reproduced LongCite-8B baseline is 73.8 average F1, while the published LongCite-8B row is 72.0. This 1.8-point gap is never explained. If the reproduction pipeline is systematically more lenient (different GPT-4o version, different judge prompt, different decoding configuration), then the absolute comparison to all published baselines is distorted. Please reconcile the gap by running the official evaluation code, documenting any pipeline differences, and showing that the 5.3-point gain over the reproduction is robust to those differences.","section":"Table 1, LongCite-8B rows"},{"comment":"The paper's central assumption is that the contributive reward transfers to the corroborative evaluation metric. The authors state this alignment is 'at least partially aligned' and 'not guaranteed,' which is honest, but the empirical evidence for transfer is currently only the five point estimates in Table 1. Without per-dataset error bars or a simple sign test over the five datasets, it is hard to know whether the consistent direction of improvement is statistically reliable or consistent with noise. Please add a per-dataset uncertainty estimate or another explicit transfer test.","section":"Section 5 and Section 3.4"}],"minor_comments":[{"comment":"The sentence 'it depends on larger proprietary models two proprietary APIs' is missing a word; it should read 'it depends on two larger proprietary APIs' or similar.","section":"Section 5, Related Work"},{"comment":"The footnote reports that after deduplication only 4.8 candidates remain per statement on average, but the text still refers to N=10 candidates. Please clarify whether the reward selection operates over the deduplicated set and whether the effective candidate count varies across statements.","section":"Footnote 1, Section 3.1"},{"comment":"The caption says citation lengths are computed across all LongBench-Cite datasets, but the table has a single 'Citation Length' column next to HotpotQA-specific R/P/F1. Please clarify in the caption that R/P/F1 are for HotpotQA while citation length is the LongBench-Cite average, or split them into separate columns.","section":"Table 3, caption"},{"comment":"The hyperparameter selection for SimPO is described as choosing a learning rate on a 100-example development set, but no development-set results are shown. Reporting the selected learning rate and the sensitivity of the final F1 to it would help reproducibility.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The paper is well-aligned with the current interest in self-supervised grounding and attribution, and the method is simple enough to be a useful baseline. The author overlap with ContextCite is disclosed and the reuse of the context-ablation idea is handled fairly. The main concern is the reliability of the headline number: no error bars and an unexplained baseline gap. I would accept a major revision that adds variance estimates and a baseline reconciliation; if the authors can show the 5.3-point gain is robust to these factors, the result would be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a solid empirical contribution. The reward is simple and interpretable, the ablations are carefully designed, and the gains are consistent across five datasets. The headline number, though, is not fully verified yet because of evaluation noise and an unexplained baseline gap. I would send it to peer review, but with a request for variance estimates and a baseline reconciliation.\n\nWhat is actually new: applying a context-ablation necessity/sufficiency reward to rerank citation candidates (best-of-N) and to train with SimPO, plus a length-balancing trick that keeps the model from just inflating citation counts. The ablation study in Table 3 is genuinely informative: reranking by LM log-prob or by max length fails, the two reward components each help, and the combination wins. The NLI-reward baseline is a fair controlled comparison, and the fully self-supervised setting (starting from Llama-3.1-8B-Instruct with ContextCite-generated SFT data) gives a useful picture of what the method does without LongCite's supervised data. The code is available, which is real evidence.\n\nThe soft spots are real. No error bars anywhere; evaluation uses stochastic generation (temperature 0.95, top-p 0.7; BoN at 1.2) and a nondeterministic GPT-4o judge. LongBench-Cite is small, so a few judge disagreements could move F1 by a point or two. More concretely, the reproduced LongCite-8B baseline is 73.8 F1, while the published number is 72.0, and the paper does not explain that 1.8-point gap. That does not invalidate the internal comparison against the authors' own reproduction, but it does mean the absolute comparison to prior published numbers is distorted, and the \"5.3-point improvement\" is measured against a reproduction whose provenance is unclear. The authors should reconcile this by reporting the exact evaluation setup or by running both pipelines on the same outputs.\n\nThe conceptual worry about contributive versus corroborative citation is acknowledged in Section 5, and the consistent gains across datasets partially mitigate it. I do not see a fatal flaw. The math is simple, the baseline choices are fair, and the paper is honest about limitations. The main missing piece is a robustness analysis.\n\nThis paper is for researchers working on attribution, citation generation, and retrieval-augmented generation. They will find the method reproducible and the ablations useful. I would cite it if I worked on this line.\n\nRecommendation: accept for peer review, with a request for multiple runs with error bars and a clear explanation of the baseline gap. The core idea is sound and deserves serious engagement.","headline":"A clever, well-ablated self-supervised citation reward that deserves review, but the headline F1 gain needs an error-bar and baseline-reconciliation pass.","tokens_in":25069,"tokens_out":3031,"would_cite":true,"duration_ms":31022,"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":"SelfCite lifts LLM citation F1 by 5.3 points with no human labels","keywords":["self-supervised alignment","context attribution","citations","LongBench-Cite","context ablation","best-of-N sampling","preference optimization","SimPO"],"falsifier":"Measure, on a held-out set of LongBench-Cite statements with human citation-quality judgments, the rank correlation between the SelfCite reward and human citation precision/recall: if the reward's ranking of alternative citation candidates does not positively correlate with human scores (or if inserting an irrelevant but causally influential sentence into a citation increases the reward), the central claim that the ablation reward tracks citation quality is falsified.","tokens_in":24022,"feed_emoji":"📚","tokens_out":5249,"duration_ms":45115,"temperature":0.7,"pith_summary":"The paper claims that citation quality in long-form question answering can be improved substantially using only the model’s own output probabilities, with no human or API annotations during the alignment stage. Its core idea is a self-supervised reward computed by context ablation: removing the cited sentences should make the response less probable (necessity), and keeping only the cited sentences should preserve it (sufficiency). Applied as best-of-N reranking over candidate citation spans, and then distilled into the model with SimPO preference optimization, the reward raises citation F1 by up to 5.3 points on LongBench-Cite across five long-context QA tasks. A fully self-supervised variant, starting from a model whose citations were generated by ContextCite rather than human data, also improves. If correct, the result means citation quality can be bootstrapped without expensive annotation, and smaller open models can approach proprietary citation pipelines.","feed_headline":"SelfCite lifts LLM citation F1 by 5.3 points with no human labels","feed_subtitle":"A context-ablation reward lets an 8B model improve citation accuracy without annotation or proprietary APIs.","key_machinery":"The load-bearing object is the context-ablation reward pair: a necessity score (probability drop when the cited sentences are removed) plus a sufficiency score (probability hold when only the cited sentences remain), which telescope into Eq. (1). This reward is computed with the language model in eval mode, requires no trained reward model or annotation, and is used in two ways: (i) best-of-N sampling reranks N citation-span candidates per statement by the reward; (ii) preference optimization with SimPO uses the BoN-improved responses as chosen examples to internalize the behavior. The token-length cap (384 tokens unless the citation is a single sentence) prevents the model from gaming the reward by citing long passages.","core_discovery":"The central discovery is that the difference between the log-probability of a response statement given only the cited sentences and given the context minus the cited sentences—$\\text{Reward}(e_i) = \\log p_{LM}(r_i \\mid E_i) - \\log p_{LM}(r_i \\mid C \\setminus E_i)$—is a sufficient reward signal to rerank candidate citations. The paper shows that selecting, for each statement, the candidate citation that maximizes this reward (with a token-length cap) improves both citation recall and precision, pushing average citation F1 from 73.8 to 77.5 on LongBench-Cite. Training with SimPO on preference pairs built from best-of-N selections matches this gain in a single forward pass (77.9 F1), and applying best-of-N again on the SimPO model reaches 79.1 F1, an improvement of 5.3 points over the base model. The authors interpret this as evidence that the contributive, causality-oriented attribution signal is at least partially aligned with the corroborative, support-based citation quality measured by the benchmark.","pith_inferences":["The reward measures causal contribution to the model's own generation, not semantic support; on claims where the model relies on spurious or biased context signals, the reward could rank a citation that triggers the response above one a human would judge as evidence. The paper itself notes this contributive/corroborative mismatch is not guaranteed to align.","Because the method requires access to token-level output probabilities, it cannot be applied directly to closed-source APIs; a testable extension would be to distill a probability-estimating surrogate or to approximate the reward with sampling-based estimates.","The ablation-based reward could serve as a general-purpose attribution quality metric beyond citations—for example, evaluating RAG faithfulness or selecting evidence in retrieval pipelines—since it is annotation-free and model-specific.","A concrete stress test: if the reward is used to filter cited sentences generated by a deliberately hallucinating model, does citation F1 still improve without answer correctness collapsing? The paper reports correctness is unchanged, but the test would bound the method's safety."],"forward_implications":["Citation alignment can be performed without human annotation or proprietary APIs, since the reward is computed from the model's own probabilities.","Best-of-N sampling alone yields immediate citation gains (73.8 to 77.5 F1) at inference time, with only short citation spans being resampled, so latency scales only modestly.","SimPO fine-tuning internalizes the BoN gains, giving equal or better citation quality in a single forward pass (77.9 F1) and eliminating BoN's latency overhead.","A fully self-supervised pipeline (ContextCite-generated SFT data plus SelfCite) cuts the gap to the supervised LongCite-8B from 8.1 to 2.3 F1 points (65.7 to 71.5 with BoN), suggesting that human or SFT annotation may be replaceable.","Iterative SimPO over three rounds keeps improving F1, indicating the reward remains reliable under the model's evolving output distribution."],"supporting_citations":[{"why":"Supplies the LongBench-Cite benchmark, the LongCite-8B base model, and the LongCite-45K data and citation format that SelfCite starts from and is evaluated on.","marker":"Zhang et al. (2024)"},{"why":"Provides ContextCite, the contributive-attribution baseline and the random-ablation method that motivates SelfCite's necessity/sufficiency reward.","marker":"Cohen-Wang et al. (2024)"},{"why":"Supplies SimPO, the reference-free preference optimization objective used to internalize the best-of-N gains into the model.","marker":"Meng et al. (2024)"},{"why":"Provides the NLI-based fine-grained citation rewards used as the strongest fine-tuned baseline, which SelfCite outperforms.","marker":"Huang et al. (2024a)"},{"why":"Justifies using best-of-N sampling as the standard oracle for validating reward design before fine-tuning.","marker":"Gao et al. (2023a)"}],"fun_headline_variants":["SelfCite: self-supervised citations boost F1 by 5.3","No labels needed: SelfCite improves citation F1 by 5.3","Context ablation reward powers SelfCite's 5.3 F1 gain","SelfCite teaches LLMs to cite via self-supervised signals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the change in the model's own next-token probability when context is ablated faithfully tracks whether a citation is necessary and sufficient, and that this contributive signal improves the external, corroborative citation F1 score measured by GPT-4o on LongBench-Cite; it also assumes the base model can already generate parseable citation candidates.","fun_headline_variants_meta":{"raw":{"variants":["SelfCite: self-supervised citations boost F1 by 5.3","No labels needed: SelfCite improves citation F1 by 5.3","Context ablation reward powers SelfCite's 5.3 F1 gain","SelfCite teaches LLMs to cite via self-supervised signals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000223,"raw_usage":{"total_tokens":1449,"prompt_tokens":928,"completion_tokens":521,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":442}},"tokens_in":544,"tokens_out":521,"duration_ms":5419,"temperature":1.0,"reasoning_tokens":442,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T20:52:17.077876+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure, on a held-out set of LongBench-Cite statements with human citation-quality judgments, the rank correlation between the SelfCite reward and human citation precision/recall: if the reward's ranking of alternative citation candidates does not positively correlate with human scores (or if inserting an irrelevant but causally influential sentence into a citation increases the reward), the central claim that the ablation reward tracks citation quality is falsified.","supporting_citations":[{"cited_title":"Contextcite: Attributing model generation to context","cited_arxiv_id":null,"evidence_quote":"Provides ContextCite, the contributive-attribution baseline and the random-ablation method that motivates SelfCite's necessity/sufficiency reward."}],"review_version":1}