{"id":"5e422d94-56f8-4ac2-adef-f2271a63a193","arxiv_id":"2506.14755","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LC-R1 compresses reasoning traces of DeepSeek-R1-Distill models by about 50 percent while keeping accuracy nearly unchanged, by rewarding early termination once the correct answer appears.","lead":"A new post-training method, LC-R1, uses two rewards to make reasoning models stop thinking right after they reach the correct answer, cutting output length by about half with roughly a 2 percent drop in accuracy. It matters because large reasoning models currently waste computing on redundant self-checks, so a reliable compression method could cut inference costs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper validates LC-Extractor on only 100 samples, but its prompt finds the first mention of the answer, not the first derivation; if these disagree, the Compress Reward and VT metric both rest on an unproven boundary.","rationale":"The paper has real strengths: two backbones, seven benchmarks, released code, ablations, and a direct Pass@k analysis. I do not dispute the measured ~40-50% length reduction or the ~2% average accuracy drop, since these are read from model outputs and are not circular. The load-bearing issue is narrower: the method's definition of 'invalid thinking' and its main efficiency metric are produced by the same extractor, and the extractor's prompt operationalizes 'valid' as first mention rather than first derivation. The 100-sample validation does not close this gap. A careful reader should therefore treat the Sufficiency and 'truly redundant' claims as conditional, not established. This does not change the reader's conditional verdict, but it sharpens the reason: the requested revisions should include direct validation of the extraction boundary as a genuine derivation point, with a human-annotated sample and a retraining check on disagreement cases. Without that, the method could be teaching models to stop at the first plausible answer rather than removing truly redundant verification.","tokens_in":21648,"tokens_out":16513,"duration_ms":188159,"concrete_test":"On 500 correct base-model traces (MATH500 and GSM8K, both backbones), have three human annotators mark the first token where the correct answer is derived, not merely mentioned, and measure token-level agreement with LC-Extractor. If agreement is below 95%, retrain the 7B LC-R1 with human-corrected boundaries on the disagreement subset and rerun Table 2; if accuracy, length, or VT shifts by more than one point, the first-mention versus first-derivation gap is load-bearing and the claims need qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"LC-R1's core claim is that post-first-answer tokens are 'invalid' and can be removed with negligible loss. That claim enters through LC-Extractor, which defines valid thinking in Eq. 1, constructs compressed trajectories in Eq. 2, and supplies the Compress Reward in Eq. 11. The extractor's prompt in Figure 5 asks for 'the first sentence that reveals the Ground Truth Answer,' while the text in Sections 2.1 and 3.1 says 'first derived.' These are not the same: a trace can mention a candidate answer before deriving it. The only validation in Appendix B is 98% on 100 MATH samples, measuring agreement with human labels of extraction position, not whether the boundary is a true derivation. If the extractor cuts at an early tentative mention, training rewards premature termination (violating Sufficiency) and the VT rates in Tables 1-3 are inflated for every model, including baselines. Section 5's Pass@k evidence tests a retrained model, so it does not directly show that the removed suffix of the original trace was redundant. The central claim is therefore conditional on an operational definition that is not yet established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LC-R1, a GRPO-based post-training method for compressing the reasoning chains of large reasoning models. The authors introduce a \"Valid Thinking\" (VT) rate (Eq. 1) that measures the fraction of a trace before the correct answer is first derived, and a compression function f (Eq. 2) that truncates the thinking process at that boundary using a learned parser, LC-Extractor. Training combines a Length Reward (Eq. 3) with a Compress Reward (Eq. 11) applied to the </think> token, and the GRPO objective is computed on the compressed trajectories (Eqs. 7-10). Experiments on DeepSeek-R1-Distill-Qwen-7B/1.5B across seven benchmarks report roughly 46-52% average length reduction with about 2% average accuracy drop, and ablations show that both rewards contribute. The paper also claims that compression does not impair exploration ability based on Pass@k curves and that the removed \"invalid thinking\" segments are truly redundant.","tokens_in":21859,"tokens_out":11851,"duration_ms":122988,"significance":"If the results are confirmed, LC-R1 is a practical and inexpensive method for roughly halving the inference cost of small reasoning models with only a small accuracy penalty. The paper deserves credit for releasing code, for evaluating on seven diverse benchmarks, and for including ablations and comparison with several baselines. The central quantitative result in Table 2 is measured on raw token counts and external answer checking, so the headline length-reduction number is not an artifact of the extractor. However, the paper's conceptual claims about \"invalid thinking,\" the VT rates, and the training signal all rest on the LC-Extractor boundary, which is validated on only 100 samples; the training objective is also not derived as a standard policy-gradient surrogate. The lack of error bars and possible train/test overlap with MATH500 make the \"marginal (~2%) drop\" claim difficult to assess. These issues are fixable, so the contribution is promising but needs substantial revision.","major_comments":[{"comment":"The LC-Extractor's boundary definition is ambiguous and its validation is too thin to support the load-bearing \"valid thinking\" construct. The text in Sections 2.1 and 3.1 says the valid prefix runs \"until the correct answer is first derived,\" but the prompt in Figure 5 asks for \"the first sentence that reveals the Ground Truth Answer.\" A trace can mention an answer before deriving it, so these are not equivalent. The only validation in Appendix B is 98% agreement on 100 samples, and the metric is agreement on extraction position, not whether the extracted prefix is sufficient for derivation. Because the Compress Reward in Eq. (11) and the VT metric in Eq. (1) both depend on this boundary, a systematic early-cut error would inflate VT rates and could penalize valid reasoning during training. Please provide a larger human study with a precise definition of \"first derived,\" report inter-annotator agreement, and analyze cases where the first mention and first derivation differ.","section":"Sections 2.1, 3.1, 3.2, and Appendix B; Eq. (1), Eq. (11), Fig. 5"},{"comment":"The GRPO objective in Eq. (7) is computed on compressed trajectories o'_i = f(o_i), but the samples are drawn from the old policy as o_i. The likelihood ratio R_t(θ) = π_θ(o'_{i,t} | ...)/π_θ_old(o'_{i,t} | ...) is not the likelihood ratio of the actually sampled action sequence, because o'_i is a counterfactual sequence that the sampling policy did not generate. The expectation in Eq. (7) is therefore not the standard policy-gradient surrogate for the reward under π_θ, and no off-policy correction or derivation is provided. The paper should either derive the correct objective, or explicitly frame LC-R1 as a heuristic and provide analysis of the induced bias. This matters because the training objective is the core of the proposed method.","section":"Section 3.2, Eqs. (7)-(9)"},{"comment":"All experimental results are single-run, with no error bars or multiple seeds. The central claim of a \"marginal (~2%) drop\" in accuracy is an average of per-benchmark relative changes; on AIME25 (30 problems), the reported difference is under one problem, which is within plausible sampling noise. The paper should report variance across at least three seeds, or bootstrap confidence intervals, and ideally a significance test for the accuracy differences. The length-reduction numbers are large and likely robust, but the accuracy-drop claim needs statistical support.","section":"Section 4.3, Table 2"},{"comment":"The training data is described as \"past AIME competition problems with the MATH dataset\" (Section 4.1), and the evaluation uses AIME25 and MATH500, which are subsets of the AIME and MATH problem distributions. The paper does not state that these evaluation benchmarks were excluded from the training set. If MATH500 (or AIME25) overlaps with training, the corresponding numbers in Table 2 are in-distribution and the generalization claims are weakened. Please clarify the exact train/test split and the contamination controls used.","section":"Sections 4.1 and Appendix C.2"},{"comment":"The Pass@k comparison between the original model and the retrained LC-R1 model does not establish that the removed suffix of the original trace is \"truly redundant.\" The LC-R1 model is a different policy; its similar Pass@k curve reflects the new policy's exploration ability, not the redundancy of the original model's post-answer tokens. To support the redundancy claim, the authors should evaluate the original model with its own traces truncated at the extracted boundary, or otherwise directly measure the contribution of the removed suffix. In addition, the Pass@k curves in Figure 4 are based on 30 AIME25 problems and are presented without error bars, so \"almost perfectly overlaps\" is not quantitatively established.","section":"Section 5, Figure 4"}],"minor_comments":[{"comment":"The heading \"Compression and Efficienct Reasoning Models\" contains a typo: \"Efficienct\" should be \"Efficient.\"","section":"Section 2.1 heading"},{"comment":"The description of AIME25 as \"International Conference on Artificial Intelligence in Medicine\" is incorrect; AIME here stands for the American Invitational Mathematics Examination. This is a factual error that should be corrected.","section":"Appendix C.2, AIME25 description"},{"comment":"The text says \"We use averaged Pass@1 as our primary metric,\" but Table 2 reports pass@10 for LiveCodeBench. Please clarify the exact evaluation metric used for each benchmark.","section":"Section 4.1 and Table 2"},{"comment":"The acknowledgment \"Many thanks to Yao Wan and for his invaluable support\" appears to have a missing name after \"and.\" Please complete the sentence.","section":"Acknowledgment"},{"comment":"The notation \"ans(q)∈t(o'_i)\" uses set-membership for a token sequence; please define it explicitly as \"the ground-truth answer appears in the thinking token sequence t(o'_i).\"","section":"Eq. (11)"},{"comment":"The DPO baseline is described as being applied to the SFT-tuned model, but Table 2 lists \"DPO\" as a standalone method; the naming may confuse readers about the actual training pipeline.","section":"Section 4.2 and Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is formatted as an ICML workshop paper and the core idea is simple and potentially useful. The main issues are the thin validation of the extractor, the unaddressed theoretical status of the training objective, and the lack of statistical rigor in the reported accuracy drops. These are fixable with additional experiments and analysis, so I recommend major revision rather than rejection. I encourage the editor to verify that the authors address the MATH500/training overlap issue explicitly, as this is a standard contamination concern for reasoning benchmarks."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look, but read Table 2 with variance in mind. LC-R1 is a practical, incremental advance in the efficient-reasoning line: the combination of an extractor-defined valid prefix, a compress reward applied only at the </think> token, and GRPO training on the compressed trajectories is new relative to O1-Pruner, ThinkPrune, L1, and DAPO. The headline numbers — roughly 46-52% sequence-length reduction with about 2% average accuracy drop on the 1.5B and 7B R1 distills — are supported across seven benchmarks, and because length is counted on raw tokens and correctness is verified externally by Math-Verify, that central claim does not depend on the extractor being right. The ablations show both rewards matter, and they release code.\n\nThe soft spots are real but not fatal. First, no error bars or multiple seeds anywhere, so the 'marginal' accuracy loss is credible but unquantified; on 30-problem AIME25 the drops are within noise. Second, the 97-98% VT headline is circular — the same LC-Extractor produces the training target and measures the outcome, and its only validation is 100 samples. The stress-test concern that the extractor finds 'first mention' rather than 'first derivation' is fair; if it errs early, training rewards premature stopping. What keeps that in check is the external accuracy signal: if the extractor were badly wrong, you would expect more than a 2% hit. Third, Section 5's 'truly redundant' claim is over-stated: the Pass@k curves show the retrained model still explores, which is good evidence on its own, but they do not show the removed suffix was worthless in the original trace. Also, MATH500 is drawn from the same distribution as the training set, so treat the seven-benchmark average as partly in-distribution; GPQA-D and LiveCodeBench are the cleaner OOD evidence.\n\nI would send this to review: the main quantitative claim is solid enough to deserve referee time, and a good referee would ask for seeds, an OOD-only accuracy summary, and extractor validation that is not self-referential. I would cite it if I worked on reasoning efficiency; probably not otherwise.","headline":"LC-R1 delivers a real, incremental win — extractor-guided </think> compression cuts ~50% of tokens at ~2% accuracy cost — but the VT headline is circular and the evidence lacks variance bars.","tokens_in":22443,"tokens_out":6424,"would_cite":true,"duration_ms":58429,"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 new post-training method identifies and cuts \"invalid thinking\" tokens that appear after the correct answer is first derived, halving output length while losing roughly 2% accuracy.","keywords":["large reasoning models","invalid thinking","valid thinking rate","reasoning compression","GRPO","length reward","compress reward","efficient reasoning"],"falsifier":"Take a fresh sample of reasoning traces from the base and compressed models, have independent annotators mark the first point where the correct answer is actually established, and compare those marks with LC-Extractor's outputs; if agreement is low, or if retraining LC-R1 with a hand-corrected or differently seeded extractor changes the reported length/accuracy trade-off, the central claim is not stable.","tokens_in":21408,"feed_emoji":"✂️","tokens_out":6099,"duration_ms":56841,"temperature":0.7,"pith_summary":"The paper is trying to establish that a specific inefficiency in long-thinking models, which it calls \"invalid thinking,\" can be pruned in post-training without meaningful accuracy loss. Invalid thinking is defined as the tokens produced after the model first states the correct answer, and the paper argues that this tail is largely redundant double-checking. It introduces LC-R1, a GRPO-based method with two rewards, one for overall brevity and one for stopping at the first correct derivation. On 7B and 1.5B reasoning models, the reported result is roughly half the output length with only about a 2% drop in accuracy, plus a large increase in the \"valid thinking\" rate. If true, this is a practical recipe for cutting the inference cost of reasoning models while preserving their problem-solving potential.","feed_headline":"Reasoning models shed half their output for a 2% accuracy cost","feed_subtitle":"A GRPO reward scheme cuts post-answer double-checking while preserving exploration and out-of-distribution accuracy.","key_machinery":"The load-bearing object is the compression function $f$, realized by a finetuned parser called LC-Extractor, which locates the first occurrence of the ground-truth answer inside a thinking trace and truncates the trace at that point. Around this function the method builds a dual reward: a Length Reward, $1 - |o'_i|/\\max_{j\\in C}|o'_j|$, for overall conciseness, and a Compress Reward applied only to the final `</think>` token, equal to the fraction of tokens removed when the answer is correct and $-1$ if the model stops before finding the correct answer. GRPO is then run on the compressed trajectories $o'_i = f(o_i)$, so policy updates favor terminating thinking at the first correct derivation. LC-Extractor performs double duty, generating training targets and measuring the valid-thinking metric, so its reliability is central to both the training signal and the headline efficiency numbers.","core_discovery":"The paper's central claim is that the verbosity of large reasoning models is not diffuse: a substantial fraction of generated tokens occurs after the correct answer has already been derived, and this tail can be removed without harming the model's ability to solve problems. LC-R1 reports average length reductions of 46.3% on DeepSeek-R1-Distill-Qwen-7B and 51.9% on the 1.5B variant, with average accuracy drops of 1.84 and 2.14 percentage points respectively across seven benchmarks. The method also raises the valid thinking rate to above 97%, meaning nearly all thinking tokens precede the first correct answer. The paper further argues that pass@k performance on AIME25 is nearly unchanged after compression, concluding that the pruned \"invalid thinking\" segments are truly redundant rather than exploratory.","pith_inferences":["An extrapolation: the same redundancy pattern appears in the paper's measurements of much larger models, where 35-45% of tokens are invalid thinking, so similar training could plausibly yield comparable savings on models beyond the 7B scale, though LC-R1 itself is only trained and tested up to 7B.","Because LC-Extractor defines both the training signal and the headline metric, a natural testable extension is to use one extractor for training data and an independent hand-annotated extractor for evaluation, which would separate real efficiency gains from artifacts of the measurement.","The \"stop at the first correct derivation\" reward could generalize beyond the `</think>` token to other stopping decisions, such as terminating a tool-use loop or ending a search once a verified answer is found.","Compression-trained models are natural teachers for shorter-chain distillation, potentially letting smaller models inherit brevity without needing reinforcement learning themselves."],"forward_implications":["On seven benchmarks spanning math, general knowledge, and code, LC-R1 reduces average token count by 46.3% on the 7B model and 51.9% on the 1.5B model, with average accuracy drops of 1.84% and 2.14% respectively.","Valid thinking rate rises to 97.1% (7B) and 98.6% (1.5B), meaning almost all generated thinking tokens now occur before the first correct answer.","Pass@k on AIME25 up to $k=128$ is nearly unchanged after compression, so repeated sampling still finds correct solutions at the same rate.","Compression is uniform across problem difficulty, and out-of-distribution benchmarks such as GPQA-Diamond and LiveCodeBench hold up better under LC-R1 than under SFT-based baselines.","Ablations show that both rewards are necessary: removing the Compress Reward lowers the valid thinking rate sharply, while removing the Length Reward reduces overall compression."],"supporting_citations":[{"why":"Supplies the base reasoning models, DeepSeek-R1 and its Qwen distillations, whose thinking traces are compressed and whose length and accuracy define the baseline.","marker":"DeepSeek-AI et al., 2025"},{"why":"Documents the overthinking phenomenon and motivates the SFT baseline that prunes to the initial correct solution.","marker":"Chen et al., 2024"},{"why":"ThinkPrune is the reproduced RL baseline that applies a length-truncation reward, providing the main comparison for compression and accuracy.","marker":"Hou et al., 2025"},{"why":"O1-Pruner is the offline pruning baseline and also the base for the SFT+O1-Pruner hybrid comparison.","marker":"Luo et al., 2025b"},{"why":"Supplies the group-mean subtraction used in the reward normalization to obtain unbiased GRPO advantages.","marker":"Liu et al., 2025"},{"why":"DAPO's group-token objective is adopted so short-correct and long-incorrect sequences do not bias training.","marker":"Yu et al., 2025"},{"why":"Provides the DPO baseline, which constructs chosen and rejected pairs from the shortest and longest correct traces.","marker":"Rafailov et al., 2023"},{"why":"Qwen2.5-3B-Instruct is the backbone for LC-Extractor, the parser that truncates traces at the first correct answer.","marker":"Team, 2024"}],"fun_headline_variants":["Reasoning models cut thinking length in half for 2% accuracy","LC-R1: 50% shorter reasoning, 2% less accurate","Trim the redundant tail: LRMs compress by half with minimal loss","Half the tokens, nearly same accuracy: new LRM method"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the LC-Extractor parser reliably finds the first moment the model states the correct answer, because every training reward and the reported valid-thinking metric depend on that cut point; it was validated on only 100 human-annotated samples.","fun_headline_variants_meta":{"raw":{"variants":["Reasoning models cut thinking length in half for 2% accuracy","LC-R1: 50% shorter reasoning, 2% less accurate","Trim the redundant tail: LRMs compress by half with minimal loss","Half the tokens, nearly same accuracy: new LRM method"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1288,"prompt_tokens":932,"completion_tokens":356,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":280}},"tokens_in":548,"tokens_out":356,"duration_ms":3902,"temperature":1.0,"reasoning_tokens":280,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:46:43.441988+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fresh sample of reasoning traces from the base and compressed models, have independent annotators mark the first point where the correct answer is actually established, and compare those marks with LC-Extractor's outputs; if agreement is low, or if retraining LC-R1 with a hand-corrected or differently seeded extractor changes the reported length/accuracy trade-off, the central claim is not stable.","supporting_citations":[],"review_version":2}