acceptodds
Under review as a conference paper at ICLR 2027

Reasoning, Answer Delivery, and Budgeted Accuracy

Abstract

A late answer cue can raise a reasoning model's score, but its interpretation depends on the comparison. We compare direct continuation D with a four-token cue L on the same prompt and sampled prefix, under one output cap that includes forced tokens. Correct delivery requires a natural end of sequence (EOS), a complete final boxed answer and acceptance by the fixed scorer, which allows numeric tolerance on MATH. On 200 unused MATH questions, preregistered L raises delivery by 7.33 percentage points, 95% interval [5.00, 10.00], for Qwen3.5-9B at 2048 tokens with 45 rescues and one harm, by 6.83 for Ministral-3-8B and by 4.50 for Qwen3.5-9B at 4096 tokens. Registered GPQA-Diamond transfers give +20.20 [16.50, 24.07] with Qwen and +4.55 [2.69, 6.57] with Ministral without harms, and L generates 1–4% fewer tokens. Only runs active at the cue can change, and across nine settings 23–49% of them convert net of harms. L exceeds a budget-charged variant of the closest protocol's conclusion message by 3.50 points [1.50, 5.67], largely through greater answer room, and a numeric-budget system prompt by 9.50 [6.17, 13.00]. The comparison changes reserve selection: the earliest cue looks best against the stopped stream in three MATH settings, but against continuation to the same cap a 128-token reserve beats it by 15–21 of 600 runs at 2048 tokens. Prospectively, at a 1024-token cap, a 128-token reserve beats a 512-token one by 14.33 [10.67, 18.33] and 10.33 [6.50, 14.33] points for the two models, while matched no-cue continuations deliver within 0.67 points of D. Continuing D also beats restarting thinking at 1024 tokens by 8.25 points at the same cap.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, reasoning, test-time compute, token budgets

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_AjkKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper looks at what happens when a model runs out of its token budget before writing a final answer. The proposed fix is simple: if the model is still generating about 130 tokens before a hard output limit, the text "Final answer:" is inserted and the model continues, leaving room for the answer. This is compared with letting the same run continue to the limit. Both versions share the same prompt and the same sampled text up to the insertion point, and the inserted tokens count against the budget, so any difference comes only from runs still going at that point.

The main experiments use Qwen3.5-9B and Ministral-3-8B in their direct (non-thinking) mode, mostly with a 2048-token limit, on MATH and GPQA-Diamond, with three sampled runs per question. An answer counts only if the model stops on its own and gives a complete boxed answer that the scorer accepts. In a preregistered study on 200 unseen MATH questions, the cue improves correct answers by about 7 points for both models at 2048 tokens and by 4.5 points at 4096. On GPQA-Diamond the gains are about 20 points for Qwen and 4.5 for Ministral. The cue also uses slightly fewer tokens overall.

The paper also compares the cue with an adapted version of Sun et al. (2025)'s concluding message and with budget-aware system prompts, and tests which parts of the cue matter. It studies how much room to leave for the answer: a short reserve (64–128 tokens) works better than a long one (512) when compared against letting the model continue. Continuing the original run also works better than discarding it and restarting in thinking mode. Simple accounting identities explain why only unfinished runs can change and why the choice of baseline affects these conclusions.

The evidence includes a development study, a fixed-policy rerun, and several preregistered follow-up studies, analysed with question-level bootstrap intervals and Holm correction.

Critical strengths
  1. The experiments are set up carefully. The version with the cue and the version without it start from exactly the same text, and the extra cue words count against the same token limit, so any difference can only come from runs that were still going when the cue was added. The authors check this instead of assuming it. They also repeat the main experiment on new questions that were fixed in advance, and get about the same gain. This makes the main result easy to believe.

  2. The paper makes a useful point about how to compare methods. If you compare the cue against a run that was cut off at the same point, leaving a lot of room for the answer looks best. If you compare against a run that was allowed to continue to the full limit, leaving a small amount of room works better, and a follow-up test on new settings confirms this. That is a good lesson for anyone evaluating methods that stop or shorten model outputs.

Critical weaknesses
  1. The idea is not new and the gain is narrow. Forcing a final answer near the token limit is already used in s1 budget forcing, Ma et al. (2025) and Sun et al. (2025). When Sun et al.'s message is placed at the same position as the proposed cue, it comes within about two points on MATH, and slightly beats the cue at the 4096-token limit, so most of the advantage comes from leaving more room for the answer rather than from the cue itself. The gain also only comes from runs that would otherwise run out of tokens, so it shrinks as the limit grows. The main 2048-token limit is far below the 32k the model card recommends, and at 2048 tokens the cue does about as well as plain generation with twice the limit. The main comparison also runs with thinking turned off, which doesn't match the "reasoning" framing of the title.

  2. Some results are stated more strongly than the evidence supports. On GPQA-Diamond the cue mostly turns runs with no answer into a forced letter, and wrong answers go from 21 to 184. On four-option questions, giving a random letter to those runs would already recover about three quarters of the reported gain. Several preregistered tests also did not hold: the two prose "time's up" cues showed no gain, the short-versus-long reserve test on GPQA with Ministral found no difference, and the first rerun's planned analysis did not meet its own decision rule. The main text and abstract report only the tests that passed. Since the paper relies on preregistration for credibility, every outcome should be stated in the main text.

Questions for authors

Clarifications:

  1. How does the cue compare with simply raising the token limit, at the same number of generated tokens? What happens at the model card's recommended budget, where few runs hit the limit?
  2. On GPQA-Diamond, what is the gain over giving a random letter to runs with no answer, or after a standard guessing correction?
  3. Can the failed preregistered tests be reported in the main text? And why is the first rerun called a replication when its planned analysis did not meet its decision rule?
  4. Does a late cue also help in thinking mode, for example closing the thinking block and forcing "Final answer:" near the limit?
  5. Does a lower presence penalty change how often runs hit the limit, and so the size of the gain?

Suggestion (future work): choosing the cue position adaptively, based on whether the run looks unlikely to finish in time.

6 comments
AuthorsOpenReview ↗

Response to Reviewer AjkK

Thank you for the detailed review. We answer your questions in order. For the first one we ran a new experiment: the 200 MATH questions of the paper's first repeat study, three new sampling streams, the same 2048-token cap, on Qwen3.5-9B and the larger Qwen3.8-27B, with our cue and the Sun et al. message at equal answer reserves of 64 to 512 tokens and 600 runs per arm.

"most of the advantage comes from leaving more room for the answer rather than from the cue itself" At equal reserve we find no delivery difference between the two cues: the gain comes from a late cue with an adequate answer reserve, under either wording. Cue minus message is +0.33, +0.83, -0.33 and +0.50 points for 9B and -0.33, +1.00, +0.17 and +0.50 for 27B, and every 95% interval lies within [-1.67, +2.17]. At equal reserve the message starts 10 tokens before our cue, so this compares two policies with the same reserve. The revised PDF reports this in Sec. 6.2, Table 5 and Appendix Table 36.

On the gain shrinking as the limit grows: Prop. 2 limits changes to runs that are still generating at the cue, and this is not the same set as the runs that finally hit the cap. Their share can be measured from plain continuation alone and it bounds the gain. Within that share, the converted fraction varied from 23% to 49% across the settings of Table 7. In the new experiment 27B has 121 of 600 such runs against 141 for 9B, and gains +5.33 [3.33, 7.50] against +8.00 [5.33, 10.83].

"How does the cue compare with simply raising the token limit, at the same number of generated tokens?" The cue at cap 2048 comes close to continuation at cap 4096 and uses fewer tokens. On the questions of the second repeat study, the cue at cap 2048 delivers 507 of 600 (84.50%) with 951.8 generated tokens per run, and continuation at cap 4096 delivers 509 (84.83%) with 1226.1, so the cue uses 22.4% fewer generated tokens. At cap 4096 the cue adds a further +4.50 [2.50, 6.83]. Our follow-up comment below gives continuation up to 8192 and 16384 tokens.

At larger budgets the cue keeps a positive gain: +7.67 [5.17, 10.33] at cap 2048, +4.83 [3.00, 7.00] at 4096 and +1.50 [0.67, 2.50] at 8192, and it generates no more tokens than continuation at any cap. The details are in the follow-up comment below.

"Does a late cue also help in thinking mode" In thinking mode the paper tests an early close: forcing the close-and-answer suffix at token 1024 gains +45.79 points [40.07, 51.52] over uninterrupted thinking at the same 2048 cap. The late close is in our follow-up comment below.

On GPQA-Diamond against a random letter: uniform guessing on the 360 runs where continuation delivers no answer gives 303 expected correct answers. The cue delivers 333, which is +5.05 points above that. With a penalty of 1/3 per wrong letter the gain is +11.05 points, and with a penalty λ the cue is ahead for λ < 120/163 ≈ 0.736.

On reporting the tests that were not met: yes. Table 4 of the revised main text lists every planned test with its outcome. The paper no longer calls the first study a replication, and its +7.41 under the delivery contract is reported as a separate analysis. The second study has delivery as its primary endpoint, and its four primary tests are all met.

On the presence penalty: the experiments used presence penalty 2 in direct mode. Penalties 0 and 1 are in our follow-up comment below.

Choosing the cue position adaptively is a natural extension of the fixed position studied in the paper.

AuthorsOpenReview ↗

Follow-up for Reviewer AjkK: presence penalty, late cue in thinking mode, and larger caps

Here are the three experiments you asked about: a lower presence penalty, a late cue in thinking mode, and larger caps. All use Qwen3.5-9B on the same 200 MATH questions and three sampling streams as the new experiment of our first reply, with 600 runs per arm. Counts are correct deliveries out of 600, gains are in points, and brackets are 95% question-level bootstrap intervals.

Presence penalty (direct mode, cap 2048). The cap-hit rate changes little with the penalty, and the cue gains at every penalty. Without the cue, continuation reaches the cap in 114 of 600 runs (19.0%) at penalty 0, 124 (20.7%) at penalty 1 and 131 (21.8%) at penalty 2, and 122, 132 and 141 runs are still generating at the cue. The cue gains +5.67 [3.67, 8.00] at penalty 0, with 36 gains and 2 harms, +8.00 [5.17, 11.17] at penalty 1, with 49 and 1, and +8.00 [5.33, 10.83] at penalty 2. Every run whose outcome changes was still generating at the cue.

Late cue in native thinking mode. A late cue helps in thinking mode at every cap. Thinking runs uninterrupted to 8192 tokens and is cut at each cap. The late cue is a six-token close-and-answer cue forced 134 tokens before the cap, only in runs whose thinking is still open there.

CapUninterrupted thinkingWith the late cueGainThinking open at the cue
204890391+50.17 [44.50, 55.67]496 runs (82.7%)
4096206439+38.83 [33.67, 44.17]368 runs (61.3%)
8192360495+22.50 [17.83, 27.33]216 runs (36.0%)

The cue turns 301, 233 and 135 of the open runs into correct deliveries and harms none. The gain follows the share of runs whose thinking is still open at the cue. The paper's early close, at token 1024 with cap 2048, delivers 377 on the same streams, against 391 for the late cue (+2.33 [-2.17, 7.00]).

Larger caps (direct mode, penalty 2). Continuation is one stream run to 16384 tokens and cut at each cap. The cue is placed 132 tokens before each cap.

CapContinuation: correct, mean tokensCue: correct, mean tokensRuns still generating at the cueGain
2048441, 1027487, 1001143+7.67 [5.17, 10.33]
4096491, 1369520, 135680+4.83 [3.00, 7.00]
8192535, 1709544, 170525+1.50 [0.67, 2.50]

At cap 16384 every run has finished before the cue position, so the cue is never applied and both arms deliver 549 of 600 with 1775 tokens. The gain is bounded by the number of runs still generating at the cue (Prop. 2), and the cue generates no more tokens than continuation at any cap. The cue at cap 2048 reaches 487 correct deliveries with 1001 mean generated tokens, where continuation needs cap 4096 and 1369 tokens to reach 491.

Thank you for the detailed response, and for running every experiment I asked about. I went through the revised PDF.

A lot is fixed now. The equal-reserve comparison answers my question about where the gain comes from: your cue and the Sun et al. message do the same at every reserve, and the abstract now says this. Table 4 shows every registered test, including the ones that were not met, and the first rerun is no longer called a replication. Section 6.1 gives the random-letter baseline for GPQA-Diamond. Section 6 is also easier to read in three parts.

But my main concern is still there, and the new results make it clearer.

First, if the wording does not matter, then the intervention is the same as in prior work. What is left is the evaluation setup and the reserve finding. I find both useful, but it is a narrow contribution.

Second, in direct mode the gain goes from 7.67 points at 2048 tokens to 1.50 at 8192, and to zero at 16384. That is still only half of the 32768 tokens the model card gives as typical.

Third, the thinking-mode gains look large, but the baseline there is weak. At every cap, thinking with the cue (391, 439 and 495 of 600) is still below plain direct mode with no cue (441, 491 and 535). Also, the late close and the early close do about the same (391 vs 377), so the reserve finding is only shown in direct mode so far.

One small thing: the abstract still gives +20.20 on GPQA-Diamond without the guessing baseline. Only about 5 points of that are above random letters, so the abstract should say it.

I am raising Soundness to 3, since the reporting issues are fixed. I am keeping my rating at 2 because the contribution is still narrow.

Reviewer_6BmgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper tests whether inserting a late final-answer cue helps models return correct answers within a fixed output limit. Cue and control runs share the same generated prefix, and forced tokens count toward the limit. Gains repeat on MATH and GPQA-Diamond with two model families. Comparing with continuation to the same limit also changes which answer reserve performs best.

Critical strengths
  1. The paired design is well controlled. The cue and baseline share exactly the same output before the intervention, and both receive the same total allowance. Forced tokens and discarded work are counted. This makes the comparison more informative than comparing separately generated responses or allowing extra answer tokens outside the limit.
  2. The delivery gain is supported by several replications, including unused MATH questions, a second model family, different limits, and GPQA-Diamond. The registered follow-ups and prospective reserve test strengthen the evidence. The analysis also correctly treats questions, rather than their three sampled responses, as the units of statistical inference.
Critical weaknesses
  1. The methodological contribution is incremental. Sun et al. (2025) already interrupt generation near an output limit, insert a concluding message, and compare against generation continued to that limit. The main additions here are exact shared prefixes, stricter token accounting, a delivery endpoint, and a detailed reserve study. These are useful refinements, but the core intervention is established. The accounting identities clarify the comparisons, but mainly partition outcomes rather than provide a new method for choosing when to intervene.
  2. The headline comparison with the closest prior cue changes both the cue and the space available for answering. The budget-charged Sun variant leaves only 11 answer tokens, compared with 128 for the proposed cue. The authors acknowledge this and test the older message at their earlier insertion point. That control closes much of the gap and even exceeds the proposed cue’s count with Qwen at 4096 tokens. The original protocol’s advance notice is also omitted to preserve shared prefixes. These controls support late answer forcing, but leave the improvement over the prior procedure unclear. Compare both cues across the same answer-reserve choices and include the full prior prompting procedure under the same total limit.
Questions for authors
  1. Can the late-cue comparison be repeated in the models’ native thinking mode? The existing thinking and forced-close controls are useful, but close thinking at an earlier fixed position. A late close-and-answer cue would test whether the same benefit holds when the intervention is applied near the limit of a thinking-mode response.
  2. The GPQA results show that forcing an answer increases both correct and incorrect deliveries. Using the reported correct, wrong, and missing-answer counts, under what penalties for wrong answers does the cue still outperform direct continuation? This would clarify when the delivery improvement is useful in applications where withholding an answer is preferable to giving a wrong one.
  3. For presentation, please group Section 6 into delivery gains, comparisons with prior methods, and reserve selection. It currently switches repeatedly between models, cue variants, reserves, scoring rules, and registered studies. The abstract could focus on fewer results and leave secondary counts to the tables.
3 comments
AuthorsOpenReview ↗

Response to Reviewer 6Bmg

Thank you for the review. We ran the comparison you asked for. It uses the 200 MATH questions of the paper's first repeat study, three new sampling streams and one 2048-token cap that includes forced tokens, on Qwen3.5-9B and the larger Qwen3.8-27B. Each model has 13 arms of 600 runs: continuation without a cue; our cue at answer reserves of 64, 128, 256 and 512 tokens; the Sun et al. message at the same four reserves; and the message at our cue's four positions.

"Compare both cues across the same answer-reserve choices" At equal answer reserve we find no delivery difference between the two cues. Cue minus message is +0.33, +0.83, -0.33 and +0.50 points for 9B and -0.33, +1.00, +0.17 and +0.50 for 27B, and every 95% interval lies within [-1.67, +2.17]. The earlier comparison (+3.50) set 11 against 128 answer tokens, and the revised paper now leads with the equal-reserve comparison. At equal reserve the 14-token message starts 10 tokens before our 4-token cue, so this compares two policies with the same reserve. With the message at our cue's position, which leaves it 10 fewer answer tokens, the differences range from -0.50 to +1.17 and every interval includes zero. Against continuation, at a 128-token reserve, the cue gains +8.00 [5.33, 10.83] and the message +7.17 [4.67, 10.00] for 9B, and +5.33 [3.33, 7.50] and +4.33 [2.33, 6.50] for 27B. The revised PDF reports this in Sec. 6.2, Table 5 and Appendix Table 36.

These numbers score all 600 runs per arm with the paper's numeric-tolerant scorer. The first study excluded two questions whose official solutions contain more than one boxed value, which left 594 runs. On those 594 runs the cue gains +8.08 [5.39, 10.94] for 9B and +5.39 [3.37, 7.58] for 27B, and the cue-minus-message differences change by about 0.01 points.

On including the full prior prompting procedure: the advance notice changes every token before the cue, so it cannot share a prefix with continuation. The shared-prefix comparison therefore uses their message at matched answer reserves (Sec. 6.2, Table 5).

"The methodological contribution is incremental" Interrupting near the limit and comparing with generation run to that limit is Sun et al.'s, and Sec. 2 says so. Our comparison shares the prefix and charges every token to one cap, and this adds two things. First, only runs still generating at the cue can change, which we check token by token. In the new experiment every run whose outcome changes is such a run, for both models, so their share bounds the gain (Prop. 2). Second, the best reserve depends on the comparison: against the stopped stream a 512-token reserve looks best in three MATH settings, while against continuation to the same cap a short reserve delivers more. In two 1024-token MATH conditions a 128-token reserve delivers more than a 512-token one by +14.33 [10.67, 18.33] and +10.33 [6.50, 14.33] points. In the new experiment 128 minus 512 is +3.67 [1.83, 5.67] for 9B and +3.00 [1.67, 4.50] for 27B.

"late-cue comparison ... in the models' native thinking mode" The paper's thinking arms close thinking early: forcing the close-and-answer suffix at token 1024 gains +45.79 points [40.07, 51.52] over uninterrupted thinking at the same 2048 cap. The late version you describe, a close-and-answer cue 134 tokens before caps of 2048, 4096 and 8192 applied only while thinking is open, is in our follow-up comment below for Qwen3.5-9B.

On penalties for wrong answers: with utility 1 for a correct answer, -λ for a wrong one and 0 for none, the GPQA-Diamond counts are 213 correct, 21 wrong and 360 none for continuation and 333, 184 and 77 for the cue, out of 594 runs. The difference is (120 - 163λ)/594, which is positive for λ < 120/163 ≈ 0.736. At λ = 1/3 the gain is +11.05 points. On MATH in the new experiment the thresholds are 48/73 ≈ 0.6575 for 9B and 32/67 ≈ 0.4776 for 27B. Sec. 6.1 of the revised PDF reports this.

On the organisation: the revised Section 6 has three parts, 6.1 delivery gains, 6.2 comparisons with prior procedures and 6.3 reserve selection, and the abstract is shortened accordingly.

AuthorsOpenReview ↗

Follow-up for Reviewer 6Bmg: a late cue in native thinking mode

A late cue also helps in native thinking mode, at every cap we tried. This is the experiment you asked for, on Qwen3.5-9B with the same 200 MATH questions and three sampling streams as the new experiment of our first reply, 600 runs per arm.

Thinking runs uninterrupted to 8192 tokens and is cut at each cap. The late cue is a six-token close-and-answer cue forced 134 tokens before the cap, only in runs whose thinking is still open there. Counts are correct deliveries out of 600, gains are in points, and brackets are 95% question-level bootstrap intervals.

CapUninterrupted thinkingWith the late cueGainThinking open at the cue
204890391+50.17 [44.50, 55.67]496 runs (82.7%)
4096206439+38.83 [33.67, 44.17]368 runs (61.3%)
8192360495+22.50 [17.83, 27.33]216 runs (36.0%)

The cue turns 301, 233 and 135 of the open runs into correct deliveries and harms none. The gain follows the share of runs whose thinking is still open at the cue, and every run whose outcome changes is such a run.

AuthorsOpenReview ↗

Update for Reviewer 6Bmg: the late thinking-mode cue on the larger Qwen3.8-27B

The late close-and-answer cue in native thinking mode also gains on the larger Qwen3.8-27B, at every cap. The revised PDF reports it in Sec. 6.1, Appendix J and Table 37; our earlier follow-up in this thread gave the same experiment for Qwen3.5-9B.

The setting is the same as in that follow-up: 200 MATH questions, three sampling streams, 600 runs per arm. Thinking runs uninterrupted to 8192 tokens and is cut at each cap. The late cue is a six-token close-and-answer cue forced 134 tokens before the cap, only in runs whose thinking is still open there. Counts are correct deliveries out of 600, gains are in points, and brackets are 95% question-level bootstrap intervals.

CapUninterrupted thinkingWith the late cueGainThinking open at the cueGains, harms
2048452517+10.83 [7.50, 14.50]110 runs (18%)65, 0
4096508543+5.83 [3.33, 8.67]56 runs (9%)35, 0
8192529552+3.83 [1.83, 6.17]30 runs (5%)23, 0

A gain is a run that the late cue delivers correctly and uninterrupted thinking does not; a harm is the reverse.

The benefit near the limit of a thinking-mode response. On both checkpoints every run whose outcome changes is one the cue touches. The net conversion of those runs, gains minus harms over touched runs, is 0.59, 0.62 and 0.77 at the three caps for Qwen3.8-27B and 0.61, 0.63 and 0.62 for Qwen3.5-9B. The overall gain equals the touched-run share times this net conversion, and the touched shares are 83%, 61% and 36% for Qwen3.5-9B and 18%, 9% and 5% for Qwen3.8-27B.

Early versus late close at the same cap (Appendix J.1, Table 37b). The late close delivers more correct answers than the paper's early close at token 1024 at every cap. The early close was generated with a 2048-token cap, and a run that ends within 2048 tokens reads the same at a larger cap. For Qwen3.8-27B no early-close run reaches the cap: the late cue delivers 517, 543 and 552 of 600 at caps 2048, 4096 and 8192 against 503 for the early close, +2.33 [0.33, 4.50], +6.67 [3.67, 10.00] and +8.17 [4.83, 11.83]. For Qwen3.5-9B the late cue delivers 391 against 377 at cap 2048, +2.33 [-2.17, 7.00]. On this checkpoint 17 of the 600 early-close runs reach the 2048-token cap, so at caps 4096 and 8192 the early-close count lies between 377 and 394: 377 leaves the 17 runs undelivered and 394 counts all 17 as correct. The late cue delivers 439 at cap 4096, +10.33 [5.33, 15.50] against 377 and +7.50 [2.50, 12.50] against 394, and 495 at cap 8192, +19.67 [14.83, 24.67] against 377 and +16.83 [12.00, 21.83] against 394. These intervals are descriptive.

The late cue generates 1846, 3404 and 5468 tokens per run on average at the three caps for Qwen3.5-9B and 793, 1085 and 1374 for Qwen3.8-27B; the early close generates at least 1064 for Qwen3.5-9B and 577 for Qwen3.8-27B. The comparison is between the two closing positions at the same cap.

Reviewer_Bxo8OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies answer-delivery interventions in reasoning models under strict generation budgets. The authors compare direct continuation with a late final-answer cue while preserving the same sampled prefix and accounting for all injected tokens within a shared output budget. Experiments on MATH and GPQA-Diamond demonstrate consistent improvements in correctly delivered answers across two model families and several token budgets. The paper further examines answer-reserve allocation, alternative cues, budget-aware prompts, and restart strategies. A central contribution is showing that conclusions about answer interventions depend critically on whether the baseline is stopped at the intervention point or allowed to continue to the same total budget.

Critical strengths
  1. Rigorous experimental controls for inference-time interventions.

The shared-prefix comparisons and explicit accounting of generated, injected, and discarded tokens address important confounding factors in evaluating reasoning-time interventions. In particular, the distinction between stopped-prefix and equal-total-budget comparisons is methodologically important and can change conclusions about the effectiveness of an intervention.

  1. Strong empirical validation and reproducibility.

The paper includes preregistered experiments, unused evaluation questions, multiple model and budget configurations, and a prospective study of answer-reserve selection. The improvements remain positive under alternative answer readouts and scorer audits. These controls make the central empirical findings convincing within the studied settings.

Critical weaknesses
  1. The relationship to existing research on answer-phase behavior and control could be better established.

Prior work has already investigated how reasoning budgets affect answer generation and how answer-stage interventions can improve accuracy and efficiency. For example, Contrastive Thinking Decoding (ICLR 2026 submission) analyzes thinking-answer disagreement, observes that constrained thinking can shift additional reasoning into the answer phase, and studies training-free answer-stage control under reasoning budgets.

The present work differs methodologically: it introduces a fixed late answer cue and emphasizes shared-prefix comparisons, fully charged output budgets, and correct answer delivery, whereas CTD uses contrastive logit correction to improve answer alignment with a given thinking trace. Nevertheless, the overlap in motivation and empirical phenomena deserves discussion. More clearly distinguishing answer-delivery failures from thinking-answer disagreement, and specifying which conclusions are new relative to prior answer-phase analyses, would strengthen the paper's positioning.

  1. The practical scope remains relatively narrow.

The experiments focus on two relatively small model families and tasks with short, well-defined final answers. Although improvements persist under relaxed answer-extraction rules, it remains uncertain whether the intervention provides comparable benefits for stronger reasoning models, structured outputs, or deployment settings in which answers can be recovered without requiring natural EOS. The reported execution-time comparisons also do not directly establish end-to-end latency improvements in a deployed serving system.

Questions for authors
  • Relationship to prior answer-phase research. Could the authors discuss how their findings relate to Contrastive Thinking Decoding and similar work on budgeted reasoning and answer-stage control? In particular, what new conclusions does the proposed late-cue analysis establish beyond prior observations about reasoning spilling into the answer phase and answer generation being sensitive to the available thinking budget?

  • Nature of the improvements. To what extent do the observed gains represent successful delivery of answers already supported by the preceding reasoning, as opposed to correction or additional reasoning induced by the final-answer cue? Clarifying this distinction would help connect the results to previous studies of thinking-answer alignment.

  • Generalization and practical deployment. How does the intervention behave for stronger reasoning models with lower output-cap hit rates, and what additional overhead would arise when native mid-generation cue insertion is unavailable?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Bxo8

Thank you for the assessment and for pointing us to Contrastive Thinking Decoding.

"how their findings relate to Contrastive Thinking Decoding and similar work" The revised Sec. 2 discusses CTD. As you describe it, CTD treats thinking-answer disagreement: an answer is produced, and a contrastive logit correction aligns it with a given thinking trace. We study non-delivery: 137 of 138 capped direct outputs on unused MATH questions and 357 of 359 on GPQA-Diamond end without a complete final box. We leave the logits untouched and force four tokens into the stream. CTD's observation that constrained thinking shifts reasoning into the answer phase is related to what our capped runs show, and we credit CTD for it. The late-cue analysis adds three results. Under a shared prefix and one fully charged cap, only runs still generating at the cue can change, so their share bounds the gain (Prop. 2). The preferred answer reserve depends on whether the comparison stops the stream or continues it to the same cap: in two 1024-token MATH conditions a 128-token reserve delivers more than a 512-token one by +14.33 [10.67, 18.33] and +10.33 [6.50, 14.33] points. And at equal cap, continuing the direct stream beats restarting in thinking mode by 8.25 points [4.71, 11.95].

On delivery as opposed to correction: the paper's claim is improved delivery. The cue rescues 309 of the 915 still-generating MATH runs that continuation does not deliver correctly, and in 189 of them the correct value is already in the prefix. In the new experiment described next, all 49 gains for Qwen3.5-9B and all 33 for Qwen3.8-27B come from a continuation run that reached the cap without a complete final box. In 45 and 31 of them the capped output has no correct answer even under permissive extraction.

"stronger reasoning models with lower output-cap hit rates" We ran the same paired design on Qwen3.8-27B and Qwen3.5-9B, with the 200 MATH questions of the paper's first repeat study, three new sampling streams and the 2048-token cap. Continuation reaches the cap in 107 of 600 runs (17.8%) for 27B against 131 (21.8%) for 9B. The cue gains +5.33 points [3.33, 7.50] for 27B, with 33 gains and one harm, against +8.00 [5.33, 10.83] for 9B. Every run whose outcome changes was still generating at the cue. That share is 20.2% for 27B against 23.5%, and the converted fraction is .26 against .34. Larger caps for Qwen3.5-9B are in our follow-up comment below.

"deployment settings in which answers can be recovered without requiring natural EOS" The gain stays. In the paper's studies, dropping the EOS requirement, reading the last box anywhere, or using a permissive read-out keeps every reported cue-minus-continuation interval above zero (appendix table of contract read-outs). The Read-outs column of Table 7 in the revised PDF gives the range of point estimates per setting, for example 6.67 to 7.33 points in one setting. In the new experiment the permissive read-out rises from 450 to 492 of 600 for 9B and from 470 to 501 for 27B.

On structured outputs: the reported experiments use boxed final answers (MATH) and option letters (GPQA-Diamond).

On latency and overhead: we report batch-policy time. With native insertion the cue costs four forced tokens, and the policy generates between 0.8% and 6.1% fewer tokens across the settings of Table 7. Without native insertion, the policy needs a second call that prefills the prompt and the generated prefix once, and only for runs still generating at the cue: 20.2% of runs in the 27B MATH condition, 23.5% for 9B, and 63.1% (375 of 594) on GPQA-Diamond with Qwen. The revised scope paragraph states this two-call accounting.

AuthorsOpenReview ↗

Follow-up for Reviewer Bxo8: larger caps and lower presence penalties

On settings with lower output-cap hit rates, we ran the same checkpoint with larger caps and with lower presence penalties. The cue keeps a positive gain at every cap up to 8192 and at every penalty. Both experiments use Qwen3.5-9B on the same 200 MATH questions and three sampling streams as the new experiment of our first reply, with 600 runs per arm. Counts are correct deliveries out of 600, gains are in points, and brackets are 95% question-level bootstrap intervals.

Larger caps (direct mode, penalty 2). Continuation is one stream run to 16384 tokens and cut at each cap. The cue is placed 132 tokens before each cap.

CapContinuation: correct, mean tokensCue: correct, mean tokensRuns still generating at the cueGain
2048441, 1027487, 1001143+7.67 [5.17, 10.33]
4096491, 1369520, 135680+4.83 [3.00, 7.00]
8192535, 1709544, 170525+1.50 [0.67, 2.50]

At cap 16384 every run has finished before the cue position, so the cue is never applied and both arms deliver 549 of 600 with 1775 tokens. The gain is bounded by the number of runs still generating at the cue (Prop. 2), and the cue generates no more tokens than continuation at any cap. The cue at cap 2048 reaches 487 correct deliveries with 1001 mean generated tokens, where continuation needs cap 4096 and 1369 tokens to reach 491.

Presence penalty (direct mode, cap 2048). The cap-hit rate changes little with the penalty. Without the cue, continuation reaches the cap in 114 of 600 runs (19.0%) at penalty 0, 124 (20.7%) at penalty 1 and 131 (21.8%) at penalty 2, and 122, 132 and 141 runs are still generating at the cue. The cue gains +5.67 [3.67, 8.00] at penalty 0, with 36 gains and 2 harms, +8.00 [5.17, 11.17] at penalty 1, with 49 and 1, and +8.00 [5.33, 10.83] at penalty 2. Every run whose outcome changes was still generating at the cue.

AuthorsOpenReview ↗

Update for Reviewer Bxo8: lower penalties, larger caps and thinking mode on Qwen3.8-27B

On stronger reasoning models with lower output-cap hit rates: the cue also gains on the larger Qwen3.8-27B, at lower presence penalties, at larger caps and in thinking mode. The revised PDF reports these results in Sec. 6.1, Appendix J and Table 37; our earlier follow-up in this thread gave the direct-mode results for Qwen3.5-9B. The setting is the same: 200 MATH questions, three sampling streams, 600 runs per arm. Counts are correct deliveries out of 600, gains are in points, and brackets are 95% question-level bootstrap intervals.

Presence penalty (direct mode, cap 2048). Without the cue, continuation reaches the cap in 107 of 600 runs at penalty 0 and 114 at penalty 1, and 113 and 120 runs are still generating at the cue. Continuation and the cue deliver 469 and 515 at penalty 0, a gain of +7.67 [5.00, 10.67] with 46 gains and 0 harms, and 467 and 504 at penalty 1, a gain of +6.17 [3.83, 8.67] with 37 gains and 0 harms. At the penalty of 2 used elsewhere the gain is +5.33 [3.33, 7.50].

Larger caps (direct mode, penalty 2). Continuation is one stream run to 8192 tokens and cut at each cap. The cue is placed 132 tokens before each cap.

CapContinuation: correct, mean tokensCue: correct, mean tokensRuns still generating at the cueGains, harmsGain
2048468, 980500, 95812133, 1+5.33 [3.33, 7.50]
4096518, 1233530, 12244813, 1+2.00 [0.67, 3.50]
8192541, 1393543, 1391102, 0+0.33 [0.00, 0.83]

The gain follows the number of runs that reach the cue (Prop. 2): the net conversion of those runs is 0.26, 0.25 and 0.20 at the three caps. On average the cue generates no more tokens than continuation at any cap.

Native thinking mode. Thinking runs uninterrupted to 8192 tokens and is cut at each cap. The late cue is a six-token close-and-answer cue forced 134 tokens before the cap, only in runs whose thinking is still open there.

CapUninterrupted thinkingWith the late cueGainThinking open at the cueGains, harms
2048452517+10.83 [7.50, 14.50]110 runs65, 0
4096508543+5.83 [3.33, 8.67]56 runs35, 0
8192529552+3.83 [1.83, 6.17]30 runs23, 0

The net conversion of the runs the cue touches is 0.59, 0.62 and 0.77.

Overhead without native insertion. Every run whose outcome changes is one the cue touches. In direct mode, a serving stack without mid-generation insertion needs two calls and one extra prefill (Sec. 7), and the second call is needed only for the runs still generating at the cue: 121, 48 and 10 of 600 at caps 2048, 4096 and 8192, and 113 and 120 at penalties 0 and 1. In thinking mode the cue touches 110, 56 and 30 of 600 runs at the three caps (18%, 9% and 5%).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.