{"id":"3b219bff-0ffc-4039-87f6-67fffae5cd98","arxiv_id":"2412.07289","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A supervised rationale verifier trained on contrastively collected biased and unbiased rationales improves few-shot relation extraction by retrieving corrective demonstrations for the LLM.","lead":"This paper introduces a small rationale-supervisor model that checks whether a large language model's explanation for a relation extraction answer is biased, and if so, feeds it better example demonstrations to correct itself. The method improves few-shot relation extraction accuracy across several models and datasets, offering a cheap self-correction module for information extraction.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The biased/unbiased contrast may encode intervention artifacts (golden label supplied vs wrong-label demonstrations) rather than the actual bias that occurs at test time, so the supervisor and feedback retrieval could be solving the wrong contrast.","rationale":"The reader's weakest assumption focuses on the label-guided intervention: that prompting with the golden label yields a genuinely unbiased rationale. I agree this is fragile, but the more serious gap is symmetric: the biased rationales are also collected under an artificial intervention (wrong-label demonstrations). The entire contrastive training set is generated under two do-operations that are unlike the natural inference-time condition, so the supervisor could be learning to separate 'label was shown' from 'a contradictory label was shown' rather than 'reasoning is biased' from 'reasoning is unbiased'. This is load-bearing because both the verification and the feedback retrieval depend on the supervisor's representation space generalizing to natural test-time rationales. The paper provides no direct evidence of this generalization: no analysis of the supervisor's own accuracy, no distributional comparison between DI rationales and naturally biased rationales, and no analysis of whether retrieved feedback demonstrations are actually more informative than random demonstrations with the same predicted label. The end-task F1 gains and ablations are consistent with the method working, but they are also consistent with a shallower mechanism, e.g., the supervisor detecting a style difference and the feedback demonstrations simply providing a new ICL context that nudges the model toward a different label. The proposed concrete test would settle this by measuring whether the supervisor and retriever perform their intended function on natural biased rationales. I still view the paper as a solid engineering contribution with extensive experiments, which is why the verdict remains conditional rather than reject, but the central causal claim is not yet established. The reader and I overlap on the LGI concern, but my concern extends to the DI procedure and to the lack of a direct test of the supervision signal, hence 'partial' agreement.","tokens_in":24972,"tokens_out":5513,"duration_ms":65946,"concrete_test":"On a held-out subset (e.g., SemEval 5-shot with the random backbone, greedy decoding), record all initial predictions and rationales. For every sample whose initial prediction is wrong, treat the rationale as a naturally biased rationale with known ground truth. Then: (1) compute the supervisor's verification accuracy on these natural biased rationales using Eq. 5; (2) for samples the supervisor flags as biased, measure how often the top-1 feedback demonstration retrieved by Eq. 7 has the same golden relation as the test sample. Compare both metrics with the same metrics computed on the artificially collected DI rationales in the training set.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that the rationale supervisor detects and corrects the LLM's relation bias. The training signal for this supervisor, however, is constructed by two interventions that may not reproduce the test-time phenomenon. In 'Induce Unbiased Rationale', the golden label is explicitly supplied in the prompt (do(Y = yi)), and in 'Observe Biased Rationale', the LLM is deliberately pushed toward a wrong label by providing demonstrations whose labels differ from yi (do(I = {xj, rj, yj})). The contrastive loss in Eq. 1 therefore separates rationales generated under these two artificial conditions. At inference, the initial rationale is produced by standard ICL with correct demonstrations but without the golden label, so it lies outside both training conditions. If the supervisor's representation space is dominated by surface cues such as 'the correct label was stated in the prompt' versus 'a conflicting label appeared in the demonstrations', then the verification score in Eq. 5 and the feedback retrieval in Eqs. 6-7 will not generalize to naturally occurring biased rationales. The paper never measures the supervisor's accuracy on real test-time biased rationales, nor whether retrieved feedback demonstrations actually share the true relation with the test example. The appendix quality analysis of unbiased rationales scores plausibility, not absence of bias, and the consistency check do(R) -> Y only establishes sufficiency of the rationale for the golden label, not that the rationale is free of the stereotype bias B that the SCM allows to influence R even under do(Y). If the contrast is an artifact of the collection protocol, the ablation gains in Table 2 (w/o DI, w/o LGI) would reflect sensitivity to the intervention procedure rather than evidence that the method corrects genuine relation bias.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SRVF, an automated feedback framework for relation extraction (RE) with large language models. The method trains a BERT-based 'rationale supervisor' by contrastive learning on rationales collected via two prompt interventions: 'unbiased' rationales are induced by supplying the golden relation label to the LLM, and 'biased' rationales are obtained by providing in-context demonstrations whose labels differ from the golden label. At inference, the supervisor scores the LLM's generated rationale; if it is judged biased, the supervisor retrieves feedback demonstrations from the labeled set and the LLM regenerates the prediction, iterating until the rationale is deemed unbiased. Experiments cover three sentence-level RE datasets (SemEval, TACRED, Re-TACRED) with three demonstration-selection backbones and four few-shot settings, five LLM backbones, two document-level datasets, and one event-detection dataset, with ablations and sensitivity analyses.","tokens_in":25226,"tokens_out":7143,"duration_ms":75297,"significance":"If the reported gains are reliable, the paper fills a genuine gap: bringing automated verification-and-feedback correction to relation extraction. The design is thoughtful—using the supervisor both to detect bias and to retrieve corrective demonstrations is a novel combination, and the paper includes a broad experimental matrix, ablations, and sensitivity studies. The code is released. The main empirical claim is large and consistent across settings (e.g., up to 10.65 absolute micro-F1 improvement over random ICL). However, the absence of statistical validation and the lack of direct evidence that the supervisor's bias detection generalizes to test-time rationales are significant caveats.","major_comments":[{"comment":"The contrastive training separates rationales produced under two artificial interventions: with the golden label explicitly provided (do(Y=yi)) and with demonstrations whose labels differ from yi (do(I=Ddii)). At inference, the initial rationale is generated by standard ICL without either intervention, so it lies outside both training conditions. The paper never measures whether the supervisor's decision pb>0 actually matches ground-truth bias on real test-time rationales, nor whether the retrieved feedback demonstrations share the true relation with the test example. The end-task F1 gains are consistent with a supervisor that acts as a noisy gate, and the 'w/o FDR' ablation does not isolate verification accuracy. Please add a direct evaluation of the supervisor on held-out test rationales (e.g., rationales from the ICL baseline on the test set), report detection precision/recall, and show that correction success is concentrated in examples flagged as biased rather than in a random subset.","section":"Rationale Verification and Feedback, Eq. (5)"},{"comment":"All reported micro-F1 numbers are point estimates; no standard deviations, confidence intervals, or significance tests are given. The abstract claims that the method 'significantly outperforms' existing methods, but without repeated runs (e.g., different random demonstration samples for the Random backbone, or different seeds) it is impossible to assess whether the gains over Self-Consistency and GRACE are statistically reliable. Please provide mean±std over multiple seeds for at least the main comparisons and run a paired significance test (e.g., paired bootstrap).","section":"Table 1 and Experiments"},{"comment":"The label-guided procedure guarantees only that the generated rationale is sufficient for the golden label under the LLM (the do(R)->Y consistency check). It does not ensure that the rationale is free of the bias that occurs at test time; a rationale could simply restate or be conditioned on the supplied label and still pass step 2. The quality scores in Table 12 measure plausibility, not absence of bias. This matters because the 'unbiased' rationales are used as positive anchors in Eq. (4)–(5). Please provide a bias-specific analysis, for example by comparing induced 'unbiased' rationales with actual test-time correct rationales, or by showing that the supervisor's Su anchors are not identifiable solely through surface cues such as the presence of the golden label in the prompt.","section":"Induce Unbiased Rationale and Appendix Quality Analysis"},{"comment":"The paper states that the temperature τ is set to 0.2 for all experiments, but the sensitivity analysis in Table 13 shows that τ=0.50 yields a higher average micro-F1 (41.49 vs. 40.68). This inconsistency needs an explanation: if the main results were obtained with a suboptimal τ, the hyperparameter selection procedure should be reported (e.g., chosen on a development set rather than on the test set). Otherwise, please confirm that the main conclusions are unchanged when τ=0.50 is used.","section":"Impact of the Hyper-parameter τ (Appendix)"}],"minor_comments":[{"comment":"The section header 'Evaluation Protocal' should be 'Evaluation Protocol'.","section":"Experiments (typo)"},{"comment":"The SCM diagrams are used to motivate do(Y=yi) and do(I=...), but these operations are implemented as prompt modifications, not true causal interventions that cut all back-door paths (e.g., B can still influence R through X→B→R). The causal language should be tempered to avoid overclaiming.","section":"Figure 2 and Method"},{"comment":"The paper uses both 'Reasoning Explanations' (e.g., Table 6) and 'rationale' (in the method text) to refer to the same output. Please unify the terminology.","section":"Prompt examples"},{"comment":"The description of the GRACE baseline adaptation is terse. Please specify what the discriminator's positive and negative training samples are in the RE setting and how the 'best rationale' is selected during inference, so that the comparison is reproducible.","section":"Appendix, GRACE adaptation"},{"comment":"The error matrices show only the six non-Other relation types; please clarify whether predictions of 'Other' are omitted or merged, and whether the row sums are consistent with the total number of test examples.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical contribution with broad experiments and a clear application. The main risk is not the end-task results per se, but the lack of validation that the rationale supervisor actually detects the intended phenomenon (relation bias) rather than surface cues from the intervention prompts. If the authors can provide direct evidence of supervisor generalizability and add statistical rigor, the paper would be much stronger. The hyperparameter inconsistency (τ=0.2 vs. τ=0.5) is also worth watching, as it may indicate test-set peeking in the sensitivity analysis. Overall, I would support revision rather than rejection because the central idea is defensible and the empirical findings are extensive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it delivers a practical plug-in that improves few-shot relation extraction across a wide range of settings: three sentence-level datasets, shot counts from 5 to 50, three demonstration-selection backbones, five LLMs, plus document-level RE and event detection. The absolute F1 gains are large, especially with random demonstrations (about 10.7 points on average) and still meaningful with task-specific retrieval (about 3.2). Second, the mechanism story is the weakest part of the paper. The rationale supervisor is contrastively trained on \"unbiased\" rationales produced by giving the LLM the golden label, and \"biased\" rationales produced by pushing it with wrong-label demonstrations. At test time the initial rationale is generated with neither intervention. The paper never measures whether the supervisor's verdict (biased vs unbiased) actually matches the true correctness of the prediction, nor whether the retrieved feedback demonstrations share the true relation with the test example. The consistency check for unbiased rationales (do(R) -> Y) tests sufficiency, not absence of stereotype bias. So the contrast may be encoding intervention artifacts—label supplied vs not, conflicting demo vs not—rather than the relation bias they claim.\n\nDo not overcorrect: the empirical result is robust and internally consistent. The ablations show each component matters, the method beats the adapted baselines by wide margins, and the sensitivity analyses show the method isn't finicky about prompts or hyperparameters. The novelty is real: nobody has built an automated feedback loop for RE that re-selects demonstrations as correction, and the contrastive rationale collection strategy is new. The code is public.\n\nThe weaknesses are mostly about interpretation, not execution. I'd like to see error bars or significance tests, a direct analysis of the supervisor's verification accuracy on test-time rationales, and an analysis of what the retrieved demonstrations have in common with the test samples. Those are addressable in revision. The paper deserves a serious referee, not a desk reject; the idea is timely and the evidence is broad.\n\nFor you: it's worth knowing about this paper if you track self-correction or information extraction. I'd bring it to reading group for a discussion of intervention design.","headline":"The empirical claims hold up, but the paper's central bias-detection mechanism is under-validated: the supervision signal is collected under different conditions than the test-time predictions, and the authors never show the supervisor is actually separating unbiased from genuinely biased rationales at inference.","tokens_in":25829,"tokens_out":3751,"would_cite":true,"duration_ms":39952,"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 small rationale supervisor can catch the biased relation predictions of large language models and correct them with re-selected demonstrations, adding up to 10.65 points of micro-F1 over standard in-context learning.","keywords":["relation extraction","large language models","rationale verification","automated feedback","in-context learning","causal intervention","contrastive learning","few-shot learning"],"falsifier":"A decisive test would be to hold out a set of samples that each have both a label-guided 'unbiased' rationale and a demonstration-guided 'biased' rationale for the same sentence, then measure whether the trained supervisor ranks the biased rationale above the unbiased one with above-chance accuracy; if it does not, the collected rationales are not encoding relation bias in the intended way.","tokens_in":24767,"feed_emoji":"🔗","tokens_out":17886,"duration_ms":160237,"temperature":0.7,"pith_summary":"Relation extraction is the task of deciding which semantic relation holds between two entities mentioned in a sentence, and large language models often get this wrong because their pre-trained knowledge nudges them toward stereotyped relations. The paper claims that this bias shows up in the model's rationale, the explanation it gives for its prediction, and that the bias can be detected and corrected without retraining the large model. It builds a small rationale supervisor that is contrastively trained on two kinds of rationales: unbiased ones, produced by asking the LLM to explain the golden relation label, and biased ones, produced by feeding demonstrations with deliberately different labels. At inference the supervisor flags rationales that resemble known biased ones and retrieves new in-context demonstrations from the labeled set as feedback, repeating until the rationale looks unbiased. The paper reports average micro-F1 gains of 10.65 points over random-demonstration in-context learning, 6.49 points over semantic-similarity retrieval, and 3.24 points over task-specific retrieval, with the same loop transferring to document-level relation extraction and to event detection.","feed_headline":"Rationale verifier lifts LLM relation extraction by 10.65 points","feed_subtitle":"A small verifier catches biased rationales and feeds corrective demonstrations back, lifting few-shot relation F1.","key_machinery":"The load-bearing object is the rationale supervisor, a small pre-trained encoder trained with a supervised contrastive loss over rationale pairs. The training data come from a causal intervention-and-observation step: label-guided intervention blocks the model's bias path and asks it to explain the golden label, producing the unbiased rationale pool, while diversified intervention feeds demonstrations with different labels to elicit a spread of biased rationales. At inference the supervisor computes $p_b = \\max_{r^b \\in S_b} \\operatorname{sim}(r, r^b) - \\max_{r^u \\in S_u} \\operatorname{sim}(r, r^u)$, where $S_b$ and $S_u$ are the stored biased and unbiased rationales for the predicted relation; $p_b > 0$ means the current rationale is closer to the biased region. Feedback is then generated by taking the $k$ labeled samples whose biased rationales are most similar to the current rationale and using their unbiased versions as new in-context demonstrations, so the correction signal is a demonstration set rather than a textual error message.","core_discovery":"On the paper's own terms, the central claim is that the relation bias of an LLM is visible in the rationale it produces and can be exploited by a lightweight supervised model. The paper induces a label-guided unbiased rationale for each labeled sample by prompting the LLM to explain why the golden relation holds, and it elicits deliberately wrong rationales by feeding demonstrations with different labels, so that both correct explanations and typical error directions are collected. A contrastive loss pulls together unbiased rationales with the same golden label and biased rationales under the same error direction, while pushing apart biased and unbiased rationales for the same sample. At inference, the supervisor scores a new rationale by the difference between its maximum similarity to the stored biased pool and its maximum similarity to the stored unbiased pool for the predicted relation; a positive difference flags the prediction as biased and triggers retrieval of feedback demonstrations from labeled samples in the most similar biased cluster. The LLM is then prompted again with those demonstrations, and the loop repeats until the rationale passes verification, which the paper reports converges in about four feedback rounds.","pith_inferences":["An implicit consequence is that the supervisor's representation could serve as a general bias probe, flagging ambiguous or low-confidence predictions for human review even when corrective demonstrations are unavailable.","A clean way to isolate the mechanism would be to strip the rationale out of the feedback prompt and keep only the retrieved demonstrations; if the F1 gain persists, the correction is coming from demonstration selection rather than from rationale verification.","The supervised signal is generated by the very LLM it corrects, so the quality of the 'unbiased' rationales is bounded by that model's ability to justify a supplied label; a weaker generator should narrow the representation gap between the two pools, and the paper does not directly measure this dependence."],"forward_implications":["Because only samples flagged as biased are regenerated, the correction loop is selective; the paper reports that SRVF is the second-most efficient method in its comparison while giving the largest F1 gains.","The framework acts as a prompt-level plug-in rather than a fine-tune, and the paper reports gains on Llama-2-chat, Meta-Llama-3-Instruct, and GPT-3.5-turbo, so the correction mechanism is not tied to one model family.","The same verification-feedback procedure improves document-level relation extraction on DocRED and Re-DocRED and transfers to event detection on ACE05, indicating the supervision is not limited to sentence-level relation labels.","Convergence is fast: one correction round yields up to 11 points of absolute micro-F1, and performance saturates after about four rounds, so the number of extra LLM calls is bounded in practice.","Ablation results show that removing the label-guided intervention, the diversified intervention, the contrastive training, the feedback retrieval, or the regeneration step each lowers performance, so every component of the loop contributes to the reported gains."],"supporting_citations":[{"why":"Defines the Self-Refine automated feedback baseline that the paper adapts to relation extraction and outperforms.","marker":"Madaan et al. 2023"},{"why":"Provides the GRACE verifier-guided baseline, representative of verification without demonstration-based feedback.","marker":"Khalifa et al. 2023"},{"why":"Provides the Self-Consistency majority-voting verification baseline used for comparison.","marker":"Wang et al. 2023b"},{"why":"Supplies the task-specific demonstration-retrieval backbone that SRVF is applied to and one of its strongest comparison settings.","marker":"Wan et al. 2023"},{"why":"Supplies the sentence-embedding retriever used as the semantic-similarity demonstration-selection backbone.","marker":"Gao, Yao, and Chen 2021"},{"why":"Supplies the pre-trained language model used to initialize the rationale supervisor.","marker":"Devlin et al. 2018"},{"why":"Provides the Llama-2-chat models that carry the main relation-extraction experiments.","marker":"Touvron et al. 2023"},{"why":"Provides the SemEval benchmark used for the main relation-extraction evaluation.","marker":"Hendrickx et al. 2010"},{"why":"Provides the TACRED benchmark used in the few-shot relation-extraction evaluation.","marker":"Zhang et al. 2017"},{"why":"Provides the Re-TACRED benchmark used in the few-shot relation-extraction evaluation.","marker":"Stoica, Platanios, and Poczos 2021"}],"fun_headline_variants":["Rationale verifier feedback improves LLM relation extraction","Contrastive rationale supervisor enhances LLM relation extraction","Iterative rationale check and feedback refine LLM relation extraction","Supervised rationale verification corrects LLM relation extraction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes that asking the LLM to explain the golden relation label produces a genuinely unbiased rationale about the evidence, rather than a post-hoc rationalization of the supplied label; if that assumption fails, the supervisor learns label-conditioned wording instead of true relation bias, and the feedback loop no longer corrects the underlying error.","fun_headline_variants_meta":{"raw":{"variants":["Rationale verifier feedback improves LLM relation extraction","Contrastive rationale supervisor enhances LLM relation extraction","Iterative rationale check and feedback refine LLM relation extraction","Supervised rationale verification corrects LLM relation extraction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00036,"raw_usage":{"total_tokens":1913,"prompt_tokens":877,"completion_tokens":1036,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":972}},"tokens_in":493,"tokens_out":1036,"duration_ms":11050,"temperature":1.0,"reasoning_tokens":972,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:55:34.150794+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test would be to hold out a set of samples that each have both a label-guided 'unbiased' rationale and a demonstration-guided 'biased' rationale for the same sentence, then measure whether the trained supervisor ranks the biased rationale above the unbiased one with above-chance accuracy; if it does not, the collected rationales are not encoding relation bias in the intended way.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the GRACE verifier-guided baseline, representative of verification without demonstration-based feedback."},{"cited_title":"N.; Kozareva, Z.; Nakov, P.; \\'O S \\'e aghdha, D.; Pad \\'o , S.; Pennacchiotti, M.; Romano, L.; and Szpakowicz, S","cited_arxiv_id":null,"evidence_quote":"Provides the SemEval benchmark used for the main relation-extraction evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the TACRED benchmark used in the few-shot relation-extraction evaluation."},{"cited_title":"A.; and P \\'o czos, B","cited_arxiv_id":null,"evidence_quote":"Provides the Re-TACRED benchmark used in the few-shot relation-extraction evaluation."}],"review_version":1}