Construting Reverse Thinking: Developing Large Language Models' Reverse Thingking Ability
Xin Liu, Yunhai Li, Chunfu Jia, Ziliang Chen, Jisen Song
cs.AI
Quick verdict
Proposes backward-reasoning SFT+RL for LLMs; gains on math benchmarks with shorter chains, but small margins, self-built tests, no significance testing.
Summary
The paper argues that long chain-of-thought LLMs (GPT-o1/o3, DeepSeek-R1) default to forward reasoning and proposes a training pipeline to add backward reasoning as an alternative mode. Analyzing failures on 171 'backward questions' drawn from GSM8K, MATH-500 and AIME, the authors categorize errors into five types (solution-space coverage, computational mistakes, unverified assumptions, ignored constraints, length limits). They use DeepSeek-R1 to generate three types of backward questions (condition-swap, reverse-intermediate, error-diagnosis) with backward reasoning paths from MATH-lighteval and OpenR1-Math-220k, verified by R1 itself, split into easy/hard subsets, and fine-tune DeepSeek-R1-Distill-1.5B and Qwen3-1.7B/8B/14B via two-stage LoRA SFT followed by Reinforce++ RL with a BERT-based mode-classifier reward, cosine length-smoothed answer reward, and linearly decaying balanced sampling. Reported results: average accuracy gains over base models (e.g., 1.5B: 64.6→71.0; 14B: 87.4→87.9) with 24.6–37.5% shorter responses after RL, and adaptive mode selection (backward usage 12% on GSM8K vs 79% on their self-built BQ set). Limitations: no statistical significance testing, the BQ test set comes from the same generation pipeline as training data, reward weights are unreported, and no code is released.
Review complete
Assessment at a glance
5
risk flags identified
● high 2
● medium 2
● 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
- Long-CoT model failures on 171 identified backward questions fall into five categories, three of which persist even in GPT-5/GPT-o3 (Fig 1).
- Synthesized, R1-verified backward-reasoning data (condition-swap, reverse-intermediate, error-diagnosis) can teach small open models backward thinking via two-stage SFT.
- Adding Reinforce++ RL with a mode-classifier reward and cosine length-smoothed answer reward improves average accuracy (1.5B: 64.6→71.0; 14B: 87.4→87.9) while cutting response length 24.6–37.5%.
- The trained model autonomously selects reasoning mode per problem, using backward reasoning on 12% of GSM8K but 79% of the backward-question set BQ (Fig 4).
Strengths
Error analysis grounds the method
Fig 1's five-category taxonomy comes from statistics over 171 backward questions across ten model scales (1.5B–32B plus GPT-5/o3), and the three residual error types in frontier models (coverage, unverified assumptions, constraints) directly motivate backward reasoning.
Full pipeline with stage-wise ablation
Table 3 evaluates merged SFT, stage 1, stage 2 and RL for all four base models, isolating that RL contributes most of the gain (Qwen3-8b: 81.1–82.0 avg SFT-only vs 84.7 with RL; AIME drops to 48.3 under SFT stage 1 alone).
Accuracy and efficiency reported together
Adaptive mode selection is quantified
Multi-scale, multi-benchmark evaluation
Weaknesses
Single-run results without significance testing
Evaluation circularity in BQ and backward subsets
Closest prior never compared empirically
Method under-specification blocks reproduction
Judge-dependent answer scoring
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper proposes a training pipeline that adds backward (goal-to-condition) reasoning as an alternative mode to forward chain-of-thought in small open LLMs (DeepSeek-R1-Distill-1.5B, Qwen3-1.7B/8B/14B). The motivating analysis (§1, Fig 1) examines 171 'backward questions' identified in GSM8K, MATH-500 and AIME and categorizes model failures into five types: insufficient solution-space coverage, computational mistakes, unverified assumptions, ignored constraints, and max-response-length truncation; the first, third and fourth persist even in GPT-5/GPT-o3. Central claims: (1) backward-reasoning data can be synthesized at scale and verified automatically; (2) two-stage easy-to-hard SFT plus Reinforce++ RL with a mode-selection reward teaches models when to reason backward; (3) the method improves average accuracy while shortening responses (avg 64.6→71.0 at 1.5B, 87.4→87.9 at 14B; lengths −24.6% to −37.5% after RL, Table 3). Scope is exclusively mathematical reasoning; cross-domain transfer is deferred to future work.
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 backward (reverse) reasoning in LLMs?
Instead of deriving an answer step-by-step from the given conditions, the model starts from the target or conclusion and backtracks to the known conditions — for example, assuming a conclusion and working out the premises that must hold. This paper trains small open models (1.5B–14B) to choose between forward and backward modes depending on the problem.
How much does the method improve performance?
On GSM8K, MATH-500, AIME-2024 and a self-built 300-item backward-question set (BQ), the authors report average accuracy rising from 64.6 to 71.0 for their R1-Distill-1.5B variant and from 87.4 to 87.9 for the Qwen3-14B variant, while average response length falls by 24.6–37.5% after the RL stage. These are single-run results without significance testing.
Is the code or dataset available?
The paper does not mention a code or data release. It quotes the generation prompts verbatim and builds on public sources (MATH-lighteval, OpenR1-Math-220k), but key details — reward weights α and β, final dataset sizes, and the BERT discriminator's training data — are unreported.
Which models were fine-tuned and how?
DeepSeek-R1-Distill-1.5B and Qwen3-1.7B/8B/14B. Each gets two-stage LoRA supervised fine-tuning (easy backward-reasoning data first, then hard), followed by Reinforce++ reinforcement learning with a BERT-based mode-classifier reward, a cosine length-smoothed answer reward, and a linearly decaying balanced sampling strategy.
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.
- Figures 1–4 are not analyzable from the extracted text; only in-text numeric summaries (e.g., the Fig 4 percentages quoted in §4.3) are available.
- Equation 2 (format reward) is corrupted in extraction, so its exact form is inferred rather than verified.
- The metadata title contains typos ('Construting', 'Thingking') absent from the paper's own title block; this review refers to the corrected title.
- The metadata abstract is truncated mid-sentence, but the full abstract appears in the paper body.
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.