{"id":"c1e6068c-4e4b-4e2c-922f-d383da2bf814","arxiv_id":"2506.12860","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Removing the question from fine-tuning data preserves a model's concise reasoning style while adding reflective long reasoning, halving token use without losing math accuracy.","lead":"QFFT is a fine-tuning trick for reasoning language models: train on long, step-by-step answers only, with the question removed, so the model keeps its short, efficient style for easy problems and switches to deeper thinking when it hits trouble. On math benchmarks the model keeps about the same accuracy while generating roughly half as many tokens, which would cut the cost of running reasoning AI.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 2's transfer of reflective Long CoT to Short CoT is the load-bearing mechanism, but the QFFT model's own error-triggered switching is never tested; Appendix B.5 only tests DeepSeek-R1-Distill-Qwen-7B.","rationale":"The reader's weakest-assumption pick is the same one I reached. The QFFT training objective (Eq. 2) deliberately removes Q, so the model is never trained on a term that conditions the Long CoT response on the question. That is exactly why the method avoids overriding Short CoT, but it also makes the inference-time switch a pure extrapolation. The pilot (Fig. 2b) and the difficulty-stratified Long CoT proportions (Fig. 5a) are consistent with the claimed behavior, but consistency is not causation. I credit the paper for the pilot, the case study, the phi-backbone replication (Table 7), and the low-resource/noise robustness evidence, which all strengthen the empirical base. The remaining gap is decisive: the only direct test of the transfer assumption uses a different model, so the central explanation of hard-problem parity is currently an assertion. A conditional revision that reruns Appendix B.5 on the actual QFFT checkpoints would resolve it. I considered making the absence of error bars or the overstated '>50%' token-reduction headline (the averages in Tables 2 and 6 show reductions of roughly 15-41% on the reported averages) the primary concern, but those are quantitative caveats; the transfer assumption is the point where the mechanism itself could be false, not just the headline overstated.","tokens_in":22611,"tokens_out":12151,"duration_ms":118063,"concrete_test":"Run the Appendix B.5 continuation experiment with the actual QFFT-trained models (e.g., S1.1-QFFT-7B/32B and LIMO-QFFT-7B/32B) instead of DeepSeek-R1-Distill-Qwen-7B. Use ProcessBench short-CoT step sequences, construct matched pairs where an arithmetic step is correct versus contains an error, truncate at that step, and measure the frequency of reflective markers (e.g., 'wait', 'let me recheck', backtracking sentences) in the model's continuation. If QFFT models do not show a significantly higher reflection rate after erroneous prefixes than after correct prefixes, Assumption 2 is unsupported and the adaptive-switching explanation of hard-problem parity fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism, stated as Assumption 2 in Section 3.4, is that reflective behavior learned as P_theta(Br | U_L) in Long CoT contexts transfers to Short CoT contexts as P_theta(Br | U_S). This transfer is what lets the model default to Short CoT and then spontaneously switch to Long CoT when the quick answer is wrong or uncertain; without it, hard-question accuracy should fall below SFT. The paper's only direct evidence, Appendix B.5, runs a continuation experiment on DeepSeek-R1-Distill-Qwen-7B, not on any QFFT-trained model. Showing that an already-long-CoT model reflects more often after erroneous prefixes does not show that a model trained question-free on Long CoT responses has acquired the same conditional trigger and can apply it inside its own Short CoT generation. The case study (Figure 3) and GPT-4o behavior classification (Figure 4) show that when QFFT does generate Long CoT, it is often preceded by 'wait' and classified as verification/backtracking, but there is no control comparison: we never see QFFT responses on matched questions with error-free Short CoT prefixes, nor a measure of the pre-switch error rate. Main-table differences on the hardest benchmarks are negative in most QFFT rows (AIME25: -1.0, -1.5, -1.8, -0.8 across S1.1/BS/32B S1.1/32B LIMO; AIME24: -0.2, -2.1 for 32B), so without testing the actual QFFT model's switching, the 'comparable to SFT on hard questions' part of the claim is not secured. If transfer is weak in QFFT models, these small deficits may be real and the central adaptive-reasoning story fails.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Question-Free Fine-Tuning (QFFT), which removes the input question from the training data and fine-tunes a base model on Long CoT responses alone. The central claim is that this preserves the model's default Short CoT patterns while teaching it reflective Long CoT behaviors, so the resulting model adaptively uses short reasoning for simple questions and switches to long reasoning when it detects errors or uncertainty. Experiments on Qwen2.5-7B/32B and Phi-4-mini backbones over three distillation datasets and six math benchmarks report accuracy comparable to standard SFT while reducing average token length by more than 50%. Additional experiments report advantages for QFFT over SFT in noisy, out-of-domain, and low-resource settings, and the authors introduce a new metric, Reasoning Adaptability Cohen's Kappa (RAK), to quantify adaptive reasoning. The method is simple, the code is released, and the empirical pattern of token reduction with roughly matched accuracy is consistent across several configurations.","tokens_in":23007,"tokens_out":4788,"duration_ms":45389,"significance":"If the findings hold, the main contribution is conceptually interesting and practically useful: it suggests that a reasoning style can be injected from answer text alone, without learning question-to-response mappings, thereby avoiding the override of an existing concise style. The paper ships reproducible experiments across multiple backbones, distillation datasets, and benchmarks, and the question-free training recipe is simple enough to be broadly adopted. However, the evidence for the central adaptive-switching mechanism is incomplete: the transfer assumption in Section 3.4 is load-bearing but is only tested on a different model in Appendix B.5, and the RAK metric used to measure adaptive reasoning lacks an operational definition of Short vs Long CoT. The reported accuracy differences on the hardest benchmarks are small and mostly negative, yet no error bars or significance tests are provided. The noisy-data advantage is partly a design artifact because QFFT is never exposed to the mismatched question-answer pairs. These issues are fixable, but they currently leave the strongest claims not fully secured.","major_comments":[{"comment":"Assumption 2, that reflective behavior learned as P_theta(Br | U_L) in Long CoT contexts transfers to Short CoT contexts as P_theta(Br | U_S), is the mechanism that explains why QFFT can match SFT on hard questions. The only direct evidence is a continuation experiment on DeepSeek-R1-Distill-Qwen-7B, not on any QFFT-trained model. Showing that an already-long-CoT model reflects more often after erroneous prefixes does not show that a model trained question-free on Long CoT responses has acquired the same conditional trigger and can apply it inside its own Short CoT generation. The case study in Figure 3 and the GPT-4o classification in Figure 4 do not include a control comparison with error-free Short CoT prefixes or a measure of the pre-switch error rate. Please probe the QFFT models themselves, e.g., by feeding matched correct and erroneous Short CoT prefixes and measuring the switch-to-Long-CoT rate, or by comparing the QFFT model's own error-triggered switching against the SFT baseline on the same questions. Without this, the 'comparable on hard questions' part of the central claim is not secured, especially since the AIME25 and AIME24 differences in Tables 2 and 6 are mostly negative.","section":"§3.4, Appendix B.5"},{"comment":"RAK is the central evidence for the adaptive-reasoning claim, but the paper never specifies an operational rule for classifying a generated response as Short CoT versus Long CoT. Section 5.1 uses the first occurrence of 'wait' as a boundary, but it is not stated that the RAK computation in Tables 2 and 6 uses this rule, a length threshold, or an external classifier. Without a reproducible classification rule and some reliability check, RAK values such as 47.7 versus 3.5 on MATH cannot be interpreted as measuring adaptability rather than whatever the classifier happens to capture. Please specify the exact classification procedure used to compute RAK, report the classifier's agreement with human judgment, and make the annotation code available.","section":"§2.2, Tables 2 and 6"},{"comment":"The central claim that QFFT achieves performance comparable to SFT rests on 16-sample averages with no error bars, confidence intervals, or significance tests. The differences on the hardest benchmarks are small and mostly negative, e.g., AIME25 accuracy for 32B S1.1 is 48.6 for SFT versus 46.8 for QFFT (-1.8), and AIME24 accuracy for 32B LIMO is 56.7 versus 54.6 (-2.1). Given these magnitudes, the word 'comparable' is not statistically supported. Please report variance, confidence intervals, or pairwise significance tests for the main accuracy and token comparisons, at least for the headline tables.","section":"§4.1, Tables 2 and 6"},{"comment":"The Level IV noise comparison is confounded by design. In the QFFT format the question is removed, so a dataset in which each question is paired with another question's answer is not noisy for QFFT at all; only the SFT model is exposed to the mismatched question-answer pairs. The result that QFFT maintains 78.6% accuracy at Level IV is therefore partly a consequence of the method's question-free formulation rather than evidence of robustness to noisy data in the usual sense. Please add a control in which QFFT is trained on internally corrupted responses (e.g., truncated or wrong-answer responses while the question remains absent), or a condition with a present but uninformative question, or soften the robustness claim accordingly.","section":"§6.1, Figure 6"}],"minor_comments":[{"comment":"The notation with the struck-through Q in Equation (2) is not defined; please clarify that it denotes removal of the question and check the typesetting of the conditioning context.","section":"§3.2, Eq. (2)"},{"comment":"The x-axis labeled 'Ratio of Data with Question (%)' includes 0% and 100% but the text describes a logarithmic scale; a log scale cannot include 0%. Please use a broken axis or a consistent scale.","section":"§2.3, Figure 2b"},{"comment":"The header 'A VG' appears to be a typo for 'Average', and the tables are hard to read because column separators are missing. Please reformat them.","section":"Tables 3 and 9"},{"comment":"The subsets 'MATHSystem 1' and 'MATHSystem 2' are used without definition; please define them explicitly as the Short-CoT and Long-CoT response subsets described in the surrounding text.","section":"§5.2"},{"comment":"Footnote 4 says the reason for the remaining Long CoT patterns will be discussed in Section 4, but the relevant discussion appears in Section 5.1; please fix the cross-reference.","section":"Footnote 4"}],"recommendation":"major_revision","confidential_remarks":"The paper's efficiency and accuracy results are consistent across multiple configurations, and the core idea is worth publishing if the mechanism evidence is strengthened. The main gap is evidentiary rather than conceptual: Assumption 2 needs to be tested on QFFT models, RAK needs an operational definition, and the noisy-data claim needs a fairer control. I do not see grounds for rejection, but the listed issues are load-bearing for the adaptive-reasoning claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe main claim is real: train on Long CoT responses with the question removed, and you get about half the tokens at accuracy within 1–2 points of SFT across the main benchmarks. That holds across two backbones, three distillation datasets (S1.1, LIMO, Bespoke-Stratos), and six math benchmarks. The method itself is simple to the point of being obvious—the authors acknowledge it is effectively response-only continued pre-training—but the empirical demonstration is genuinely useful: even 0.1%–1% of question–response pairs in the data overrides the model's Short CoT default, while 0% preserves it. That is a clean observation and the paper deserves credit for making it systematically.\n\nThe soft spots are in the interpretation, not the headline numbers. The RAK metric is a creative attempt to measure adaptive reasoning, but it lacks an operational rule for what counts as Short vs Long CoT in a response, and the 16-sample averages are reported without error bars. The more substantive issue is Assumption 2: the claim that reflective Long CoT behavior learned on response-only data transfers to Short CoT contexts, so the model spontaneously switches when it makes an error. The appendix test for this uses DeepSeek-R1-Distill-Qwen-7B, not a QFFT-trained model; there is no control for error-free Short CoT sequences, no pre-switch error-rate measurement, and no direct test on the QFFT model itself. The hard-benchmark numbers (AIME24/25) are consistently 1–2 points below SFT, which is consistent with a weaker transfer than advertised.\n\nThe OOD and noise comparisons are also a bit oversold. SFT-7B's failure to extract answers on MMLU-Pro and LLM-AggreFact is an evaluation artifact, and the \"robustness\" to Level IV noise is partly trivial because QFFT never reads the question. These sections need a fairer baseline before claiming superiority.\n\nNone of this sinks the paper. The token reduction and matched accuracy are the contribution, and they are reproducible (code is public). The authors are also honest about the remaining overthinking on hard problems. This deserves a serious referee. My recommendation: send it out, and ask for revisions that specify the Short/Long classification rule, add variance, and test Assumption 2 directly on a QFFT model. With those, the adaptive-reasoning story would be much more convincing. Without them, the core efficiency result still stands on its own.","headline":"Question-free fine-tuning on long CoT responses cuts tokens ~50% at roughly matched accuracy—the core claim is solid; the adaptive switching story is plausible but under-tested, and RAK needs tightening.","tokens_in":23551,"tokens_out":6158,"would_cite":true,"duration_ms":55383,"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":"Removing the input question from fine-tuning lets a model keep its concise answers and switch to long, reflective reasoning only when needed, matching full fine-tuning accuracy while using about half the tokens.","keywords":["question-free fine-tuning","adaptive reasoning","chain-of-thought","overthinking","reasoning distillation","token efficiency","mathematical reasoning","kappa statistic"],"falsifier":"Run the Appendix B.5 continuation test on the QFFT-trained model itself, not on a different distilled model: feed it correct and erroneous Short CoT prefixes and compare how often reflection keywords appear. If the rates are equal, or if the QFFT model's hard-question accuracy on AIME24 and AIME25 falls materially below the SFT model's, the transfer assumption fails. A simpler check: classify the QFFT model's responses by whether the first reflective keyword follows a wrong step; if reflection is not concentrated after errors, the adaptive-switching claim is contradicted.","tokens_in":22436,"feed_emoji":"🧠","tokens_out":10860,"duration_ms":87454,"temperature":0.7,"pith_summary":"Question-Free Fine-Tuning (QFFT) is a training recipe that drops the input question entirely and fine-tunes a language model only on long chain-of-thought responses. The paper claims this preserves the model's original concise, direct answering style while teaching it the reflective \"wait, check, revise\" behavior that long reasoning traces contain. On six mathematics benchmarks, QFFT is reported to match standard Supervised Fine-Tuning in accuracy while cutting average response length by more than half, and it outperforms SFT on noisy, out-of-domain, and low-resource data. The broader point is that a reasoning style can be added to a model from answer text alone, without learning a question-to-long-answer mapping that would override the model's default behavior. If correct, this gives a cheap, robust way to make reasoning models efficient without sacrificing hard-problem accuracy.","feed_headline":"Removing the question from training halves reasoning length","feed_subtitle":"Fine-tuning on answer text alone keeps easy answers short and switches to deep reflection on hard ones.","key_machinery":"The machinery is the question-free objective over the response sequence alone, $L_{\\mathrm{QFFT}} = -\\frac{1}{|R|}\\sum_{t\\in R}\\log P_\\theta(R_t | R_{<t})$. This is the whole method; it can be implemented by deleting the question from a standard SFT template. The paper identifies two equivalences that carry the argument: QFFT is SFT with null questions, so no concrete $Q \\to R$ association is learned, and it is a form of continued pre-training on reasoning traces, so the model's existing abilities are extended rather than replaced. A second load-bearing piece is the transfer assumption that reflection learned for Long CoT contexts, written $P_\\theta(B_r | U_L)$, also applies in Short CoT contexts, written $P_\\theta(B_r | U_S)$, so the model spontaneously starts reflecting when its quick answer hits an error. The RAK score, a chance-corrected agreement statistic between whether a question is hard for the reference Short CoT model and whether the evaluated model chooses Long CoT, is the metric used to demonstrate that the switching actually tracks difficulty.","core_discovery":"The central claim is that the overthinking seen in long-reasoning models is not a property of the long reasoning style itself but of the supervised mapping from questions to long responses. Standard Long CoT SFT teaches the model $P_\\theta(R | Q)$ for a long reflective response $R$, and this override makes the model produce long answers even for simple questions. QFFT instead optimizes $L_{\\mathrm{QFFT}} = -\\frac{1}{|R|}\\sum_{t \\in R}\\log P_\\theta(R_t | R_{<t})$, a standard causal language modeling objective over the response tokens only, with the question removed. The paper argues that this removes the $Q \\to R$ mapping, so the model's default Short CoT behavior survives, while the response-only training still installs Long CoT reflective patterns. At inference the model is claimed to start with a concise answer and switch into Long CoT when it detects an error or uncertainty, a behavior the paper calls adaptive reasoning. The empirical claim is that QFFT matches SFT accuracy on six math benchmarks, reduces average response length by more than 50%, raises a proposed Reasoning Adaptability Kappa (RAK) score, and beats SFT in noisy, out-of-domain, and low-resource settings.","pith_inferences":["A natural next test is whether the same question-free recipe can inject other behavioral patterns—tool-calling, code formatting, cautious verification—into a model without erasing its default style; the paper gestures at this idea in its closing section but does not test it.","If the noise-robustness claim generalizes, QFFT could let practitioners use raw, uncurated reasoning traces from logs as training signal, since the method does not depend on question-answer alignment; this is an extension, not something the paper demonstrates beyond its synthetic noise levels.","The adaptive-switching explanation predicts that QFFT's advantage over SFT should shrink as question difficulty rises and grow as data quality falls; this is a testable pattern across future benchmarks.","Because QFFT preserves the base model's knowledge, it may be a safer starting point than SFT for continual learning in specialized domains, where later training should add skills without forgetting earlier ones."],"forward_implications":["A model can learn a long, reflective reasoning style from responses alone, so question-answer pairing is not required to transfer reasoning behavior.","Token cost adapts to difficulty: on simple datasets like GSM8K response length falls by roughly three-quarters, while on AIME the reduction is smaller, consistent with the model keeping long reasoning where it is needed.","Because no question-to-response mapping is learned, corrupted or mismatched training responses do much less damage; QFFT holds accuracy under noise levels that make SFT collapse.","Out-of-domain generalization and hallucination behavior are preserved or improved relative to SFT, since the base model's knowledge and instruction-following are not overwritten by a new question-to-response mapping.","QFFT composes with length-reduction methods such as DPO and SimPO, further cutting the tokens spent on long reasoning portions without large accuracy loss."],"supporting_citations":[{"why":"Supplies the override hypothesis: implicit inference from question-to-answer mappings causes the original Short CoT behavior to be forgotten.","marker":"[18]"},{"why":"Provides the S1.1 dataset of 1,000 curated question-long-response pairs used in training and comparison.","marker":"[9]"},{"why":"Provides the LIMO dataset and the observation that math reasoning distillation transfers to nearby domains.","marker":"[10]"},{"why":"Provides the Bespoke-Stratos-17k distillation dataset used for training.","marker":"[8]"},{"why":"Supports the premise that the structure of long reasoning demonstrations, not the question content, is what matters for learning the reasoning style.","marker":"[6]"},{"why":"Also supports the claim that Long CoT capability can be learned from response structure alone.","marker":"[23]"},{"why":"Documents the overthinking problem and supplies the SFT-Shortest, DPO, and SimPO baselines QFFT is compared against.","marker":"[11]"},{"why":"Provides the O1-Pruner baseline and the Accuracy-Efficiency Score metric used to compare length-accuracy trade-offs.","marker":"[14]"},{"why":"Supplies the chance-corrected agreement statistic behind the RAK metric used to measure adaptive reasoning.","marker":"[21]"},{"why":"ProcessBench supplies the correct-versus-erroneous reasoning chains used in Appendix B.5 to test whether reflective behavior transfers to Short CoT contexts.","marker":"[49]"}],"fun_headline_variants":["Drop the question, halve reasoning length","Question-free fine-tuning halves CoT length","Adaptive reasoning: no question, shorter answers","Train on answers alone to cut overthinking","QFFT: match accuracy with half the tokens"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument hinges on the claim that reflective behavior learned inside long, deliberate reasoning chains will transfer to the model's own quick, concise answers, so that when a quick answer contains an error the model spontaneously switches into checking and correcting mode; if that transfer is weak, accuracy on hard questions would drop below SFT.","fun_headline_variants_meta":{"raw":{"variants":["Drop the question, halve reasoning length","Question-free fine-tuning halves CoT length","Adaptive reasoning: no question, shorter answers","Train on answers alone to cut overthinking","QFFT: match accuracy with half the tokens"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000873,"raw_usage":{"total_tokens":3800,"prompt_tokens":987,"completion_tokens":2813,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":2744}},"tokens_in":603,"tokens_out":2813,"duration_ms":20441,"temperature":1.0,"reasoning_tokens":2744,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:06:05.348374+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the Appendix B.5 continuation test on the QFFT-trained model itself, not on a different distilled model: feed it correct and erroneous Short CoT prefixes and compare how often reflection keywords appear. If the rates are equal, or if the QFFT model's hard-question accuracy on AIME24 and AIME25 falls materially below the SFT model's, the transfer assumption fails. A simpler check: classify the QFFT model's responses by whether the first reflective keyword follows a wrong step; if reflection is not concentrated after errors, the adaptive-switching claim is contradicted.","supporting_citations":[{"cited_title":"Sky-t1: Fully open-source reasoning model with o1-preview performance in 450 budget","cited_arxiv_id":null,"evidence_quote":"Provides the Bespoke-Stratos-17k distillation dataset used for training."},{"cited_title":"Note on cohen’s kappa.Psychological reports, 65(1):223–226, 1989","cited_arxiv_id":null,"evidence_quote":"Supplies the chance-corrected agreement statistic behind the RAK metric used to measure adaptive reasoning."},{"cited_title":"let me double-check","cited_arxiv_id":null,"evidence_quote":"ProcessBench supplies the correct-versus-erroneous reasoning chains used in Appendix B.5 to test whether reflective behavior transfers to Short CoT contexts."}],"review_version":1}