{"id":"d9a85443-25a8-4509-bb3d-05242c9bacbf","arxiv_id":"2506.10822","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ReCUT combines stepwise long-short sampling, dual DPO training, and DARE-Ties parameter interpolation to compress LLM reasoning chains by 30-50% without sacrificing accuracy on math benchmarks.","lead":"A new fine-tuning recipe, ReCUT, generates diverse reasoning traces by alternating long and short prompts, then uses preference optimization and model merging to make LLMs reason in fewer tokens. Across math benchmarks, it cuts token use by roughly 30-50% while keeping accuracy close to that of the base model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Accuracy-maintenance claim unsupported: ReCUT's Pass@1 trails Direct-DPO on both backbones and OnlyM_acc on Qwen, inside the noise floor of 30-40-problem sets, with no significance tests/CIs/seeds reported; the 30-50% length claim also fails vs CoD on Qwen (5.4x longer).","rationale":"The paper's central claim is that ReCUT reduces reasoning length by ~30-50% while maintaining or improving accuracy. The length component is broadly corroborated against Vanilla, SFT, Direct-DPO, and L1 on both backbones (roughly 42-65% token reduction), and the w/o Explore ablation shows the stepwise mechanism contributes. The accuracy component is where the argument is least secure, and it is load-bearing because the claim is explicitly about preserving accuracy, not only about compression. The evaluation sets driving the difficulty comparisons are tiny: AIME24 and AIME25 each contain 30 problems, and AMC23 contains 40, so each correct answer shifts Pass@1 by 2.5-3.3 points. Within that noise floor, ReCUT's average accuracy is numerically below Direct-DPO on both backbones (Qwen 45.7 vs 46.4; Llama 29.1 vs 30.8), below SFT on Llama (29.1 vs 31.5), and below the paper's own OnlyM_acc on Qwen (45.7 vs 46.9). Because §4-§5 report no confidence intervals, significance tests, or multiple seeds, 'maintaining or improving accuracy' is an overinterpretation of noise-level differences, several of which trend negative. The abstract also frames the length reduction as 'compared to various baselines' without excluding CoD, yet on Qwen ReCUT is 5.4x longer than CoD (1,298 vs 239 average tokens, Table 1). A further aggravator is data overlap: the 8,000 training questions are sampled from DeepScaleR-Preview, which §4 describes as containing AIME/AMC pairs, so exact evaluation problems may appear in training; any such overlap contaminates the absolute accuracy numbers used to justify the accuracy-parity claim. I weigh the reader's greedy step-selection concern (Eq. 5) as real but secondary: a bad early step degrades candidate-pool diversity and can propagate through the iterative conditioning, yet the w/o Explore ablation shows the overall stepwise design helps, and this failure mode would not by itself falsify the length result. Two smaller internal issues: Eq. (11) writes the merge as θ_acc + α·Top_x(θ_len), which is not the DARE-Ties procedure from Yu et al. 2024 (top-k task-vector merging with rescaling), and the Limitation section concedes the interpolation parameters are fixed at 0.25 on empirical grounds without analysis, against the 0.3-0.4 range it cites. Reproducibility also suffers from unspecified hyperparameters (DPO β, LoRA rank, sampling temperature, decoding strategy) and the absence of released code at the cited URL. The paper has genuine strengths: consistent large token reductions across two backbones, an ablation isolating the exploration mechanism, and honest limitation statements. These do not cure the statistical gap in the accuracy claim.","tokens_in":12325,"tokens_out":19576,"duration_ms":199733,"concrete_test":"Obtain the per-problem evaluation logs behind Tables 1-2 and the list of the 8,000 sampled training questions from the authors. (1) For Pass@1, compute paired bootstrap 95% confidence intervals for ReCUT minus Direct-DPO and ReCUT minus OnlyM_acc, aggregating per-problem correctness across all five benchmarks; if the intervals contain zero, the accuracy-maintenance claim is noise-limited and must be downgraded. (2) Compute ReCUT/CoD token ratios per benchmark and check the AIME24/AIME25/AMC23 test problems for exact overlap with the training sample; any overlap invalidates the absolute accuracy numbers on those sets, and a Qwen ratio above 1 contradicts the 30-50% framing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The accuracy-maintenance half of the central claim — that ReCUT 'maintains or even surpasses' baseline accuracy while cutting length 30-50% — is not supported by the reported evidence. In Table 1, ReCUT's average Pass@1 is numerically below Direct-DPO on both backbones (Qwen: 45.7 vs 46.4; Llama: 29.1 vs 30.8) and below SFT on Llama (29.1 vs 31.5); in Table 2, it is below OnlyM_acc on Qwen (45.7 vs 46.9). AIME24 and AIME25 contain only 30 problems each and AMC23 only 40, so each problem shifts Pass@1 by 2.5-3.3 points: every reported difference in these comparisons sits inside that noise floor. No confidence intervals, significance tests, or multiple seeds are reported (§4, §5.1), so 'maintaining or improving accuracy' is an overinterpretation of noise-level differences, several of which trend negative. The length claim is also over-scoped: the 30-50% reduction holds against Vanilla/SFT/Direct-DPO/L1 but not against CoD, a listed baseline — on Qwen, ReCUT averages 1,298 tokens vs CoD's 239, a 5.4x increase (Table 1). Additionally, the 8,000 training questions are sampled from DeepScaleR-Preview, which §4 states contains AIME/AMC pairs; if any of the 30 AIME24/30 AIME25/40 AMC23 evaluation problems fall in that sample, the absolute accuracy numbers (e.g., the Qwen AIME25 gain of 13.3 vs Vanilla's 6.7) are contaminated for every trained model. The greedy step-selection rule in Eq. (5) is a legitimate but secondary risk: a bad early step degrades candidate-pool diversity and propagates through the iterative conditioning, yet the w/o Explore ablation (§5.2) provides aggregate evidence that stepwise exploration helps, and this failure mode would not by itself refute the length-reduction result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ReCUT, a method to reduce the reasoning length of LLMs while maintaining accuracy. The method (i) constructs a diverse trajectory pool via stepwise exploration with long/short switching instructions, (ii) trains two DPO models—one favoring accuracy, one favoring shorter reasoning—using preference pairs built from this pool, and (iii) merges them via DARE-Ties parameter interpolation. Experiments on Qwen2.5-7B and Llama-3.1-8B across five math benchmarks report token reductions of roughly 30–50% versus SFT/Direct-DPO/Vanilla/L1 baselines, with comparable or slightly lower Pass@1. The main contributions are the stepwise long-short sampling strategy and the two-model interpolation approach.","tokens_in":12861,"tokens_out":3068,"duration_ms":33942,"significance":"If the claims hold, ReCUT is a practically useful, architecture-agnostic method for reducing inference cost in reasoning LLMs, and the stepwise preference-data construction is a reasonable alternative to whole-trajectory sampling. The paper's strengths are a clear, simple pipeline and direct comparison against several relevant baselines. The length-reduction result is consistently demonstrated. However, the accuracy-maintenance claim is not statistically supported, and a potential training/evaluation data overlap threatens the validity of the absolute accuracy numbers. The method's theoretical grounding is limited—the stepwise greedy selection is a heuristic without analysis. Overall, the idea is promising but the evidence in its current form does not fully support the central 'maintaining accuracy' claim.","major_comments":[{"comment":"The claim that ReCUT 'maintains or even surpasses' baseline accuracy is not supported by the reported results. ReCUT's average Pass@1 is below Direct-DPO on both backbones (Qwen: 45.7 vs 46.4; Llama: 29.1 vs 30.8) and below SFT on Llama (29.1 vs 31.5). For AIME24, AIME25, and AMC23—which contain 30, 30, and 40 problems, respectively—each problem shifts Pass@1 by 2.5–3.3 points, so the observed differences are within the noise floor. No confidence intervals, significance tests, or multiple seeds are reported, making the numerical gaps uninterpretable. The paper should either soften the accuracy claim to 'comparable' with appropriate statistical qualification or provide error bars and significance testing.","section":"§5.1, Table 1"},{"comment":"The training data are sampled from DeepScaleR-Preview-Dataset, which the paper states comprises AIME, AMC, Omni-Math, and STILL questions. The evaluation sets include AIME24, AIME25, and AMC23. If any of these evaluation problems appear in the 8,000-question training sample, the absolute accuracy numbers for every trained model (including ReCUT and all baselines) are contaminated. The paper must clarify whether the evaluation sets were excluded from the training sample, and if not, rerun the experiments with a disjoint split or report the overlap.","section":"§4, Dataset"},{"comment":"The optimal-step selection rule assumes that taking the first step from the continuation with the higher full-trajectory reward is the best next step. This greedy rule is unproven: a longer continuation that corrects an early mistake would receive a high reward, yet its first step may be suboptimal when extended alone. Since Eq. (5) directly determines the partial trajectories used to train the preference models, this heuristic could inject low-quality steps into the candidate pool. The paper should provide an analysis of when this rule is justified, or at least an ablation comparing Eq. (5) against alternatives such as random step selection or selecting based on step-level rewards.","section":"§3.1, Eq. (5)"},{"comment":"The L1 baseline appears to be mis-tuned. On Qwen2.5-7B, L1 achieves only 57.6% Pass@1 on GSM8K, whereas Vanilla reaches 87.4%; similarly low L1 scores appear on MATH500 and AMC23. This suggests the L1 implementation does not match its original paper's hyperparameters or training procedure. As a result, the claim that ReCUT 'significantly outperforms' L1 is not a meaningful comparison. The authors should verify the L1 baseline setup and report the settings used, or replace it with a correctly tuned variant.","section":"§5.1, Table 1 (L1 baseline)"},{"comment":"The abstract's claim of 'reducing reasoning lengths by approximately 30–50%' is over-scoped. While ReCUT does reduce token count substantially versus Vanilla, SFT, Direct-DPO, and L1, it does not hold against the Chain of Draft (CoD) baseline: on Qwen2.5-7B, ReCUT averages 1,298 tokens versus CoD's 239, a 5.4x increase. The paper should either exclude CoD when making the length-reduction claim or state the comparison explicitly with the baselines to which the 30–50% figure applies.","section":"Abstract; §5.1"}],"minor_comments":[{"comment":"The Limitation section contains the typo 'stepwise seasoning trajectory exploration'—presumably 'stepwise sampling trajectory exploration'. This should be corrected.","section":"Limitation"},{"comment":"The paper names the two specialized models 'Gemini LLMs', which is potentially confusing given Google's Gemini models. Consider using a different name such as 'twin models' or 'complementary models'.","section":"§3.2 and §4, Implementation Details"},{"comment":"The text states that DARE-Ties density Topx and weight α are set to 0.25, but the Limitation section later says prior work typically uses 0.3–0.4. This inconsistency should be clarified.","section":"§4, Implementation Details"},{"comment":"The reward function in Eq. (4) uses the raw token count |Y[t]|, but the paper does not specify how tokenization is performed (e.g., whitespace, subword units). This should be stated for reproducibility.","section":"§3.1, Eq. (4)"},{"comment":"For the Direct-DPO baseline, the paper says it generates 20 trajectories per question; for L1, no training details are given beyond 'GRPO'. Please provide the full hyperparameter settings for all baselines, including learning rates, LoRA rank, and number of steps, so the comparisons are reproducible.","section":"§4, Baseline setup"},{"comment":"The paper reports 'reasoning steps' but does not define how a 'step' is counted, particularly for the Vanilla LLM which may not produce explicit [STEP] markers. Please clarify the step-extraction procedure.","section":"§5.3, Figure 4"},{"comment":"The paper claims 'All codes and data will be released via https://github.com/NEUIR/ReCUT', but the repository is not yet available. Please either provide a link to a working repository or state the release date.","section":"Overall"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and practical problem, and the stepwise exploration idea is a credible contribution. My main concern is that the accuracy-maintenance claim is not established with the current experimental evidence: the differences are within noise, no statistical measures are given, and the potential training/evaluation data overlap could invalidate the absolute results. These issues are fixable by additional experiments and careful re-claiming, so I recommend major revision rather than rejection. The authors should be encouraged to add multi-seed runs, significance tests, and a data-leakage check, and to rephrase the central claim to match the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"ReCUT is a new assembly of known pieces—stepwise long-short sampling, dual DPO preference sets, and DARE-Ties interpolation—and the token-compression effect is real: on Qwen it cuts average output from 2,233 (Vanilla) to 1,298 tokens, and on Llama from 5,331 to 2,058. The ablations show the stepwise exploration and the long-short prompts each contribute. That part is worth a look.\n\nThe soft spot is the accuracy-maintenance claim. In Table 1 ReCUT's average Pass@1 is below Direct-DPO on both backbones (45.7 vs 46.4 on Qwen; 29.1 vs 30.8 on Llama) and below SFT on Llama. AIME24, AIME25 and AMC23 contain 30-40 problems, so each problem moves Pass@1 by 2.5-3.3 points. Every gap here is inside that noise floor, and no seeds or confidence intervals are reported. 'Maintaining or even surpassing accuracy' is an overreading of numbers that actually trend negative. The honest claim is 'roughly comparable accuracy with 30-50% fewer tokens,' and even that needs a caveat: the length reduction does not hold against CoD, which is 5.4x shorter on Qwen (at a big accuracy cost).\n\nTwo smaller issues. The training sample is 8,000 questions from DeepScaleR-Preview, which includes AIME/AMC; the evaluation uses AIME24, AIME25 and AMC23, and the paper does not say whether overlap was removed. If any evaluation problems appear in training, every trained model is inflated relative to the zero-shot baselines. And the greedy step-selection rule in Eq. (5) is a heuristic, but the w/o Explore ablation gives at least aggregate evidence that stepwise sampling helps, so I'd treat that as secondary.\n\nThe paper is clearly written, the limitation section is honest about the instruction-following dependence and the ad hoc DARE-Ties settings, and the pipeline is reproducible in principle (though the code link is not yet live). I'd send it to review. The authors should be asked to soften the accuracy claim, add uncertainty estimates or multiple seeds, report train/eval overlap, and qualify the length comparison. A solid, incremental contribution with an overstated headline.","headline":"Token reduction is real and useful, but 'maintains accuracy' is not supported by the table; worth a revision and a serious referee.","tokens_in":13380,"tokens_out":3211,"would_cite":true,"duration_ms":34204,"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":"ReCUT uses stepwise trials and preference learning to cut LLM reasoning length by 30-50% while keeping accuracy.","keywords":["reasoning compression","chain-of-thought","preference optimization","DPO","parameter interpolation","overthinking","LLM efficiency","mathematical reasoning"],"falsifier":"Run ReCUT's exploration but replace the Eq. (5) selector with a coin flip between the long and short first steps, keeping the rest of the pipeline identical; if the 30-50% token reduction survives, the stepwise optimal-step mechanism is not what drives compression. Alternatively, inspect the chosen first steps and measure how often they are steps that the final correct trajectory later abandons.","tokens_in":12177,"feed_emoji":"✂️","tokens_out":7719,"duration_ms":76845,"temperature":0.7,"pith_summary":"ReCUT is a method for making LLMs reason much more tersely without giving up accuracy. The idea is to generate a pool of reasoning paths of very different lengths by stepping through a problem: at each step the model is asked to continue the current best partial solution twice, once with a 'think deeply' instruction and once with a 'be concise' instruction, and the better continuation is chosen using a reward that rewards correct answers and penalizes length. The pool is converted into two preference datasets, one designed to reward accuracy and one to reward shortness, and two models are fine-tuned with DPO. Merging those two models by parameter interpolation gives a single model that keeps the accuracy of the first and the brevity of the second. On math benchmarks with two 7B/8B backbones, the paper reports reasoning lengths roughly 30-50% shorter than a vanilla step-by-step model while Pass@1 accuracy is maintained or improved.","feed_headline":"ReCUT cuts LLM reasoning length 30–50% and keeps accuracy","feed_subtitle":"Stepwise trials and two preference-tuned models, merged by weight interpolation, let a model think shorter and still solve math.","key_machinery":"Three components carry the argument. Long-Short Switched Sampling prompts the model twice at every step, once for a long and once for a short continuation of the current best partial trajectory, and stores both completed paths in a candidate pool. Optimal Reasoning Step Selection scores each completed candidate by $\\pm 1/|Y_{[t]}|$ depending on whether its final answer matches the gold answer, and in Eq. (5) takes the first step of the higher-scoring continuation as the next step of the growing optimal path. Gemini-LLM training then builds two DPO datasets from the pool, both with the shortest correct trajectory as the positive sample but with the longest incorrect trajectory as the accuracy-focused negative and the longest correct trajectory as the length-focused negative; the two resulting models are merged with DARE-Ties parameter interpolation to balance the objectives.","core_discovery":"The paper's central claim is that stepwise exploration with long-short switching yields a candidate pool diverse enough in both length and correctness that preference optimization alone can compress reasoning by 30-50% without sacrificing accuracy. On the reported benchmarks, ReCUT lowers average generated tokens from 2,233 to 1,298 for Qwen2.5-7B and from 5,331 to 2,058 for Llama-3.1-8B, while the average Pass@1 rises from 43.9 to 45.7 on Qwen and from 27.6 to 29.1 on Llama. The paper also argues that the stepwise selector, not just the length prompting, is responsible: ablations that remove stepwise exploration or the long-short prompts both lose accuracy, and the full model compresses both correct and incorrect trajectories, shrinking the gap in length between them.","pith_inferences":["Beyond the paper, the same stepwise pool construction should transfer to any domain with a cheap correctness signal, such as code generation or constrained QA, where correct and incorrect completions can be scored automatically.","Beyond the paper, the reward in Eq. (4) only distinguishes correct from incorrect at a fixed $\\pm 1/|Y|$ scale; a graded reward that gives credit for shorter correct paths would likely sharpen the length-accuracy frontier.","The authors themselves flag that exploration quality depends on instruction-following and that the DARE-Ties merge parameters are taken from prior practice; both are natural points to vary in follow-up work."],"forward_implications":["On the reported math benchmarks, ReCUT uses roughly half the tokens of vanilla step-by-step prompting while matching or beating its Pass@1, so per-query inference cost drops by about 30-50%.","ReCUT compresses incorrect trajectories more than correct ones, so the overthinking pattern in which wrong answers run much longer than right ones is reduced.","The method is model-agnostic and preference-based, demonstrated on two different 7B/8B backbones, so it can be layered onto other instruction-following LLMs without architectural changes.","Because the final model is a parameter interpolation of an accuracy-optimized and a length-optimized model, the trade-off can in principle be dialed by choosing the interpolation weight rather than retraining."],"supporting_citations":[{"why":"Provides the DPO objective used to fine-tune the accuracy- and length-specialized models.","marker":"Rafailov et al. (2023)"},{"why":"Provides the DARE-Ties parameter interpolation used to merge the two specialized models.","marker":"Yu et al. (2024)"},{"why":"Defines the L1 baseline and the joint length-plus-accuracy reward perspective that ReCUT extends.","marker":"Aggarwal and Welleck (2025)"},{"why":"Supplies the Chain-of-Draft prompt-only compression baseline that motivates training-based compression.","marker":"Xu et al. (2025)"},{"why":"Motivates stepwise decoding as a way to search for less redundant reasoning paths.","marker":"Wang et al. (2025)"},{"why":"Supplies the GRPO algorithm used by the L1 baseline.","marker":"Shao et al. (2024)"},{"why":"Is the source of the DeepScaleR training question-answer pairs.","marker":"Luo et al. (2025)"},{"why":"Provides the Qwen2.5-7B backbone for the first set of experiments.","marker":"Yang et al. (2024)"},{"why":"Provides the Llama-3.1-8B backbone for the second set of experiments.","marker":"Grattafiori et al. (2024)"}],"fun_headline_variants":["ReCUT trims LLM reasoning 30–50% without losing accuracy","Stepwise trails let ReCUT shorten LLM reasoning, hold accuracy","ReCUT compresses LLM thoughts 30–50%, accuracy steady","Preference tuning slashes reasoning length by 30–50%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the greedy step-selection rule: the first step of whichever full continuation earns the higher reward is assumed to be the right next step, and the paper's own Limitation section concedes that the entire exploration depends on the base model faithfully following the long and short instructions.","fun_headline_variants_meta":{"raw":{"variants":["ReCUT trims LLM reasoning 30–50% without losing accuracy","Stepwise trails let ReCUT shorten LLM reasoning, hold accuracy","ReCUT compresses LLM thoughts 30–50%, accuracy steady","Preference tuning slashes reasoning length by 30–50%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000135,"raw_usage":{"total_tokens":1156,"prompt_tokens":968,"completion_tokens":188,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":107}},"tokens_in":584,"tokens_out":188,"duration_ms":2987,"temperature":1.0,"reasoning_tokens":107,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:17:27.736882+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ReCUT's exploration but replace the Eq. (5) selector with a coin flip between the long and short first steps, keeping the rest of the pipeline identical; if the 30-50% token reduction survives, the stepwise optimal-step mechanism is not what drives compression. Alternatively, inspect the chosen first steps and measure how often they are steps that the final correct trajectory later abandons.","supporting_citations":[{"cited_title":"Manning, Stefano Ermon, and Chelsea Finn","cited_arxiv_id":null,"evidence_quote":"Provides the DPO objective used to fine-tune the accuracy- and length-specialized models."}],"review_version":1}