{"id":"1366efc8-142c-4869-9add-a74ef77c4890","arxiv_id":"2505.05315","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Training reasoning models with a fixed two-part token budget, thinking plus solution, makes them robust to truncated thinking and more concise overall.","lead":"A Salesforce AI Research team proposes splitting a reasoning model's token budget into a thinking part and a solution part, and training the model with reinforcement learning to keep working well when the thinking part is cut short. The method improves accuracy under tight token budgets on math and coding benchmarks and also shortens answers when no budget is imposed.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The budget-control comparison may not hold total token budgets fixed across methods, so the central 'reliability under tight budgets' claim is not yet established.","rationale":"The reader's weakest assumption was that the thinking/solution segmentation holds; the paper's own Appendix A acknowledges this limitation and it is not the main risk for the math/code benchmarks studied. I agree the paper has loose ends (AMC train/eval overlap, missing error bars, unspecified reward details), but the more load-bearing issue for the central claim is that the budget-control comparison may not hold total token budgets fixed. The method fixes the solution budget at 1K and only varies the thinking budget, so the 'unseen budget constraints' are not arbitrary total budgets. A matched-budget re-evaluation is cheap because the code is public and would settle whether the headline efficiency results are artifacts of unequal token allocation. The core training idea remains plausible and the code is available, so this does not warrant rejection; it warrants an additional condition on the evaluation protocol.","tokens_in":14199,"tokens_out":12648,"duration_ms":126520,"concrete_test":"Recompute the Figure 4 and Table 1 comparisons under a strict matched-budget protocol: for each total token budget c in {1K, 1.5K, 2K, 4K}, evaluate E1 with several (t,s) splits satisfying t+s=c (including s values other than 1K, e.g., 0.25K, 0.5K, 1K), evaluate S1/L1 with a hard cap of exactly c tokens, and report Pass@1 at exactly c (padding/truncating identically if needed). If E1's advantage persists at matched c for c≤1.5K and for small s, the claim holds; if it shrinks or reverses, the 'tight budget' and 'arbitrary budget' claims are overstated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's headline claim is that Elastic Reasoning 'significantly improves reliability under tight resource constraints' (Abstract) by splitting a total budget c into t thinking tokens and s solution tokens. However, the evaluations in Figures 1 and 4 plot Pass@1 against 'Tokens Used' without specifying whether this is the configured maximum budget or the actual generated tokens. For E1, the configured budget is c=t+s with s fixed at 1K in all training and ablations; Figure 7 shows E1-Math at (0.5K+1K) actually uses 1,340 tokens and at (1K+1K) uses 1,799 tokens, i.e., less than the configured maximum. S1/L1 under budget forcing enforce a hard cap of c, so their actual usage equals c. If the x-axis is actual tokens, E1's curve is shifted left and the comparison is efficiency-fair but not aligned at identical budgets; if the x-axis is the configured budget, E1 is allocated a larger total allowance whenever it does not exhaust its budget. More importantly, because s is fixed to 1K, the method only varies t; at a claimed 1K total budget t=0, and no experiment varies s or evaluates total budgets below 1.5K. Thus the 'generalizes to unseen budget constraints' claim (Section 3.2.3) is only demonstrated for unseen t with s frozen at its training value, not for arbitrary total budgets. This directly affects the central efficiency and length-control claims.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Elastic Reasoning, a framework for length-controllable reasoning in large reasoning models. It consists of two components: separate budgeting at inference, which divides a total token budget c into t thinking tokens and s solution tokens and forcibly inserts </think> after t tokens to protect solution generation; and budget-constrained rollout during GRPO training, which trains the model under a fixed budget pair (t*=1K, s*=1K) so that it learns to answer correctly even when thinking is truncated. The method is evaluated on mathematical benchmarks (AIME2024, MATH500, AMC, Olympiad-Bench, Minerva Math) and coding benchmarks (LiveCodeBench, Codeforces, HumanEval+). The paper claims robust performance under strict inference-time budgets, generalization to unseen budget configurations without additional training, lower training cost than L1 baselines, and more concise unconstrained outputs, including a 32.1% token reduction on AIME2024 and a 37.4% token reduction on LiveCodeBench.","tokens_in":14498,"tokens_out":6863,"duration_ms":69441,"significance":"If the results hold, this is a practically valuable contribution. The method is simple and cheap: it requires only 200 training steps for the math model and 30 steps for the code model, and the code is publicly released. The budget-constrained rollout idea is clearly described and the ablation in Table 2 provides evidence that both thinking and solution generation improve, with the solution component benefiting most under short thinking budgets. The evaluation spans multiple math and coding benchmarks, and the unconstrained length reductions are large and potentially useful for deployment. However, two issues substantially weaken the current evidence: AMC appears in both the training and evaluation sets, and the comparison in Figures 4 and 5 does not clearly control total token budgets across methods. These issues are fixable but need to be addressed before the headline claims can be accepted.","major_comments":[{"comment":"The x-axis label of Figures 4 and 5 is 'Tokens Used', but the paper does not state whether this is the configured maximum budget or the actual average number of generated tokens. This matters because Figure 7 shows that E1-Math at a configured budget of (0.5K+1K) actually uses 1,340 tokens and at (1K+1K) uses 1,799 tokens, i.e., less than the configured caps of 1.5K and 2K; S1 and L1 under budget forcing enforce a hard cap, so their actual usage equals the configured budget. If the x-axis is actual tokens, the curves compare methods at different total budgets; if it is the configured budget, E1 receives a larger effective allowance whenever it emits </think> early or finishes the solution early. The headline claim of 'significantly improves reliability under tight resource constraints' therefore needs either an explicit fixed-budget comparison with equal configured budgets for all methods, or a clear statement that the comparison is in actual token usage.","section":"§4.2, Figures 4 and 7"},{"comment":"The math training set explicitly includes AMC, and the math evaluation set also includes AMC. The AMC panel of Figure 4 is therefore not an external evaluation, and any statement that E1-Math-1.5B 'consistently outperforms' baselines across the five math benchmarks is contaminated by this overlap. Please remove AMC from the evaluation or retrain without it; at minimum, report all benchmark numbers separately and do not include AMC in summary claims.","section":"§4.1 and Figure 4"},{"comment":"The claim that the learned policy 'generalizes effectively to unseen budget constraints' is only demonstrated for unseen thinking budgets t with the solution budget fixed at s*=1K. Figure 6 and Figure 7 vary t in {0.5K,1K,2K,3K}, and Table 1 varies t for the code model, but no experiment varies s or considers total budgets below 1.5K. Thus the evidence supports generalization across t, not across arbitrary budget splits or total budgets c. Please either add experiments with varying s (e.g., s in {0.5K,2K} while fixing the total budget) or restrict the generalization claim accordingly.","section":"§3.2.3 and §4.4.2"}],"minor_comments":[{"comment":"The section heading contains a typo: 'mathematical reasoning rsults' should be 'results'.","section":"§4.2"},{"comment":"Please standardize model names: Figure 4 uses 'DeepScalaR' while the text uses 'DeepScaleR'; Figure 7 uses 'E1-Coder' but the model is named E1-Code; Table 2 header says 'E1-Math-1.5.'.","section":"Figures 4 and 7, Table 2"},{"comment":"The notation for the solution budget is inconsistent: Section 3.2.3 uses s*, Section 4.4.2 uses a*, and Table 1 uses a. Please unify the notation.","section":"Throughout"},{"comment":"The paper reports single-run Pass@1 values and Codeforces ratings without error bars or multiple seeds. Given the small performance differences in the unconstrained setting (e.g., +0.3% on LiveCodeBench and +42 Codeforces rating), please provide confidence intervals or results across multiple runs.","section":"Figures 4-7 and Table 1"},{"comment":"Please state explicitly how Pass@1 is computed (greedy decoding or sampling) and whether the token counts in Figure 7 are averages over the evaluation set; this information is needed to reproduce Figures 4 and 5.","section":"§4.2 and Figure 7"}],"recommendation":"major_revision","confidential_remarks":"The AMC contamination and the budget-comparison ambiguity are the two issues most likely to be raised by readers. Both are fixable, but they affect the central claims and should be resolved before acceptance. The paper is otherwise a solid empirical contribution with a clear method and useful negative results, such as the iterative training observation in Table 3."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hanze – quick take on 2505.05315. The core idea is simple and, as far as I can tell, new: split the reasoning budget into thinking and solution, and train the model under truncated thinking rollouts via GRPO. That combination is not in S1 or L1, which either prompt for special tokens or enforce a single total length. The budget-constrained rollout does real work—the ablation in Table 2 shows the solution-generation component improves substantially when thinking is cut short. And the empirical wins are meaningful: E1-Math-1.5B matches L1-Max with far fewer training steps, and E1-Code-14B reaches o1-level Codeforces ratings while cutting tokens by ~37%. The unconstrained conciseness effect is a nice bonus.\n\nThe paper is not without soft spots. The most load-bearing: AMC is in both the training set and the evaluation suite, so one of the five math benchmarks is not external. That doesn't kill the result—AIME 2024, MATH500, the future LiveCodeBench window, and Codeforces are clean—but it should be disclosed and the curves re-reported without AMC. Second, the 'generalizes to unseen budget constraints' claim is only demonstrated for unseen thinking budgets t, with the solution budget s frozen at 1K in both training and inference. That's a narrower claim than 'arbitrary budgets.' It's still useful—t is the parameter you'd most want to vary—but the phrasing should be calibrated. Third, the paper omits error bars, reward details, and leaves the VeRL citation unresolved; those are fixable but need to be addressed.\n\nThe stress-test worry about the x-axis (configured budget vs actual tokens) doesn't land hard: Figure 7 shows E1 uses fewer tokens than the configured cap, so plotting against 'Tokens Used' is efficiency-fair. If anything, E1 looks better at identical actual-token counts. The real caveat is that s is never varied, which I've already flagged.\n\nBottom line: this paper deserves a serious referee. It's a cheap, reproducible method (code is public) with a plausible mechanism and mostly clean external benchmarks. A solid conditional accept: require AMC disclosure, error bars, and a revised generalization claim.","headline":"A genuinely useful length-control method with a clean central idea, despite a few evaluation gaps that need tightening.","tokens_in":15006,"tokens_out":2340,"would_cite":true,"duration_ms":24182,"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":"Elastic Reasoning claims that separating a reasoning model's output into thinking and solution budgets, and training the model under forced truncation of the thinking phase, preserves accuracy under tight token budgets while making…","keywords":["chain-of-thought","token budget control","budget-constrained rollout","GRPO","separate budgeting","reasoning efficiency","test-time scaling","large reasoning models"],"falsifier":"On AIME 2024, evaluate the trained model at thinking budgets of 0.5K, 1K, 2K, and 3K tokens while holding the solution budget at 1K. The central claim predicts a smooth, roughly log-linear accuracy gain and a solution segment that stays near its 1K allowance; a sharp accuracy cliff below some budget, or a solution segment that collapses when thinking is truncated, would show the model has not actually learned to adapt to incomplete reasoning.","tokens_in":14026,"feed_emoji":"🧠","tokens_out":11249,"duration_ms":108359,"temperature":0.7,"pith_summary":"The paper's claim is that a reasoning model can be made budget-aware by splitting every answer into a thinking phase and a solution phase, each with its own token allowance, and training the model with forced truncation of the thinking phase. At inference, when the thinking allowance runs out, the model forcibly emits a closing tag and moves on to writing the solution, so the solution is never starved of tokens. The training method, budget-constrained rollout inside the GRPO policy-gradient loop, which scores each sampled answer against the average score of its group, uses a fixed budget pair of 1K thinking and 1K solution tokens and teaches the model to answer well even when its reasoning was cut short. The reported outcome is that one such training pass transfers to many unseen budgets and, with no budget imposed at all, the model becomes more concise while keeping accuracy—roughly 32.1% fewer tokens on AIME 2024 and 37.4% fewer on LiveCodeBench.","feed_headline":"Split reasoning budgets keep AI accuracy under tight token caps","feed_subtitle":"Reserving tokens for the answer and training on cut-off thoughts cuts output by about a third.","key_machinery":"The load-bearing mechanism is separate budgeting: the model's output is explicitly divided into a thinking segment and a solution segment, with total budget $c=t+s$, and generation always reserves $s$ tokens for the solution even if the thinking budget $t$ is exhausted. The companion training mechanism is budget-constrained rollout: a GRPO variant in which every sampled trajectory obeys the same forced `</think>` rule at a fixed budget pair $(t^*,s^*)=(1\\mathrm{K},1\\mathrm{K})$, with the advantage for each response computed as its reward minus the mean reward of the group, scaled by the group's standard deviation. Together these two mechanisms teach the policy to place useful reasoning early and to write a self-contained solution from partial thinking, which is the property the paper identifies as the source of budget generalization.","core_discovery":"The discovery the paper argues for is that forced truncation does not have to degrade reasoning quality if the model is trained under the same truncation rule it will see at test time. Elastic Reasoning fixes a training budget $(t^*,s^*)=(1\\mathrm{K},1\\mathrm{K})$, rolls out the policy so that `</think>` is appended whenever the thinking budget is exhausted, and optimizes the policy with the GRPO advantage, which compares each generated answer against the average reward of other samples from the same prompt. The paper reports that after roughly 200 steps on math and 30 on code, the policy produces coherent solutions from incomplete thinking, generalizes across thinking budgets from 0.5K to several thousand tokens, and follows a roughly log-linear accuracy-versus-token scaling curve. A further empirical finding is that the solution segment improves more than the thinking segment after training, which the paper uses to explain the generalization to unseen budgets.","pith_inferences":["If the solution segment, not the thinking segment, carries most of the benefit, then the method may transfer to tasks where the final answer obeys a structured format even when the reasoning trace looks very different; a test would be applying the same 1K+1K rollout to a non-math, non-code domain such as multi-hop reading comprehension.","The fixed training budget can be seen as a cheap form of regularization against overthinking; if that interpretation holds, budget-constrained rollout could be an alternative to explicit length penalties in RL for tasks where concise answers are preferred, such as tool-use or customer-support generation.","The clean split between thinking and solution is what makes the method cheap, and it is also its boundary: on tasks where reasoning and answering interleave, the forced `</think>` transition is artificial, so the accuracy-versus-budget curve should be measured there before assuming the gains generalize beyond math and code.","The observed log-linear scaling suggests a deployment policy: allocate larger budgets only to harder problems, since each additional token returns diminishing accuracy; Elastic Reasoning's per-request dial makes this feasible without retraining."],"forward_implications":["A single model trained once at a 1K+1K budget can serve requests at many different token budgets at inference time by adjusting only the thinking allowance, with no further fine-tuning.","The same training recipe also acts as a shortening tool: under no budget constraint, math and code models used roughly 32.1% and 37.4% fewer tokens respectively on AIME 2024 and LiveCodeBench, with accuracy roughly preserved or slightly improved.","Because training is short (about 200 steps for a 1.5B math model, 30 steps for a 14B code model), budget-awareness can be added on top of an existing reasoning model at low compute cost.","Accuracy grows approximately linearly with the logarithm of the token budget, so users can predict the accuracy cost of tightening a latency or cost target.","A second round of training with a larger thinking budget does not improve the result, so the fixed small-budget pass appears sufficient to capture the adaptive behavior."],"supporting_citations":[{"why":"The length-control RL baseline with exact and max budget variants that Elastic Reasoning must outperform under strict budgets.","marker":"[1]"},{"why":"The reasoning model family whose long token-heavy chains motivate the need for length control and whose distillation produces the base models.","marker":"[4]"},{"why":"The contamination-free code benchmark used in both training and evaluation, providing the LiveCodeBench results.","marker":"[12]"},{"why":"The 14B code base model that fails under incomplete reasoning without training, providing the key comparison for code robustness.","marker":"[20]"},{"why":"The 1.5B math base model whose training recipe and unconstrained accuracy serve as the math baseline.","marker":"[21]"},{"why":"The budget-forcing method whose failure to preserve the solution segment motivates separate budgeting.","marker":"[24]"}],"fun_headline_variants":["Budgeted thinking and solutions keep AI accurate under token caps","Train on cut-off thoughts to reason well under short budgets","Elastic Reasoning: split budgets for accurate AI under tight token limits","Truncated thinking training yields robust AI reasoning with fewer tokens","Reserve answer tokens: reasoning stays accurate even when thinking is cut"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole procedure rests on the assumption that a model's output can be meaningfully split into a thinking phase and a solution phase, and that forcing the boundary with `</think>` at any point gives a usable place to start the answer; the paper itself notes in its appendix that this can fail for highly interleaved tasks such as dialogue or commonsense reasoning, in which case separate budgeting loses its foundation.","fun_headline_variants_meta":{"raw":{"variants":["Budgeted thinking and solutions keep AI accurate under token caps","Train on cut-off thoughts to reason well under short budgets","Elastic Reasoning: split budgets for accurate AI under tight token limits","Truncated thinking training yields robust AI reasoning with fewer tokens","Reserve answer tokens: reasoning stays accurate even when thinking is cut"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000539,"raw_usage":{"total_tokens":2581,"prompt_tokens":935,"completion_tokens":1646,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":1560}},"tokens_in":551,"tokens_out":1646,"duration_ms":12267,"temperature":1.0,"reasoning_tokens":1560,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:06:52.100058+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On AIME 2024, evaluate the trained model at thinking budgets of 0.5K, 1K, 2K, and 3K tokens while holding the solution budget at 1K. The central claim predicts a smooth, roughly log-linear accuracy gain and a solution segment that stays near its 1K allowance; a sharp accuracy cliff below some budget, or a solution segment that collapses when thinking is truncated, would show the model has not actually learned to adapt to incomplete reasoning.","supporting_citations":[{"cited_title":"L1: Controlling how long a reasoning model thinks with reinforcement learning","cited_arxiv_id":null,"evidence_quote":"The length-control RL baseline with exact and max budget variants that Elastic Reasoning must outperform under strict budgets."},{"cited_title":"Deepcoder: A fully open-source 14b coder at o3-mini level","cited_arxiv_id":null,"evidence_quote":"The 14B code base model that fails under incomplete reasoning without training, providing the key comparison for code robustness."},{"cited_title":"Tang, Manan Roongta, Colin Cai, et al","cited_arxiv_id":null,"evidence_quote":"The 1.5B math base model whose training recipe and unconstrained accuracy serve as the math baseline."}],"review_version":1}