cs.CLSep 30, 2026

Settle: Learning When to Stop Reasoning

Authors: Ryan Brown, Zihao Fu, Chris Russell

Organizations: Oxford Internet Institute, University of Oxford · Department of Linguistics and Modern Languages, The Chinese University of Hong Kong

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Reasoning models often continue generating after their answers have settled. Settle learns when to stop from answer stability in completed traces. It trains the existing end-of-reasoning token while keeping other predictions close to the base model, and requires only ordinary decoding at inference. On MATH-500 with Qwen3-4B, Settle reduces token count by 40% with a 0.5-percentage-point decrease in accuracy. It gains 6.16 percentage points over supervised fine-tuning on the same traces shortened at their first stable answer, at nearly identical token counts. Its stopping score predicts whether a correct answer will remain correct. Settle extends the accuracy-token-count Pareto frontier of the evaluated stopping methods.

Figures & tables

Figure 1: From answer readouts to stopping targets. A recorded Qwen3-4B trace solves a pie-chart problem whose answer is 107∘107^{\circ} . (a) Prompting a saved prefix gives a short answer without changing the recorded continuation. (b) A first-correct-answer rule would stop at 1,235 tokens, but that answer changes later. Settle instead labels the stable suffix: from 1,573 tokens onward, every sampled answer agrees with the final readout. These boundaries receive stop targets; earlier boundaries receive continue targets. Ellipses mark omitted readouts; the final readout defines agreement and is not a training target.
Figure 2: Settle improves on imitation and extends the Pareto frontier. Qwen3-4B, MATH-500, 16k; three-run means ±\pm sample standard deviations. (a) β\beta controls SFT regularization. (b) The dashed line joins nondominated means. Natural accuracy; token counts include DEER checks. Settle uses λ=0.2\lambda=0.2 . Baselines: shortest-correct SFT ( Munkhbat et al., 2025 ) , LSTM and token adjustment ( Liu & Wang, 2025 ) , Halt Vector ( Jayabahu & Adeleke, 2026 ) , and DEER ( Yang et al., 2025b ) . Settings and scores appear in Appendix B .
Training procedureReference weightAccuracy (%)Token count
Stability-truncated SFTβ=0\beta=087.45±2.5687.45\pm 2.562,345±752,345\pm 75
Stability-truncated SFTβ=0.2\beta=0.288.25±2.4888.25\pm 2.482,564±922,564\pm 92
Stability-truncated SFTβ=1.0\beta=1.090.50±1.9490.50\pm 1.943,130±1223,130\pm 122
Stability-truncated SFTβ=5.0\beta=5.092.96±0.5092.96\pm 0.503,884±1213,884\pm 121
Settle, w=0.25w=0.25λ=0\lambda=00.03±0.060.03\pm 0.066,289±6,6426,289\pm 6,642
Settle, w=0.25w=0.25λ=0.2\lambda=0.293.61±0.2493.61\pm 0.242,355±622,355\pm 62
Table 1: Reference regularization within each training procedure. MATH-500, Qwen3-4B, 16k; three-run means ±\pm sample standard deviations. Accuracy grades naturally generated final answers; token counts cover ordinary generation. SFT uses mean losses and Settle uses sums, so β\beta and λ\lambda have different scales. Other settings stay fixed within each procedure (Appendix A ).
λ\lambdaNatural accuracy (%)Forced accuracy (%)Token count
0.0593.04 ±\pm 0.1593.07 ±\pm 0.182,057 ±\pm 23
0.1092.99 ±\pm 0.2393.01 ±\pm 0.202,167 ±\pm 28
0.2093.19 ±\pm 0.3393.31 ±\pm 0.312,338 ±\pm 45
0.4093.98 ±\pm 0.4894.14 ±\pm 0.502,695 ±\pm 73
Table 2: Sensitivity to Settle’s reference weight. MATH-500, Qwen3-4B, 16k; w=0.25w=0.25 and all other training settings fixed. Means ±\pm sample standard deviations across three fits per coefficient. The sweep includes three new runs at the default λ=0.20\lambda=0.20 . Token counts cover ordinary generation.
ModelCapBaseSettleChange [95% interval]Token countSaved %
Qwen3-4B8k92.9292.79-0.13 [-0.95, +0.69]4,172 / 2,65036.5
Qwen3-4B16k95.2394.73-0.50 [-1.27, +0.27]4,829 / 2,90439.9
Qwen3-8B16k95.6795.45-0.22 [-0.94, +0.50]5,112 / 3,13538.7
Table 3: MATH-500 results with agreement labels, w=0.10w=0.10 . Forced accuracy in percent; changes in points with paired 95% intervals. Tokens are base / Settle means, including additional answer tokens. Each row averages three runs per model. “8k” and “16k” denote 8,192 and 16,384.
Figure 3: Stable correct answers often precede completion. MATH-500; 500 separately sampled traces per model, with an 8,192-token cap. Left: fraction with an early stable correct answer. Right: median first stable position as a fraction of trace length, among those traces. Appendix E gives full results and a repeated Qwen3-4B sample.
ScoreOverallSame token countCorrect now, same token count
Base native score0.532 [0.508, 0.556]0.507 [0.484, 0.530]0.495 [0.445, 0.547]
Elapsed tokens0.618 [0.583, 0.655]0.500†0.500^{\dagger}0.500†0.500^{\dagger}
Time-only spline0.618 [0.583, 0.654]0.500†0.500^{\dagger}0.500†0.500^{\dagger}
Settle, w=0.10w=0.100.831 [0.811, 0.851]0.807 [0.782, 0.831]0.695 [0.624, 0.764]
Settle, w=0.25w=0.250.838 [0.819, 0.857]0.818 [0.793, 0.842]0.706 [0.635, 0.774]
Table 4: Predicting answer persistence beyond elapsed length and present correctness. Overall and conditional AUROC [95% interval] on fixed Qwen3-4B prefixes. Overall uses 17,899 boundaries; matching token counts uses 12,106. The final column also requires the present readout to be correct, leaving 2,455 eligible boundaries. Intervals resample whole traces; each weight uses one agreement-trained model. †\dagger A token-count predictor assigns the same score to equal-length prefixes, giving conditional AUROC 0.5. Appendix C defines the estimands and reports all matching ranges.
BenchmarkAccuracy (%)Token countSaved (%)
MATH, separate sample96.33 / 96.523,803 / 2,17942.7
GSM8K95.1 / 94.52,053 / 1,16743.2
OlympiadBench66.6 / 70.49,024 / 5,75036.3
GPQA-Diamond55.6 / 55.98,565 / 7,43313.2
AMC2393.4 / 90.87,617 / 4,90335.6
Minerva48.2 / 44.56,536 / 3,32749.1
Table 5: Token savings transfer across tasks. Qwen3-4B, w=0.10w=0.10 , 16k cap. Forced accuracy and token pairs are base / Settle. All rows average three trained models; MATH uses three base runs and the other tasks one. Sample sizes and protocols appear in Appendices F and G .
Appendix figures & tables27 assets

Supplementary material from the paper’s appendix.

Appendix

Training procedureNatural %Forced %Token count
SFT: stability-truncated, epoch 180.3080.442,250
SFT: stability-truncated, epoch 287.4587.532,345
Settle, w=0.10w=0.1094.3894.732,904
Settle, w=0.25w=0.2593.6193.712,355
Appendix
Table 6: SFT on stability-truncated solutions. MATH-500 at 16k, four responses per problem. Accuracy excludes the two overlapping problems; token means include all 500. All rows average three independently trained models. Epochs are selected on development data.
SFT β\betaNatural accuracy changeForced accuracy changeTokens saved (%)
Settle w=0.10w=0.10
0+6.93 [+5.52, +8.43]+6.93\,[+5.52,\,+8.43]+7.20 [+5.79, +8.68]+7.20\,[+5.79,\,+8.68]−23.81-23.81
0.2+6.12 [+4.77, +7.56]+6.12\,[+4.77,\,+7.56]+6.21 [+4.84, +7.65]+6.21\,[+4.84,\,+7.65]−13.26-13.26
1.0+3.88 [+2.71, +5.10]+3.88\,[+2.71,\,+5.10]+3.77 [+2.58, +5.00]+3.77\,[+2.58,\,+5.00]+7.22+7.22
5.0+1.42 [+0.42, +2.49]+1.42\,[+0.42,\,+2.49]+0.87 [−0.13, +1.92]+0.87\,[-0.13,\,+1.92]+25.25+25.25
Settle w=0.25w=0.25
Appendix
Table 7: Both Settle settings versus every SFT reference weight. Settle-minus-SFT accuracy differences in percentage points, with paired 95% problem-bootstrap intervals. Token savings compare ordinary-generation means; negative values mean Settle uses more tokens. All comparisons use three trained models per setting and the interval procedure in Appendix D .
SupervisionRunsNatural %Forced %Token count
Agreement, w=0.10w=0.10394.3894.732,904
Reference, w=0.10w=0.10393.9694.382,961
Reference, w=0.25w=0.25393.5693.692,408
Agreement, w=0.25w=0.25393.6193.712,356
Base cut at 2,560135.8980.772,314
Shuffled labels186.7586.801,296
Appendix
Table 8: Training controls on MATH-500 at 16k. Accuracy uses the same scoring and overlap exclusions as the main evaluation. Tokens include additional answer tokens under forced scoring. Reference labels use known correct answers; agreement labels use the model’s own answers.
Figure 4: Learning the stop action versus shorter budgets and demonstrations. Labels on the base curve give maximum token budgets. Every point uses forced-answer accuracy and generated-token count, with the same scoring and overlap exclusions. Each base cap uses one run; Settle and both SFT controls show three-run means with sample-standard-deviation bars at a 16k cap. SFT on stability-truncated solutions uses the development-selected setting; the arrow gives its forced-accuracy difference from Settle at w=0.25w=0.25 . The base curve uses a separate evaluation from Table 3 ; lines connect its evaluated caps.
Maximum tokensForced accuracy %Mean token count
2,04878.011,950
3,07283.842,624
4,09687.703,115
6,14491.623,781
8,19293.474,198
16,38495.734,869
Appendix
Table 9: Qwen3-4B fixed-budget curve on MATH-500. Each row is one evaluation run.
Maximum tokensTraining seedNatural accuracy %Token count
8,192086.451,767
8,192186.601,684
16,384088.151,923
Appendix
Table 10: Ending-only SFT on MATH-500. Accuracy is natural, with no forced answer at the cap.
TaskPositive supervisionNatural %Forced %Token count
MATH-500First stable boundary93.4293.932,969
 All stable boundaries94.1894.582,890
Appendix
Table 11: Placement of positive stopping supervision, Qwen3-4B. Positive loss weight per trace, negative targets, and KL positions are identical. One training seed; four responses per MATH-500 problem, with evaluation seed 42 and a 16k cap. Token counts include prompted answers for capped responses.
MethodNatural (%) ↑\uparrowForced (%) ↑\uparrowToken count ↓\downarrow
(a) Training objectives and reference anchoring
Base ( Yang et al., 2025a )93.89±0.5293.89\pm 0.5295.23±0.2395.23\pm 0.234,828±284,828\pm 28
SFT: shortest correct ( Munkhbat et al., 2025 )94.13±0.3394.13\pm 0.3395.30±0.14\mathbf{95.30}\pm 0.144,493±704,493\pm 70
SFT: stability-truncated87.45±2.5687.45\pm 2.5687.53±2.5887.53\pm 2.582,345±75†\mathbf{2,345}\pm 75^{\dagger}
SFT: stability-truncated, β=0.2\beta=0.288.25±2.4888.25\pm 2.4888.52±2.4988.52\pm 2.492,564±922,564\pm 92
SFT: stability-truncated, β=1.0\beta=1.090.50±1.9490.50\pm 1.9490.96±1.8390.96\pm 1.833,130±1223,130\pm 122
Appendix
Table 12: Learning when to stop: training controls and existing stopping methods. MATH-500, Qwen3-4B, 16k cap; three-run means ±\pm sample standard deviations. Natural accuracy grades generated final answers; forced accuracy also prompts budget-limited responses to answer. Token counts cover ordinary generation before additional answer forcing, including intermediate checks for DEER. Bold marks the best mean within each panel. Unregularized SFT and controller settings are selected on development data; regularization controls keep other training settings fixed (Appendix A ).
θ\thetaTrialsρ\rhoMATH %Token countExtra
0.95100.8094.833,688118
0.90100.8094.783,530115
0.85100.8094.833,561115
0.80100.8094.883,475115
0.70100.8094.633,480114
0.50100.8094.933,470114
Appendix
Table 13: DEER settings on Qwen3-4B at 16k. θ\theta is the confidence threshold, ρ\rho the reasoning-budget fraction, and trials the checking parameter defined above. Accuracy is natural; token counts include checks. Extra counts trial answers, added trial prompts, and discarded generated tokens on MATH-500. Each row reports one run; replicated means appear in Table 12 . MATH-500 uses the main overlap exclusions.
Settle weightNatural differenceForced differenceTokens saved (%)
DEER θ=0.50, ρ=0.5\theta=0.50,\ \rho=0.5
0.10+0.42 [−0.18, +1.07]+0.42\,[-0.18,\,+1.07]+0.72 [+0.08, +1.42]+0.72\,[+0.08,\,+1.42]−3.88-3.88
0.25−0.35 [−1.07, +0.38]-0.35\,[-1.07,\,+0.38]−0.30 [−1.00, +0.42]-0.30\,[-1.00,\,+0.42]+15.73+15.73
DEER θ=0.95, ρ=0.25\theta=0.95,\ \rho=0.25
0.10+1.66 [+0.82, +2.56]+1.66\,[+0.82,\,+2.56]+1.97 [+1.09, +2.93]+1.97\,[+1.09,\,+2.93]−9.26-9.26
0.25+0.89 [−0.03, +1.86]+0.89\,[-0.03,\,+1.86]+0.95 [+0.03, +1.94]+0.95\,[+0.03,\,+1.94]+11.37+11.37
Appendix
Table 14: Settle versus development-selected DEER on MATH-500. Accuracy differences are Settle minus DEER in percentage points, with paired 95% intervals computed as in Appendix D . Token savings compare ordinary means, including DEER checks; negative values mean Settle uses more tokens.
Method and settingNatural %Forced %Token count
Base (repeat evaluation)77.9586.1110,130
LSTM controller, τ=0.5\tau=0.530.3830.382,169
LSTM controller, τ=0.9\tau=0.952.0852.083,703
LSTM controller, τ=0.95\tau=0.9556.6056.604,190
LSTM controller, τ=0.99\tau=0.9964.2464.415,510
LSTM controller, τ=0.995\tau=0.99568.0668.236,095
Appendix
Table 15: Full development sweeps. All settings use the same 144 problems, four responses each, and sampling seed 42. SFT and controller evaluations use training seed 0. SFT entries specify learning rate and epoch; every SFT entry uses solutions truncated at their first stable answer. Accuracy is in percent. Panel (a) counts naturally generated tokens; base is α=0\alpha=0 . Panel (b) includes DEER checks; θ\theta is the confidence threshold and ρ\rho the reasoning-budget fraction. The final column marks selection under each Settle development ceiling. These problems exclude the training pool.
 MATH-500
MethodAccuracyTokens
Base93.32 / 94.934,904
PUMA, released replay93.42 / 93.673,433
PUMA-SFT, common data93.07 / 93.172,803
Settle, w=0.10w=0.1094.18 / 94.582,890
Settle, w=0.25w=0.2593.93 / 93.982,331
Appendix
Table 16: PUMA comparisons with Qwen3-4B. Accuracy pairs are natural / forced percentages. Four responses per MATH-500 problem, evaluation seed 42, and one training seed per learned method. MATH accuracy excludes the two training overlaps; token means cover all problems and include checks and prompted answers for capped responses. Base uses PUMA’s source generations.
Elapsed tokensBoundariesTime-only predictorSettle w=0.10w=0.10Settle w=0.25w=0.25
[0,1024)[0,1024)6,7980.680 [0.657, 0.703]0.842 [0.821, 0.863]0.853 [0.831, 0.873]
[1024,2048)[1024,2048)4,6670.534 [0.508, 0.560]0.824 [0.787, 0.858]0.832 [0.795, 0.867]
[2048,4096)[2048,4096)3,6780.452 [0.421, 0.483]0.766 [0.719, 0.814]0.771 [0.723, 0.820]
[4096,8192)[4096,8192)2,0140.489 [0.442, 0.537]0.743 [0.687, 0.798]0.755 [0.700, 0.809]
[8192,16384)[8192,16384)7420.500 [0.500, 0.500]0.681 [0.572, 0.782]0.682 [0.575, 0.782]
Appendix
Table 17: Native-score prediction conditional on elapsed length. AUROC [paired whole-trace bootstrap 95% interval] on 498 nonoverlapping MATH-500 traces. Panel (a) ranks boundaries within each broad range; panel (b) ranks pairs in the same position group. Exact matching admits only identical token counts. Every row uses the same fitted time-only predictor and one agreement-trained model per Settle column. All four matching rules include all 498 traces.
ScoreOverallExact token count
Elapsed tokens0.552 [0.479, 0.630]0.500†0.500^{\dagger}
Time-only spline0.551 [0.477, 0.630]0.500†0.500^{\dagger}
Settle, w=0.10w=0.100.702 [0.645, 0.760]0.695 [0.624, 0.764]
Settle, w=0.25w=0.250.715 [0.659, 0.771]0.706 [0.635, 0.774]
Appendix
Table 18: Predicting whether a correct answer remains correct. AUROC [95% whole-trace bootstrap interval] restricted to currently correct readouts. Overall uses 12,013 boundaries; exact-token matching uses 2,455. †\dagger Token-count predictors assign the same score to equal-length prefixes, giving conditional AUROC 0.5.
ItemSetting
Problem pool50 DAPO-Math-17K problems
Rollouts20 per problem, 1,000 total
Rollout samplingTemperature 1.0, top- pp 1.0
Generation / training length limit8,192 / 8,192 tokens
Readout budget16 greedy tokens per read
LoRA rank / alpha / dropout32 / 64 / 0
Appendix
Table 19: Qwen3-4B training settings.
EvaluationTracesSettle %MedianAccuracy %Trace tokens
Q3-4B MATH-500 8k50087.80.2185.6 / 94.44,151 / 1,704
Q3-4B MATH-500 8k repeat50086.60.2183.0 / 93.04,185 / 1,724
Q3-8B MATH-500 8k50085.40.1982.0 / 93.24,396 / 1,776
Nemotron MATH-500 8k50083.60.3290.0 / 93.02,937 / 1,535
Q3-4B MATH-500 16k50087.60.2094.0 / 95.44,809 / 2,065
Appendix
Table 20: When answers stabilize in sampled reasoning traces. Settle is the percentage with an early stable-correct answer; median is the first such position as a fraction of trace length among those that settle. Accuracy and trace-token pairs are ordinary completion / retrospective stopping. “Repeat” denotes a separate Qwen3-4B MATH-500 sample.
CapBaseSettleChange [95% interval]Token countSaved %
8k95.4595.87+0.42 [-0.68, +1.55]3,493 / 2,04541.5
16k96.3396.52+0.18 [-0.83, +1.22]3,803 / 2,17942.7
Appendix
Table 21: Separate 500-problem MATH test sample, Qwen3-4B. Forced accuracy, paired 95% intervals, and base / Settle token means, averaged over three runs.
ModelSetNaturalForcedForced change [95% CI]Saved %
Qwen3-4BMATH93.89 / 94.3895.23 / 94.73-0.50 [-1.27, +0.27]39.9
Qwen3-8BMATH93.71 / 94.6695.67 / 95.45-0.22 [-0.94, +0.50]38.7
Nemotron-Nano-8BMATH94.61 / 92.9994.81 / 93.06-1.76 [-2.51, -1.04]33.3
R1-Distill-Qwen-7BMATH92.82 / 91.0393.56 / 91.62-1.94 [-2.68, -1.22]31.6
Appendix
Table 22: Results across four models. Accuracy pairs are base / Settle, in percent; forced-accuracy changes are percentage points. Each model uses three base runs and three trained runs. Token savings include additional answer tokens under forced scoring.
SetNNSSAccuracy %ChangeToken countSaved %
GSM8K1319195.1 / 94.5-0.72,053 / 1,16743.2
AMC2340893.4 / 90.8-2.67,617 / 4,90335.6
Minerva272148.2 / 44.5-3.76,536 / 3,32749.1
OlympiadBench674166.6 / 70.4+3.89,024 / 5,75036.3
GPQA-Diamond198155.6 / 55.9+0.38,565 / 7,43313.2
Appendix
Table 23: Additional benchmarks, Qwen3-4B at 16k. Forced accuracy and token pairs are base / Settle. NN denotes problems and SS responses per problem. Results average three trained models against one base run.
MethodForced accuracy (%)
Base85.55
Settle, w=0.01w=0.0186.13
Settle, w=0.025†w=0.025^{\dagger}84.57
Settle, w=0.05w=0.0584.18
Settle, w=0.10w=0.1080.47
Appendix
Table 24: Development accuracy for Qwen3-4B stopping-pressure calibration. One training seed per weight and four responses per development problem, with evaluation seed 42 and a 16k cap. †\dagger The development-selected weight.
 MATH-500AMC23
MethodAccuracyToken countAccuracyToken count
Base93.32 / 94.934,90487.81 / 92.817,507
Settle, w=0.01w=0.0193.22 / 94.834,78990.62 / 94.067,392
Settle, w=0.025†w=0.025^{\dagger}93.98 / 95.134,41089.38 / 91.886,896
Settle, w=0.05w=0.0594.58 / 95.283,64388.44 / 90.625,897
Settle, w=0.10w=0.1094.18 / 94.582,89089.06 / 89.694,822
Appendix
Table 25: The same Qwen3-4B weight sweep on MATH-500 and AMC23. Accuracy pairs are natural / forced percentages. Four responses per MATH problem and eight per AMC problem, evaluation seed 42, and one training seed per weight. MATH base uses the source generations from Table 16 ; all rows use the shared prompts, 16k cap, and scoring. †\dagger The weight selected on hard development problems.
PoolLabelsNatural accuracy (%)Forced accuracy (%)Token count
MATH-500
OriginalAgreement94.21±0.1594.21\pm 0.1594.66±0.1494.66\pm 0.142,943±492{,}943\pm 49
OriginalReference94.31±0.2894.31\pm 0.2894.85±0.1694.85\pm 0.162,961±1092{,}961\pm 109
BroaderAgreement94.01±0.2194.01\pm 0.2194.38±0.1794.38\pm 0.173,073±733{,}073\pm 73
BroaderReference94.08±0.4694.08\pm 0.4694.98±0.3994.98\pm 0.393,169±1643{,}169\pm 164
Appendix
Table 26: Training-pool and label-rule comparison on Qwen3-4B, with w=0.10w=0.10 throughout. Means ±\pm sample standard deviations across three training runs. Four responses per MATH problem, with a 16k cap. Token counts include prompted answers for capped responses.
MethodBatchTotal time (s)Median (s)Accuracy %Response tokens
Base11999.713.25924,426
Base32323.7–944,214
Base100160.5–944,365
Settle11165.27.66912,602
Settle32301.9–942,621
Settle100120.0–902,670
Appendix
Table 27: Inference time for 100 MATH problems on one H100. Total time covers all requests; median time per request is reported only for batch size 1. Accuracy is natural accuracy on these requests. Token counts here cover responses only; the accuracy–cost comparisons above also count the additional work of DEER’s intermediate checks.
MethodSeedMATH-500
  Accuracy (%)Token count
Qwen3-4B
Base4294.33 / 95.284,816
Base4394.03 / 95.434,809
Base4493.32 / 94.984,862
Settle094.63 / 94.882,847
Appendix
Table 28: Individual runs across models. Accuracy pairs are natural / forced; token means include additional answer tokens under forced scoring.
Method and settingRunNatural (%)Forced (%)Tokens
Base (repeat evaluation)093.2294.834,905
 193.7895.234,853
 293.3294.784,874
Halt Vector093.7894.534,116
 193.8394.634,137
 293.9394.684,244
Appendix
Table 29: Individual stopping-baseline runs. Token counts cover ordinary generation. Epoch-2 unregularized truncated SFT appears in Table , with both token conventions.
Sampling seedNatural (%)Forced (%)Token count
   OrdinaryWith forced answers
Truncated SFT, β=0\beta=0
4289.3689.412,347.72,347.8
4384.5484.592,268.72,268.8
4488.4588.602,419.52,419.6
Truncated SFT, β=0.2\beta=0.2
Appendix
Table 30: Individual reference-regularization and DEER runs. Ordinary token counts precede additional answer forcing; the final column includes it. DEER counts include intermediate checks.
λ\lambdaTraining seedSampling seedNatural (%)Forced (%)Token count
0.0504292.8792.872,040.1
0.0514393.1793.222,048.3
0.0524493.0793.122,082.6
0.1004293.1293.122,135.9
0.1014392.7292.772,177.0
0.1024493.1293.122,188.3
Appendix
Table 31: Individual Settle reference-weight sensitivity runs. Qwen3-4B, w=0.25w=0.25 . These twelve fits use the same source traces and training settings as the main models, including three new runs at λ=0.20\lambda=0.20 . Token counts cover ordinary generation.
More in this topic

Early Stopping for Language Model Reasoning

See all 22 papers

Explore similar work

CardsList

  1. When Does Learning to Stop Help? A Cost-Aware Study of Early Exits in Reasoning Models

    Jun 29, 2026Zhe Dong, Fang Qin, Manish ShahEfficient Language Model ReasoningCost-Aware Inference

    University of Maine at Presque Isle · Stanford University · Independent Researcher

  2. Same Beliefs, Different Policies: Learning Reusable Representations for Early Stopping in LLM Reasoning

    Oct 21, 2026Jingwei Ni, Dmitry Abulkhanov, Vladislav Smirnov +5Early Stopping for Language Model ReasoningEarly Stopping

    ETH Zürich · MBZUAI · University of Zürich +1

  3. </think> Doesn't Stop Reasoning: Analysis of Spurious CoT Termination

    Sep 3, 2026Seunghee Koh, Sungjae Choi, Minchan Kwon +2CoT ReasoningEfficient Language Model Reasoning

    Korea Advanced Institute of Science and Technology, South Korea