{"id":"0f6efc83-1903-4085-b824-581d7f261340","arxiv_id":"2606.00020","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 4B model trained with balanced CPT, CoT-SFT, and efficiency-aware GRPO reaches 50.99 F0.5 on NACGEC and 59.61 F1 on CSCD, beating larger models and GPT-4.","lead":"CSRP trains a 4B LLM for Chinese grammar and spelling correction in three stages: domain pre-training, reasoned fine-tuning, and a reward that punishes unnecessary edits. It posts new high scores on standard Chinese correction tests while cutting the over-correction that plagues ordinary fine-tuning.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"SOTA precision/+8.44 P and claimed RL–CPT orthogonality rest on an unvalidated hand-crafted EAR whose Levenshtein thresholds risk gaming ChERRANT rather than robustly enforcing minimal intervention.","rationale":"The reader correctly isolates the hand-designed EAR (and its fixed thresholds) as the single weakest assumption underwriting both the SOTA numbers and the orthogonality narrative. The progressive ablations, public code, and CoT faithfulness study are genuine strengths and rule out gross data leakage or circular derivation; the only remaining fragility is whether the particular reward shape generalizes beyond the 67 k in-domain prompts and the Levenshtein-based scorer. Because that is precisely the concern already flagged, no verdict adjustment is required. A sensitivity re-run of the kind proposed above would be sufficient to convert CONDITIONAL into a higher-confidence ACCEPT.","tokens_in":19114,"tokens_out":613,"duration_ms":30156,"concrete_test":"From the released CPT+SFT checkpoint, re-run GRPO with two ablated rewards: (i) RI-only (drop η entirely) and (ii) the original form with the 0.99 threshold shifted ±0.1 and the perfect-match bonus reduced from +10 to +5. If either ablation erases more than half of the reported +5.54 F0.5 / +8.44 P on NACGEC, the specific EAR design is load-bearing for the headline claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim (NACGEC 50.99 F0.5 / 57.17 P; GRPO+EAR supplies an orthogonal ~8 % relative F0.5 lift and +8.44 precision with near-zero recall cost) treats the Efficiency-Aware Reward of §3.3 (Eqs. 3–5) as a faithful operationalization of “minimal intervention.” REAR is a piecewise function of two Levenshtein ratios (RI, η) with hard cut-offs (RI>0.99 → +10, RI>0 → 2+5·RI·η, RI≤0 → –3) plus an explicit special case (+2 identity / –2 any edit when d(S,G)=0). Because the primary evaluation metric (ChERRANT F0.5) is itself character-level edit-distance based, the observed precision surge can be produced by metric-aligned gaming rather than a general reduction in over-correction. Tables 4–5 demonstrate that the numerical gain appears and is additive to CPT, yet they never ablate the functional form, the numerical thresholds, or the special-case rule, nor do they supply human false-positive rates on already-correct text. Consequently the causal attribution “explicit optimization for edit efficiency is essential” remains under-supported.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes CSRP, a three-stage pipeline for Chinese grammatical error correction and spelling check: (i) balanced continual pre-training on 5.9M samples (8:2 general-to-correction mix), (ii) rationale-augmented SFT with teacher-distilled CoT of the form Localization→Classification→Rationale, and (iii) GRPO policy alignment under a hand-designed Efficiency-Aware Reward (EAR) built from Levenshtein relative improvement (RI) and edit-efficiency ratio η (Eqs. 3–5), with special-case rewards for already-correct inputs. On NACGEC the 4B model reports 50.99 F0.5 and 57.17 precision (SOTA among listed baselines); on CSCD it reports 59.61 F1, +5.20 over GPT-4. Progressive ablations (Tables 4–5) attribute an ~8% relative F0.5 gain and +8.44 precision (near-zero recall cost) to the RL stage and argue that this gain is orthogonal to CPT. Code is released.","tokens_in":19528,"tokens_out":927,"duration_ms":16628,"significance":"Over-correction under MLE is a genuine, practical bottleneck for LLM-based GEC; a curriculum that jointly injects domain priors, diagnostic CoT, and an explicit minimal-edit objective is a useful systems contribution. Strengths include: public code; a controlled SFT baseline that re-uses the RL data split (Table 4, “CPT + SFT (w/ RL data)”); isolation of CPT vs. RL via the “SFT + GRPO (w/o CPT)” row; a 1,000-sample double-blind CoT faithfulness study (95.2%, κ=0.81); and consistent precision–recall dynamics on two benchmarks. If the EAR truly operationalizes minimal intervention rather than gaming ChERRANT, the claimed orthogonality of knowledge internalization and edit-policy calibration would be a clear takeaway for the field.","major_comments":[{"comment":"§3.3, Eqs. (3)–(5): The central causal claim—“explicit optimization for edit efficiency is essential”—rests on EAR, a piecewise function of Levenshtein RI/η with hard thresholds (RI>0.99→+10; RI>0→2+5·RI·η; RI≤0→−3) plus the binary +2/−2 rule when d(S,G)=0. Because the primary scorer (ChERRANT F0.5) is itself character-level edit-distance based, the +8.44 precision jump can be produced by metric-aligned gaming rather than a general reduction in over-correction. Tables 4–5 show that some reward of this family yields additive gains, but they never ablate functional form, numerical coefficients, or the special-case rule, nor compare against simpler rewards (e.g., pure F0.5 proxy, binary identity penalty only, or length-normalized edit cost). Without such controls the attribution to “edit efficiency” remains under-supported.","section":null},{"comment":"§4.2–4.3 / Table 5: Precision gains are reported only via automatic ChERRANT/CSC metrics. There is no human false-positive audit on already-correct native text (or on a held-out “no-error” subset) that would confirm the model truly preserves well-formed prose rather than suppressing edits that ChERRANT happens to penalize. A modest human FP/FN study on, e.g., 200–500 correct and incorrect sentences before vs. after GRPO would substantially strengthen the over-correction narrative and is load-bearing for the “minimal intervention” interpretation.","section":null},{"comment":"§4.2, Tables 2–5: All headline numbers are single-run point estimates with no seeds, confidence intervals, or significance tests. Given that the claimed SOTA margin over CEC3 is +2.25 F0.5 and the RL-only lift is +3.78 over the data-matched SFT baseline, variance estimates are needed to decide whether the ranking and the orthogonality claim are stable. At minimum, report 2–3 seeds for the SFT→GRPO transition on NACGEC.","section":null}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: a 4B model trained CPT → CoT-SFT → GRPO with a Levenshtein-based Efficiency-Aware Reward hits 50.99 F0.5 / 57.17 P on NACGEC and 59.61 F1 on CSCD, beating the previous 4B SOTA and a 14B system, with the RL stage adding roughly +5.5 F0.5 and +8.4 precision at almost no recall cost. The controlled “SFT+GRPO without CPT” run shows that gain is largely additive to the CPT lift, which is the cleanest part of the paper.\n\nWhat is actually new is not the three-stage recipe itself (CPT + distilled CoT + GRPO are all known) but the specific EAR (relative improvement and edit-efficiency ratio with the identity-preserving special case for already-correct inputs) plus the explicit orthogonality ablation. Tables 4–5 and the progressive configurations do the work they claim: data quantity is controlled, CoT helps, and RL mainly moves precision. The 1 000-rationale human check (95.2 % faithful, κ=0.81) and the public code link are real positives. Limitations section is honest about teacher dependence and GRPO cost.\n\nSoft spots, in proportion. The stress-test note is right that EAR is hand-designed and never ablated on its thresholds or functional form; because ChERRANT is also edit-distance based, some of the precision surge could be metric-aligned rather than pure “minimal intervention.” They also never report human false-positive rates on clean text. That is a genuine gap in causal attribution, not a fatal flaw—the numerical gains and the additive pattern still stand. Free parameters (mixture ratio, reward cut-offs, N, β) are listed but not swept. Citation pattern is normal for the subfield; no circularity on the held-out benchmarks.\n\nThis is for people building Chinese GEC/CSC systems or anyone who cares about precision-oriented RL for editing. It deserves a serious referee. I would cite the numbers and the ablation design; I would not treat the EAR as a settled general principle without the missing sensitivity checks. Send it out.","headline":"Solid three-stage Chinese GEC pipeline with a clean CPT/CoT/RL ablation and real precision lift; the hand-crafted EAR is the soft spot but does not sink the result.","tokens_in":20067,"tokens_out":554,"would_cite":true,"duration_ms":6240,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A three-stage training pipeline teaches a 4B Chinese correction model to edit only when the gain is worth the change.","keywords":["Chinese grammatical error correction","continual pre-training","chain-of-thought","Group Relative Policy Optimization","efficiency-aware reward","over-correction","NACGEC","CSCD"],"falsifier":"Retrain or re-evaluate with the same GRPO stage but replace EAR by a pure accuracy or F0.5 reward (or alter the fixed thresholds); if precision no longer rises ~8 points while recall stays flat, the claim that edit-efficiency optimization is essential collapses.","tokens_in":20048,"feed_emoji":"✏️","tokens_out":754,"duration_ms":6189,"temperature":0.7,"pith_summary":"Chinese grammatical error correction with large language models tends to over-correct: maximum-likelihood training pushes fluent rewrites even when the input is already fine, which hurts precision-focused scores. This paper claims the fix is a staged curriculum. First the model absorbs Chinese linguistic priors from 5.9 million balanced samples; then it learns to diagnose errors with explicit chain-of-thought rationales; finally a reinforcement stage optimizes a hand-designed Efficiency-Aware Reward that scores relative improvement against edit distance and directly penalizes unnecessary changes. On the NACGEC native-speaker benchmark the resulting 4B model reaches 50.99 F0.5 and 57.17 precision, beating larger and prior specialized systems while lifting precision roughly eight points with almost no recall loss. The same pipeline also raises Chinese spelling correction to 59.61 F1, five points above GPT-4. Ablations show the reinforcement gain is largely independent of the pre-training stage, so both knowledge and edit discipline are required.","feed_headline":"4B Chinese corrector hits 50.99 F0.5 by rewarding fewer edits","feed_subtitle":"Reinforcement with an edit-efficiency reward lifts precision eight points while recall barely moves.","key_machinery":"Efficiency-Aware Reward (EAR): a piecewise scalar built from relative improvement RI = (d(S,G) − d(P,G)) / (d(S,G) + ε) and edit-efficiency η = (d(S,G) − d(P,G)) / (d(S,P) + ε), with fixed thresholds (near-perfect RI → +10, positive RI scaled by η, non-improvement → −3, empty output → −2, and a binary +2/−2 rule for already-correct inputs) that is optimized inside Group Relative Policy Optimization.","core_discovery":"The paper establishes that Chinese grammatical error correction can escape the over-correction plateau of maximum-likelihood fine-tuning by adding Group Relative Policy Optimization guided by an Efficiency-Aware Reward that rewards relative distance reduction to the gold target while penalizing excess edits; this stage alone yields an approximately 8 percent relative F0.5 gain that is orthogonal to large-scale continual pre-training, producing state-of-the-art 50.99 F0.5 and 57.17 precision on NACGEC with a 4B model.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["CSRP lifts 4B Chinese GEC to 50.99 F0.5 via edit-efficiency RL","Efficiency-Aware Reward yields 57.17 precision, curbing over-edits","RL stage alone adds 8% relative F0.5 gain over SFT for CGEC","Chain-of-Thought plus GRPO hits 50.99 F0.5 on NACGEC","4B corrector reaches SOTA by penalizing excess grammatical edits"],"cache_read_input_tokens":14720,"weakest_assumption_plain":"The hand-tuned reward thresholds and Levenshtein ratios correctly capture the linguistic idea of minimal necessary editing and will keep working outside the 67 thousand reinforcement prompts without creating new systematic biases.","fun_headline_variants_meta":{"raw":{"variants":["CSRP lifts 4B Chinese GEC to 50.99 F0.5 via edit-efficiency RL","Efficiency-Aware Reward yields 57.17 precision, curbing over-edits","RL stage alone adds 8% relative F0.5 gain over SFT for CGEC","Chain-of-Thought plus GRPO hits 50.99 F0.5 on NACGEC","4B corrector reaches SOTA by penalizing excess grammatical edits"]},"model":"grok-4.5","effort":"low","cost_usd":0.003864,"raw_usage":{"total_tokens":1286,"prompt_tokens":867,"num_sources_used":0,"completion_tokens":103,"cost_in_usd_ticks":38640000,"prompt_tokens_details":{"text_tokens":867,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":316,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":867,"tokens_out":103,"duration_ms":4355,"temperature":1.0,"reasoning_tokens":316,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T21:29:38.450895+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Retrain or re-evaluate with the same GRPO stage but replace EAR by a pure accuracy or F0.5 reward (or alter the fixed thresholds); if precision no longer rises ~8 points while recall stays flat, the claim that edit-efficiency optimization is essential collapses.","supporting_citations":[],"review_version":1}