A mixed supervised fine-tuning (SFT) strategy came out on top in a controlled comparison of ways to train reasoning-focused AI. After all five methods went through the same reinforcement learning with verifiable rewards (RLVR) stage, Mixed SFT posted an average of 67.4 across six in-domain benchmarks. NTR averaged 64.2 and NSR 63.6. NTR and NSR also used over 60 times more GPU hours than SFT on the same no-CoT data.
The result is the headline finding of an arXiv preprint dated 24 Aug 2026. Its central question was narrow: does next-chunk reasoning make better use of no-CoT data, meaning solutions without an exposed reasoning trail, than SFT when the base model, no-CoT corpus and RLVR budget are held the same? The comparison also included Sequential SFT and Reasoning SFT, alongside NTR, or next-token reasoning, and NSR, or next-sentence reasoning.
Five routes, one common test
All experiments used Qwen3-30B-A3B-Base as the base model. The retained training material included 152K correct long-CoT trajectories, about 1.95 billion tokens, and 421K no-CoT solutions, about 0.53 billion tokens. Long-CoT examples contain an extended reasoning trail, while no-CoT examples do not expose that trail. Keeping those ingredients fixed made the comparison about the training strategy rather than a changing model or corpus.
Before RLVR, each model took one of five routes: Reasoning SFT, Sequential SFT, Mixed SFT, NTR or NSR. NTR and NSR started from Reasoning SFT. The final stage was the same GRPO RLVR procedure for all five, using rule-based exact-match rewards on DAPO-Math-17K. That shared finish meant the reported differences were measured after a common downstream training stage.
Models were scored both before and after RLVR. The primary post-RLVR comparison covered six in-domain and three out-of-domain reasoning benchmarks, using avg@32, avg@4 and pass@1 reporting. In plain terms, the paper focused mainly on where each route ended after RLVR while also recording where it started.
The ranking changed after RLVR
That timing changed the picture dramatically for Mixed SFT. Before RLVR, it had the lowest accuracy at 27.5. After RLVR, it had the highest at 61.1, a gain of 33.7 points. The other methods began between 45.9 and 48.8 and improved by between 3.3 and 10.3 points.
On the six in-domain benchmarks, Mixed SFT's post-RLVR average of 67.4 was 3.1 points above NTR and 3.7 points above NSR. Their corresponding averages were 64.2 and 63.6.
The lead also appeared on each of the three out-of-domain tests. Mixed SFT recorded 9.24 on HLE, 60.98 on GPQA-Diamond and 75.84 on MMLU-Pro, the highest reported scores on all three and above both NTR and NSR.
Why the expensive route struggled
Compute told a second part of the story. NTR and NSR consumed over 60 times the GPU hours used by SFT on the same no-CoT data. That gap is important alongside the accuracy result: Mixed SFT's in-domain average was higher than both next-chunk methods despite their much larger GPU-hour use.
The paper then examined what the next-chunk objectives were actually learning. In a random sample of 2,048 high-entropy tokens, selected where the model's next choice was relatively uncertain, reasoning accuracy rose from 0.29 to 0.55 while entropy stayed almost flat. Accuracy on no-reasoning cases stayed around 0.48.
That pattern is consistent with the authors' reading that many apparently difficult targets were locally predictable without an explicit reasoning reconstruction. The trace examples added a warning sign: NTR outputs became shorter and more deterministic, converging on repetitive, template-like local completion. NSR showed a similar tendency at the sentence level.
The study also tested entropy and response-length controls. In entropy-controlled NTR, generation entropy stayed near 0.52 and response length around 640 tokens, yet the method reached a slightly lower post-RLVR ceiling than vanilla NTR on every benchmark.
Another ablation inserted NTR or NSR between Mixed SFT and RLVR. The extra stage left final performance almost unchanged and slightly lower on several tasks. In this setup, adding next-chunk reasoning after the mixed route did not produce an observed post-RLVR gain.
What survived the final stage
Finally, the study checked how much no-CoT performance remained after RLVR. Its retention probe used 100 no-CoT problems, generated eight responses per problem and reported avg@8. Mixed SFT plus RLVR scored 68.63, compared with 59.19 for Sequential SFT plus RLVR.
Taken together, the results favor Mixed SFT within this particular pipeline: it had the highest reported in-domain average, the strongest scores on the three out-of-domain tests and the lower GPU-hour burden relative to NTR and NSR. The comparison remains tied to the ingredients tested here: Qwen3-30B-A3B-Base, the named long-CoT and no-CoT corpora, the shared RLVR stage and the specified benchmark set. The document is an arXiv preprint, version 1, dated 24 Aug 2026.
Paper data and sources
Original title: Is Next-Chunk Reasoning RL Really Better than SFT? Revisiting Training Strategies under no-CoT Data
Authors: Yinhao Tang, Youqing Fang, Yanan Sun et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: Not available
Original paper · Full text