{"id":"bc9ab6a7-8a8a-4a98-895a-8504caf6f642","arxiv_id":"2508.08940","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A curriculum that starts LLMs with generous reasoning token budgets and tightens them during training produces more accurate and token-efficient math reasoning than fixed-budget training.","lead":"This paper trains language models to reason with longer outputs at first, then gradually requires shorter outputs over training, using curriculum learning. The method claims higher accuracy and lower token cost than fixed-length reasoning budgets, which could cut inference expenses for deployed reasoning assistants.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Potential confound between curriculum ordering and total RL training signal makes the reported gains non-identifiable from the abstract.","rationale":"The reader correctly flagged the exploration-to-compression premise as an unverified assumption. My stress-test focuses on a different but equally load-bearing issue: the experimental contrast may not be controlled for total RL training signal. Even if exploration-to-compression is a real phenomenon, the reported gains could arise simply because curriculum training gives the model more opportunities to encounter successful rollouts early in training. The abstract's language ('consistently outperforms fixed-budget baselines at the same final budget' and 'progressive constraint serves as a powerful inductive bias') asserts an ordering effect, but the abstract omits the necessary controls. The test I propose directly checks the ordering effect by permuting budget levels and by matching total training signal. Since the full manuscript is unavailable and the released code is not examined, I cannot resolve the concern; the correct verdict remains UNVERDICTED. I give partial agreement with the reader because we both see the exploration-to-compression dynamic as central, but my concern is more specifically about the experimental isolation of that dynamic from training-signal confounds.","tokens_in":669,"tokens_out":2288,"duration_ms":27456,"concrete_test":"Using the released code, rerun the main comparison with three additional controls: (1) a fixed-budget baseline at the final budget trained with the same number of policy updates and the same total generated token volume as the curriculum run; (2) a shuffled-curriculum condition in which the sequence of budget levels is randomly permuted but the marginal distribution over budgets is identical; (3) a fixed-budget baseline at the arithmetic mean of the curriculum's budget level sequence. If curriculum training still beats all three controls with statistical significance on all five datasets, the inductive-bias claim is supported. If not, the advantage is attributable to total training signal or the marginal budget distribution rather than the curriculum ordering.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central claim is that curriculum-based length control outperforms fixed-budget baselines at the same final inference budget. The least secure condition is whether the comparison isolates the curriculum ordering, rather than the total volume of reinforcement-learning signal. If the curriculum runs the same number of policy updates as a fixed-budget baseline at the final small budget, then early large-budget rollouts provide denser reward signals and more successful trajectories, potentially improving the final policy even if the ordering of budgets is irrelevant. The abstract does not state whether total training steps, total generated tokens, reward weights, and baseline convergence were matched across conditions. Without such controls, the conclusion that progressive constraint is a powerful inductive bias is not yet established. A second aspect is that the decay schedule is ablated but the abstract does not report how sensitive the results are to schedule hyperparameters across the five datasets; if the best schedule is tuned per task, the universality claim weakens. This concern is not about internal inconsistency but about the strength of the empirical evidence given abstract-only access.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a curriculum learning strategy for length-controlled reasoning in LLMs. Starting from generous token budgets and gradually tightening them during GRPO-based RL training, the method aims to let models first discover effective solution strategies and then compress them into concise reasoning traces. The reward function combines task correctness via verifier feedback, length efficiency, and formatting adherence via structural tags. Experiments are claimed on GSM8K, MATH500, SVAMP, College Math, and GSM+, reporting consistent gains over fixed-budget baselines at the same final inference budget, with ablations of reward weighting and decay schedule. Code and checkpoints are released.","tokens_in":862,"tokens_out":1325,"duration_ms":15635,"significance":"If the claims hold, this is a practically valuable result: it suggests a simple curriculum inductive bias that improves token efficiency without sacrificing accuracy, and it provides an explicit comparison against fixed-budget training, which is the relevant baseline for efficiency. The use of external benchmarks and verifier-based rewards makes the evaluation non-circular, and the release of code and checkpoints supports reproducibility. However, with only the abstract available, the effect sizes, error bars, and experimental controls cannot be assessed, so the significance is conditional on the full paper providing the missing details.","major_comments":[{"comment":"The central claim that curriculum-based training outperforms fixed-budget baselines at the same final budget is not identifiable from the abstract because the training-signal confound is not addressed. If the curriculum schedule runs the same number of policy updates as the fixed-budget baseline, then early large-budget rollouts provide denser reward signals and more successful trajectories, which could improve the final policy even if the ordering of budgets is irrelevant. The abstract does not state whether total training steps, total generated tokens, reward weights, or baseline convergence were matched across conditions. This control is load-bearing for the conclusion that progressive constraint is a powerful inductive bias, and the full paper must report it.","section":"Abstract (and missing Experiments section)"},{"comment":"The abstract claims ablations of reward weighting and decay schedule design but reports no quantitative results or sensitivity analyses. Since the decay schedule is a key component of the proposed method, the paper should report how sensitive the results are to schedule hyperparameters across all five datasets; if the best schedule is tuned per task, the universality claim weakens. This is essential for assessing whether the method is robust or merely a per-task tuned recipe.","section":"Abstract (ablations)"}],"minor_comments":[{"comment":"The phrase 'significantly improved token efficiency' would benefit from reporting the actual token savings or efficiency metric (e.g., average response length, tokens per correct answer) rather than leaving it as a qualitative claim.","section":"Abstract"},{"comment":"The list of benchmarks is presented without context on dataset sizes or difficulty; adding a brief characterization (e.g., grade-school vs. competition math) would help the reader gauge the scope of the evaluation.","section":"Abstract"}],"recommendation":"uncertain","confidential_remarks":"The manuscript is under review based only on its abstract, which makes it impossible to verify the central empirical claims. The stress-test concern about the confound between curriculum ordering and total RL training signal is valid and must be addressed in the full paper. I recommend that the editor obtain the full manuscript before making a decision; if the paper lacks the matched-budget controls, it would require major revision or rejection depending on the extent of the gap."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a sensible, clearly described method paper — curriculum over inference-token budgets in GRPO — and the release of code/checkpoints is a plus. The core claim, that progressive tightening beats fixed-budget training at the same inference budget, is plausible but I can't verify it from the abstract alone. The abstract gives no effect sizes, no token budget numbers, no schedule details, and no ablation results; the comparison to fixed-budget baselines could be confounded if the curriculum runs more total training signal (more successful rollouts early on) rather than benefiting from the ordering per se. That's not a strike against the idea — it's a testable, standard control — but the current text leaves it open.\n\nWhat's genuinely new: existing length-control methods use fixed budgets; a curriculum over budgets is a natural and potentially valuable twist, and framing it as exploration-then-compression is a reasonable inductive bias. The reward design (correctness + length efficiency + formatting tags) is a sensible engineering choice. The evaluation spans five math benchmarks, which is decent coverage.\n\nSoft spots: all empirical claims are unverifiable from what we have. Even if the full paper has the right controls, the abstract oversells the 'powerful inductive bias' claim without reporting sensitivity to schedule hyperparameters. The stress-test note about matching total RL signal is on point; the paper needs to state whether total generated tokens and policy updates were matched across conditions. Also, the best schedule might be per-task tuned, which would weaken the universal claim.\n\nWho is this for: people working on inference-efficiency for reasoning models, RLVR/GRPO. It's a solid contribution to that subfield, not a paradigm shift. I'd send it to a serious referee — the method is coherent, the experiments are standard benchmarks, and the code release helps reproducibility. But I'd ask the referee to check the training-signal confound and schedule sensitivity carefully. My own earlier skepticism about novelty was moderate; the abstract actually frames the novelty well, so I don't see a reason to doubt the authors' awareness of prior work.","headline":"Plausible and well-framed method for efficient reasoning, but the abstract can't support the central empirical claim — needs the full paper to check the curriculum-vs-training-signal confound.","tokens_in":1337,"tokens_out":1749,"would_cite":false,"duration_ms":17808,"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":"Shrinking reasoning budgets over training beats fixed budgets","keywords":["curriculum learning","length-controlled reasoning","reasoning efficiency","Group Relative Policy Optimization","GRPO","token efficiency","language models","mathematical reasoning"],"falsifier":"Train a model on a fixed small budget and on a shrinking-budget curriculum that ends at the same budget, controlling for total compute and data. If the fixed small model matches the curriculum model's accuracy on held-out math problems, or if a curriculum that starts small and grows large performs equally well, the specific exploration-to-compression ordering claimed here is not what drives the reported gains.","tokens_in":536,"feed_emoji":"⚡","tokens_out":2434,"duration_ms":27175,"temperature":0.7,"pith_summary":"This paper argues that the best way to train an LLM to reason efficiently is not to fix a single token budget, but to let the budget shrink as training progresses. It proposes a curriculum in which the model first gets generous token allowances to discover correct solution strategies, then is gradually forced to express those strategies in shorter traces. Across five math-reasoning benchmarks, this curriculum is reported to beat fixed-budget training at the same final budget, with both higher accuracy and fewer tokens per solution. If true, this gives a simple, general recipe for getting cheaper reasoning out of language models without task-specific prompt engineering.","feed_headline":"Shrinking reasoning budgets over training beats fixed ones","feed_subtitle":"Curriculum-trained models match or beat fixed-budget baselines while using fewer tokens per answer.","key_machinery":"The mechanism is a curriculum schedule over the token budget combined with GRPO. GRPO is an RL method for language models that updates the policy from a group of sampled responses; here it is augmented with a reward that combines verifier-based correctness, a length-efficiency term, and formatting checks. The curriculum is defined by the decay strategy that lowers the allowed response length over training, so the length constraint is not a constant hyperparameter but a gradually tightening inductive bias. That schedule is what forces the transition from exploration to compression.","core_discovery":"The central claim is that progressive constraint acts as a useful inductive bias for length-controlled reasoning. Starting with a large token budget and annealing it downward over training lets a model explore the strategy space first and then compress the strategies it has found, whereas a model trained from the start at the small budget never encounters the longer successful traces and settles for worse solutions. The paper reports that this curriculum, applied on top of Group Relative Policy Optimization (GRPO) with a reward balancing correctness, length efficiency, and formatting adherence, outperforms fixed-budget baselines matched at the same final budget on GSM8K, MATH500, SVAMP, College Math, and GSM+, improving both accuracy and token efficiency.","pith_inferences":["If the exploration-to-compression dynamic generalizes, the same curriculum should transfer to non-math domains such as code generation or tool use, where correct solutions also come in many forms before being distilled.","A natural extension not tested here is adapting the decay schedule dynamically per model, for example by monitoring validation accuracy and tightening the budget only when accuracy stabilizes; that would test whether the schedule is a universal inductive bias or a tuned hyperparameter.","The method's value depends on the search being able to find good strategies at large budgets; on tasks where longer traces do not contain better strategies, or where correctness is cheap to achieve, the curriculum may offer little over fixed-budget training."],"forward_implications":["At a fixed final budget, models trained with a shrinking budget should reach higher accuracy than models trained with that same budget throughout, on the reported math benchmarks.","Curriculum-trained models should use markedly fewer tokens per correct answer, reducing inference cost at deployment.","The result should hold across multiple datasets (GSM8K, MATH500, SVAMP, College Math, and GSM+) without task-specific prompt engineering, making the schedule a portable training recipe.","Reward weighting and decay schedule design matter: the ablation results imply these choices must be deliberate to get the full benefit.","Because the method targets the training procedure itself, it can be combined with other efficiency techniques rather than replacing them."],"supporting_citations":[],"fun_headline_variants":["Curriculum learning beats fixed token budgets for reasoning","Train with room to explore, then tighten: curriculum wins","Gradual budget annealing improves reasoning efficiency","For better reasoning, shrink token budgets progressively","Start long, end short: curriculum learning for reasoning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that giving a model generous token budgets early lets it discover solution strategies that survive when the budget is later tightened; if longer traces are not where the better strategies live, the curriculum's advantage disappears.","fun_headline_variants_meta":{"raw":{"variants":["Curriculum learning beats fixed token budgets for reasoning","Train with room to explore, then tighten: curriculum wins","Gradual budget annealing improves reasoning efficiency","For better reasoning, shrink token budgets progressively","Start long, end short: curriculum learning for reasoning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000437,"raw_usage":{"total_tokens":2195,"prompt_tokens":890,"completion_tokens":1305,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":1235}},"tokens_in":506,"tokens_out":1305,"duration_ms":12889,"temperature":1.0,"reasoning_tokens":1235,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:30:40.164585+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a model on a fixed small budget and on a shrinking-budget curriculum that ends at the same budget, controlling for total compute and data. If the fixed small model matches the curriculum model's accuracy on held-out math problems, or if a curriculum that starts small and grows large performs equally well, the specific exploration-to-compression ordering claimed here is not what drives the reported gains.","supporting_citations":[],"review_version":2}