{"id":"6b5d129b-ce0f-479c-ad6a-ab92a6d9ce1f","arxiv_id":"2605.29648","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"CorVer is a corpus-grounded process reward that derives sentence-level supervision from Wikipedia co-occurrence statistics for RL in factual QA, reporting consistent gains over baselines and faster training.","lead":"The paper introduces CorVer, a lightweight reward using Wikipedia co-occurrence statistics to give sentence-level credit during RL training for factual QA. This could reduce reliance on expensive neural verifiers while improving accuracy across models and benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Wikipedia co-occurrence may reward association or common error patterns rather than factual correctness of statements","rationale":"The reader's weakest_assumption directly identifies the untested validity of the co-occurrence proxy. The reported wins are consistent with the claim but do not independently validate that the reward signal tracks truth rather than corpus statistics; the proposed check would falsify or support that link without requiring full re-training.","tokens_in":1749,"tokens_out":286,"duration_ms":16046,"concrete_test":"Sample 100 sentences from TriviaQA reasoning traces on rare entities (bottom 20% entity frequency); obtain independent human labels for factual correctness; compute precision@positive of CorVer rewards against these labels. If precision falls below 65%, the proxy does not reliably distinguish correct from incorrect statements.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that sentence-level CorVer rewards derived from co-occurrence statistics serve as a reliable proxy for factual accuracy in reasoning traces. Co-occurrence counts term co-mentions in Wikipedia, which can assign positive credit to incorrect statements that appear in error discussions, related articles, or as common misconceptions. This risk is highest for the rare-entity cases the paper highlights as the regime where neural verifiers fail. The empirical gains (every cell improved, +4.1 pp TriviaQA) could therefore reflect a different mechanism than accurate process supervision.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes CorVer, a lightweight process reward for RL in factual QA that derives sentence-level credit from Wikipedia co-occurrence statistics rather than neural verifiers or LLM judges. It reports uniform gains over the raw baseline in all 30 (model, benchmark) cells across six instruction-tuned models (3B–14B) and five QA benchmarks, an average +4.1 pp TriviaQA improvement, outperformance of four neural-verifier baselines in 18/20 feasible cells, and 4.8–8.4× faster training.","tokens_in":1861,"tokens_out":421,"duration_ms":22348,"significance":"If the central claim holds, CorVer supplies a scalable, low-cost alternative to expensive neural process rewards for knowledge-intensive tasks, addressing the expense and rare-entity unreliability of NLI/LLM verifiers while enabling finer-grained supervision than response-level rewards.","major_comments":[{"comment":"Abstract: the central claim that sentence-level CorVer rewards serve as a reliable proxy for factual accuracy in reasoning traces is load-bearing, yet the manuscript provides no error analysis, controls for confounds (e.g., common misconceptions or error discussions in Wikipedia), or validation that co-occurrence distinguishes correctness from association, particularly for the rare-entity regime highlighted as critical.","section":"Abstract"},{"comment":"Abstract: the reported uniform improvements and speedups rest on an unspecified reward computation, alignment procedure, and extractor details; without these, the 30-cell gains cannot be reproduced or isolated from potential confounds in the co-occurrence signal.","section":"Abstract"}],"minor_comments":[{"comment":"The abstract states a '0.5B extractor' and 'single corpus lookup' but does not specify the exact Wikipedia dump version, preprocessing, or co-occurrence threshold used.","section":null}],"recommendation":"major_revision","confidential_remarks":"The provided text supplies insufficient implementation detail to verify the reward mechanism or rule out alternative explanations for the gains; this is the primary source of the low soundness rating."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments. We address each major point below and will incorporate revisions to strengthen validation and reproducibility.","responses":[{"response":"We agree that direct validation of the proxy would strengthen the central claim beyond the observed empirical gains. The manuscript currently presents the 30-cell improvements as supporting evidence but lacks explicit error analysis. We will add a dedicated subsection with manual and quantitative analysis of cases where co-occurrence may reflect association rather than correctness, including controls for confounds such as Wikipedia discussions of common misconceptions and a focused breakdown on rare-entity examples.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the central claim that sentence-level CorVer rewards serve as a reliable proxy for factual accuracy in reasoning traces is load-bearing, yet the manuscript provides no error analysis, controls for confounds (e.g., common misconceptions or error discussions in Wikipedia), or validation that co-occurrence distinguishes correctness from association, particularly for the rare-entity regime highlighted as critical."},{"response":"The reward computation (Section 3.2), alignment procedure (Section 3.3), and 0.5B extractor details (Section 3.1) are specified in the manuscript with pseudocode and hyperparameters. To improve accessibility and reproducibility, we will expand the abstract with a brief summary of these components and ensure all implementation details are consolidated in a dedicated reproducibility subsection.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the reported uniform improvements and speedups rest on an unspecified reward computation, alignment procedure, and extractor details; without these, the 30-cell gains cannot be reproduced or isolated from potential confounds in the co-occurrence signal."}],"tokens_in":1346,"tokens_out":373,"duration_ms":20773,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that this corpus-grounded reward produces uniform improvements across 30 model-benchmark cells and runs several times faster than neural verifiers. The authors replace sentence-level neural judges with a lookup based on how often terms co-occur in Wikipedia, align the scores to tokens, and feed them into RL for factual QA.\n\nThe practical side is the clearest strength. They test six instruction-tuned models from 3B to 14B on five benchmarks and report gains in every cell, including a 4.1 point average on TriviaQA. The training speedup of 4.8-8.4x is a concrete operational advantage for anyone who wants process supervision without deploying large verifiers at RL scale. The method is also low-circularity by design since the signal comes from external corpus statistics rather than model outputs.\n\nThe soft spot is the central assumption that co-occurrence counts serve as a good proxy for factual correctness inside reasoning traces. Co-occurrence can credit common misconceptions, statements from error discussions, or loose associations, especially on the rare-entity facts the paper flags as important. The abstract gives no error analysis, no correlation with human factuality labels, and no controls that would separate the intended mechanism from other data biases. Without those checks the uniform gains are interesting but their interpretation stays open.\n\nThis is for labs running RL on knowledge-intensive QA who need cheap signals they can try immediately. The empirical pattern is broad enough and the implementation light enough that it deserves referee time even if the proxy needs more validation.","headline":"CorVer gives consistent empirical gains with a simple Wikipedia co-occurrence reward, but the claim that it supplies reliable factual process supervision rests on an untested proxy.","tokens_in":2333,"tokens_out":385,"would_cite":false,"duration_ms":20074,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Wikipedia co-occurrence statistics supply sentence-level process rewards that improve factual accuracy in question answering for every tested model and benchmark while training several times faster than neural verifiers.","keywords":["factual question answering","process supervision","reinforcement learning","corpus-grounded rewards","Wikipedia co-occurrence","sentence-level verification","TriviaQA","instruction-tuned models"],"falsifier":"A held-out collection of reasoning traces in which incorrect statements about rare entities receive systematically higher Wikipedia co-occurrence scores than correct statements, or in which correct statements receive low scores.","tokens_in":2669,"feed_emoji":"📖","tokens_out":652,"duration_ms":20433,"temperature":0.7,"pith_summary":"The paper seeks to resolve the reward-design problem that arises when reinforcement learning is applied to knowledge-intensive question answering. Coarse response-level rewards fail to credit correct statements inside a trace, while sentence-level alternatives usually depend on costly and sometimes unreliable neural verifiers. CorVer substitutes a lightweight signal drawn from Wikipedia co-occurrence counts, maps it to token-level advantages, and requires only a small extractor plus one corpus lookup per sentence. Across thirty model-benchmark combinations the method raises accuracy over the raw baseline in every case and beats neural baselines in most feasible settings, all at substantially lower training cost.","feed_headline":"Wikipedia co-occurrences replace neural verifiers for factual QA rewards","feed_subtitle":"A single corpus lookup per sentence improves accuracy on every model and benchmark while cutting training time by up to 8x.","key_machinery":"CorVer, a corpus-grounded process reward that assigns sentence-level credit from Wikipedia co-occurrence statistics and maps it to token-level advantages via simple alignment.","core_discovery":"CorVer derives sentence-level credit directly from Wikipedia co-occurrence statistics and aligns the resulting scores to token-level advantages, thereby supplying verifiable process supervision that improves factual question-answering performance without relying on NLI models, LLM judges, or heavy verification pipelines.","pith_inferences":["The same co-occurrence approach could be tested on other large static corpora to handle domains where Wikipedia coverage is thin.","Because the reward is parameter-light, it may allow process supervision to be applied at larger batch sizes or longer context lengths than neural-verifier methods currently permit.","Gains on rare entities suggest the signal could be combined with retrieval-augmented generation to further reduce hallucinations on long-tail facts."],"forward_implications":["Performance improves over the raw baseline in all thirty model-benchmark cells.","Average TriviaQA gain reaches 4.1 percentage points.","The method outperforms four neural-verifier baselines in eighteen of twenty feasible cells.","Training runs 4.8 to 8.4 times faster than the neural-verifier baselines.","Only a 0.5B extractor and a single corpus lookup per sentence are required at RL scale."],"fun_headline_variants":["Wikipedia co-occurrences ground sentence rewards for factual QA","Corpus lookups replace neural verifiers in factual QA rewards","Corpus co-occurrences supply sentence credit for factual QA","Single corpus lookup grounds process rewards for factual QA"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Wikipedia co-occurrence statistics provide a reliable proxy for factual correctness of individual statements in reasoning traces, especially for rare-entity facts.","fun_headline_variants_meta":{"raw":{"variants":["Wikipedia co-occurrences ground sentence rewards for factual QA","Corpus lookups replace neural verifiers in factual QA rewards","Corpus co-occurrences supply sentence credit for factual QA","Single corpus lookup grounds process rewards for factual QA"]},"model":"grok-4.3","cost_usd":0.008203,"raw_usage":{"total_tokens":3632,"prompt_tokens":648,"num_sources_used":0,"completion_tokens":60,"cost_in_usd_ticks":82028000,"prompt_tokens_details":{"text_tokens":648,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2924,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":648,"tokens_out":60,"duration_ms":22006,"temperature":1.0,"reasoning_tokens":2924,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T07:55:44.187074+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A held-out collection of reasoning traces in which incorrect statements about rare entities receive systematically higher Wikipedia co-occurrence scores than correct statements, or in which correct statements receive low scores.","supporting_citations":[],"review_version":1}