Revisiting Complete Reasoning Traces for Post-Training
Jaehui Hwang, Sangdoo Yun, Byeongho Heo, Dongyoon Han
cs.CL
Quick verdict
Claims trimming the middle of reasoning traces (E-SFT) improves SFT/RL/distillation; gains are consistent but small and lack statistical rigor.
Summary
The paper asks whether LLMs actually benefit from post-training on complete reasoning traces. In a pilot on s1K-1.1 (Qwen2.5-32B-Instruct, Qwen3-8B-Base), keeping only the beginning and ending segments ('Both') beats full trajectories (75.19 vs 73.51 average over AIME24/GPQA-D/MATH for Qwen2.5-32B). Three analyses — attention inspection, segment-removal regeneration, and answer perplexity under segment replacement — indicate intermediate steps are weakly attended, causally less important, and carry little answer-prediction signal. The authors propose Endpoint-based SFT (E-SFT): remove ~20% of middle tokens, keeping first/last steps (n=100/200). E-SFT matches or beats standard SFT across Qwen2.5-32B, Qwen3-4B/8B, and Llama models on s1K-1.1 and OpenThoughts3-100K (e.g., 67.50 vs 65.78 for Qwen3-8B on OpenThoughts3-100K), outperforms random-, similarity-, LLM-compression-, and perplexity-based filtering baselines, and extends to masking in GRPO (+1.2 average on instruct Qwen3-1.7B; +8.7 on the base variant) and on-policy distillation (+2.2 with step-level masks). Main limitations: gains are small relative to benchmark noise (AIME24 has 30 problems, three sampled evaluations), no statistical significance testing is reported, and the mechanistic claim that models internally infer missing steps remains inferential.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Full reasoning trajectories provide limited benefit in SFT: retaining only prefix+suffix segments outperforms complete traces (Table 1: 75.19 vs 73.51; Table 3).
- Intermediate trajectory tokens are weakly attended, causally dispensable for answer generation, and replacing them even lowers answer perplexity (§4.1, Figs 2-3, Table 2).
- E-SFT (~20% middle-token removal) matches or beats standard SFT across five models and two datasets, and outperforms seven filtering-based trajectory-reduction baselines (Tables 3, 5, 6, B, C).
- Masking the middle 20% of rollouts also improves GRPO and on-policy distillation, with step-level masks beating token-level ones (Tables 7, 8).
Strengths
Triangulated evidence for the redundancy claim
§4.1 combines three independent probes: averaged + layer-wise attention over trajectories (Fig 2), Geva-style segment-removal regeneration scored by ROUGE-L/Jaccard/BLEU (Fig 3), and answer perplexity under segment replacement (Table 2). The step delimiter itself is validated (B.2: 33.5%/28.0% semantic transitions across \n\n vs 2.5%/1.0% within steps).
Simple, deployable recipe validated unusually broadly
E-SFT (remove ~20% middle tokens, keep first/last n=100/200 steps) is tested on 5+ models across Qwen and Llama families, two datasets (s1K-1.1, OpenThoughts3-100K), and beats 7+ filtering baselines including resource-intensive ones (Tables 5, 6, B, C).
Generality beyond SFT with granularity ablation
Honest scoping of when the method works
Supporting hygiene beyond the headline
Weaknesses
No statistical rigor around small deltas
Mechanistic account is inferential
Confounded training-perplexity argument
Suspiciously catastrophic LLM-compression baselines
Post-hoc hyperparameter heuristic
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets the standard practice of SFT-ing smaller LLMs on machine-generated reasoning traces (s1K-1.1, OpenThoughts3) and claims that complete traces are unnecessary: intermediate segments are redundant, and keeping only trajectory endpoints — E-SFT, removing ~20% of middle tokens (first/last n=100 steps on s1K-1.1, n=200 on OpenThoughts3) — trains better models. It further claims the principle transfers to GRPO (masking the middle 20% from the gradient, Tables 7) and on-policy distillation (Table 8). The central mechanistic claim is that models "internally infer" missing steps from internal knowledge plus known endpoints. Related-work positioning (§2) against short-m@k (Hassid et al.), ThinkPrune, CoThink, and token-level compression (Yuan et al.) is fair: the differentiator is a segment-level, scorer-free recipe at training time.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is E-SFT?
Endpoint-based SFT is a training recipe from this paper: before fine-tuning on machine-generated reasoning traces, delete the middle portion of each trace (~20% of tokens, e.g., keep the first and last 100 steps on s1K-1.1) and train only on the beginning and ending segments. The paper reports this matches or beats training on the full traces.
Does removing the middle of a chain-of-thought hurt accuracy?
In this paper's experiments across Qwen2.5-32B, Qwen3-4B/8B, and Llama models on s1K-1.1 and OpenThoughts3-100K, no — average scores over AIME24/GPQA-D/MATH were equal or higher than full-trace SFT (e.g., 75.19 vs 73.51 for Qwen2.5-32B on s1K-1.1; 67.50 vs 65.78 for Qwen3-8B on OpenThoughts3-100K), though individual gains are small and unaccompanied by significance tests.
Does the idea work for RL training (GRPO), not just SFT?
The authors report that masking the middle 20% of the rollout from the gradient (tokens stay in context and are still reward-scored, but get no per-token gradient) improves GRPO — +8.7 average for Qwen3-1.7B-Base and +1.2 for instruct Qwen3-1.7B on DAPO-17k — and on-policy distillation (+2.2 average with step-level masks, +0.7 token-level).
When does E-SFT not help?
Appendix A reports two boundaries: gains appear only when the base model already performs above roughly 60-65 on general benchmarks for s1K-1.1 (the threshold drops with the 100x larger OpenThoughts3-100K), and excessive truncation consistently degrades performance — the paper suggests removing ~20% of middle tokens as a practical default.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Table 2 as extracted is internally inconsistent: the PPL row (1.56/1.52/1.58/1.54) cannot be reconciled arithmetically with the ΔPPL row (+0.02/−0.02/+0.04) or the caption's stated direction; the review cites this table only qualitatively and flags the alignment as unreliable.
- Figure content (Figs 2-6, A-D) is only partially recoverable from text extraction; figure-based judgments rely on captions and surrounding prose, and exact axis values are not quoted.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.