RL on Incorrect Synthetic Data Scales the Efficiency of LLM Math Reasoning by Eight-Fold

Training on model-generated synthetic data is a promising approach for finetuning LLMs, but it remains unclear when it helps or hurts. In this paper, we investigate this question for math reasoning via an empirical study, followed by building a conceptual understanding of our observations. First, we find that while the typical approach of finetuning a model on synthetic correct or positive problem-solution pairs generated by capable models offers modest performance gains, sampling more correct solutions from the finetuned learner itself followed by subsequent fine-tuning on this self-generated data $\textbf{doubles}$ the efficiency of the same synthetic problems. At the same time, training on model-generated positives can amplify various spurious correlations, resulting in flat or even inverse scaling trends as the amount of data increases. Surprisingly, we find that several of these issues can be addressed if we also utilize negative responses, i.e., model-generated responses that are deemed incorrect by a final answer verifier. Crucially, these negatives must be constructed such that the training can appropriately recover the utility or advantage of each intermediate step in the negative response. With this per-step scheme, we are able to attain consistent gains over only positive data, attaining performance similar to amplifying the amount of synthetic data by $\mathbf{8 \times}$. We show that training on per-step negatives can help to unlearn spurious correlations in the positive data, and is equivalent to advantage-weighted reinforcement learning (RL), implying that it inherits robustness benefits of RL over imitating positive data alone.

Paper

Similar papers

Peer review

Reviewer iuJa8/10 · confidence 3/52024-07-09

Summary

There have been previous attempts at doing finetuning on positive synthetic examples to improve LM reasoning, but the performance gain from these attempts are generally quite limited (and possibly even negative). In this paper, the authors take a different approach which also accounts for negative examples and “critical” intermediate steps, improving upon previous approaches.

Strengths

I was really impressed by this paper – at least for me, it contains quite a lot of insights, and is substantially better than most papers on synthetic data that I’ve seen in the past. - I thought the idea of using negative data was fairly natural but also good, and I’m glad to see it implemented in practice - I found that the proposed link to model collapse seemed really interesting, which allegedly can occur if a model generating synthetic data has memorized the training data using a spurious feature. - I really liked the idea of looking at “critical steps”, and I thought the approach of relating this to credit assignment was impressive. More generally, I liked that the authors really made clear attempts to both quantify and explain their empirical observations. For example, I appreciated the section comparing the sample efficiency of self-generated and GPT-4/Gemini 1.5 Pro synthetic data, which does both.

Weaknesses

The biggest concrete thing that I felt was missing was discussion about computational costs. I think this is really important because I’d be interested in the scalability of approaches like the ones outlined in the paper. I think having some comparison of the overall FLOP required for finetuning (including the FLOP for synthetic data generation) would be very helpful. Another weakness is that in my understanding, the main approaches in the paper apply most directly to mathematics, since identifying critical steps involves checking results based on MC rollouts. There’s still a question about whether or not similar approaches can be applied to LMs in domains where outputs are harder to “verify” – though to be clear, I think it’s entirely reasonable that the authors focus on math in the context of this paper. Lines 197-199: This compares the scaling exponent for data for the MATH and GSM8K datasets with the exponent from the Chinchilla paper (on MassiveText). How do we know that this implies improving coverage over samples? Other possible reasons for the different exponent come to mind – e.g. different functional forms for the scaling law in finetuning and pretraining, use of different datasets, etc. As a minor comment, I think “It is predicted that we will run out of high-quality internet data by 2026” should be modified to be about human-generated public text data in particular, and it looks like the median year from source [53] is 2028 rather than 2026.

Questions

- Could the authors please provide some information about the computational costs of their approach? (perhaps comparing this for SFT, RFT, and per-step DPO) - Clarification for lines 197-199: How do we know that this implies improving coverage over samples, rather than the other reasons I mentioned? - Clarification for lines 203-204: Is it correct to interpret this as saying that “around 4% of sampled outputs from the SFT model are used for RFT”?

Rating

8

Confidence

3

Soundness

3

Presentation

3

Contribution

4

Limitations

I think the authors did a fair job discussing the limitations of their work. One minor weakness is that the paper is framed around when synthetic data improves LM reasoning, but the investigation itself is more narrowly focused on mathematics.

Reviewer 8A1v6/10 · confidence 4/52024-07-12

Summary

The paper investigates how to effectively leverage synthetic data to improve reasoning capability in the GSM8K and MATH datasets. The authors identify that sampling correct synthetic data from a fine-tuned model is more sample-efficient but comes with the risk of overfitting artifacts in the synthetic data. Instead, they propose a per-step DPO to leverage negative synthetic data to enhance sample efficiency.

Strengths

1. Synthetic data for reasoning is a highly important topic. 2. The idea is straightforward and effective. 3. The experiments are extensive, with comprehensive ablation studies.

Weaknesses

1. Missing Relevant References: The paper lacks references to relevant literature. Scaling laws on how synthetic data plateaus or performs slower than real data during pretraining have been observed and analyzed in several studies [1][2]. Additionally, the observation that self-generated data is more sample-efficient has been noted in research on images and machine translation [3][4][5]. 2. Unclear Per-Step DPO Definition: The per-step DPO is not fully defined. In Line 180, how is the “first pit” $\hat{y_c}$ determined? Does it involve looping over all intermediate steps from the beginning to find the smallest $c$? It would be helpful for the authors to provide an algorithmic outline in the appendix. In Theorem 6.1, the per-step DPO is trained with pairs (x, $[y_{1:i}, +y_{i+1}]$, $[y_{1:i}, -y_{i+1}]$), whereas Line 180 defines it as using $+y$ instead of $[y_{1:i}, +y_{i+1}]$. Which version is implemented in the experiments, and why? 3. Claims on Sample Efficiency: The improved performance on filtered $\mathcal{D}_{syn}$ from the fine-tuned model requires sampling 100 samples per question, and the per-step DPO samples require Monte Carlo rollouts for intermediate steps. This adds significant computational overhead during inference, much more than during the fine-tuning stage. A proper comparison of sample efficiency should include the inference compute spent generating these examples. Given the same computational resources, would DPO perform better as it does not require identifying the “first pit” while utilizing more synthetic data? 4. Explanation of Figure 7: The explanation of Figure 7 is unclear. How are the average Q-values calculated? What if a problem requires fewer than 8 steps in GSM8K? Will the average for step 8 be biased toward harder questions? ### Reference [1] Dohmatob, Elvis, et al. "A tale of tails: Model collapse as a change of scaling laws." arXiv preprint arXiv:2402.07043 (2024). [2] Fan, Lijie, et al. "Scaling laws of synthetic images for model training... for now." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024. [3] Mobahi, Hossein, Mehrdad Farajtabar, and Peter Bartlett. "Self-distillation amplifies regularization in hilbert space." Advances in Neural Information Processing Systems 33 (2020): 3351-3361. [4] Gu, Jiatao, et al. "Non-autoregressive neural machine translation." arXiv preprint arXiv:1711.02281 (2017). [5] Zhou, Chunting, Graham Neubig, and Jiatao Gu. "Understanding knowledge distillation in non-autoregressive machine translation." arXiv preprint arXiv:1911.02727 (2019).

Questions

1. Line 138: Why can we assume that the answers generated via this process are accurate? Is the model prompted with the correct solution? 2. Line 166: The sentence is not well-written. 3. DPO Suitability for Math: DPO is derived from the Bradley-Terry model, where preference follows a probabilistic distribution with softmax over reward values. For mathematical generation, however, correct answers should always be preferred over incorrect ones with probability 1. The framework, especially naive DPO, seems unsuitable for the math set, which might explain its weak performance. The KTO method, cited in the paper, proposes a different RLHF method that considers positive and negative distributions directly. How would the KTO method perform? Would the per-step version of KTO also outperform naive KTO? 4. Line 302: Why are the advantage values always non-positive? 5. Data Generation Protocols: It would be more self-contained to include critical data generation protocols in the paper rather than referencing “Training data for DPO is generated in the procedure outlined in [23]” in Appendix H at line 800. The reviewer will consider increase the score if some of the weakness and questions are addressed.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Limitations are discussed.

Reviewer 8A1v2024-08-11

I appreciate the response and the new experiments on KTO. However, I am not entirely convinced by the claim that "*per-step DPO is more computationally efficient than RFT in our results*" as the author is drawing 100 samples per prompt, which introduces the possibility of variance when drawing 10 samples per prompt while creating more prompts in RFT. This variant could potentially surpass the original RFT with 100 samples per prompt. The new experiments on the overfitting of RFT, however, demonstrate that given abundant compute resources, per-step DPO clearly outperforms. Consequently, I have increased my score.

Authorsrebuttal2024-08-12

Clarification on computational costs of per-step DPO and RFT

Thank you for your response and for increasing the score! We are not sure if we fully follow your comment on the “computational efficiency” of RFT vs. per-step DPO, and might be misunderstanding your proposed variant of RFT, and the precise notion of “variance” being referred to. Therefore we request you to kindly correct us if the following response does not address your comment. **Please let us know.** _We begin by belaboring that the goal of our work is not to propose “per-step DPO” as novel algorithm_, but more to highlight our conceptual framework of positive and negative synthetic data, where credit assignment with negative data can address issues caused by overfitting on spurious steps in positive synthetic data. This general point is also evidenced in our other comparisons: per-step RFT is better than RFT, and per-step KTO is better than KTO. **Sampling new prompts is substantially more costly in our setting than sampling new responses.** We apologize if this was not clear but our claim “per-step DPO is more computationally efficient than RFT in our results” is only meant to apply to the setting with the **same set of synthetic prompts**. This is because obtaining new synthetic prompts requires querying more-capable models (like GPT-4, Gemini 1.5 Pro) in all of our experiments, which would incur a much higher factor times the computational cost of the 7B models that we study (these models are more capable than Llama2-70B, so it is safe to say that their sizes are at least 10x larger, which means at least 10x the computational cost for sampling more questions). While we agree that in some uses cases where sampling new prompts using a 7B model might be sufficient, there can be variants of RFT with more prompts (and less solutions per prompt) which could be more computationally efficient, this is not true in our setting, and we are not aware of any work that samples synthetic math questions (of the same quality as GPT/Gemini) with smaller scale models. ___ **Empirical justification for the cost comparison given same set of prompts:** Within the context of the same set of prompts, more samples for RFT (e.g., 100) should reduce variance, it might lead to overfitting as we have shown in Section 5. Our results in Figure 1c (1-page PDF) compares the best configuration of RFT (which uses 4 diverse and correct samples out of the 100 samples) with per-step DPO. Note that these 4 out of 100 samples are selected based on the techniques prescribed by Yuan et al. [68] to obtain the most competitive RFT performance. For comparison, in our experiments on the 8k sized synthetic data which has 8k prompts, we got best performance from RFT by sampling 100 responses per prompt (total of **8k * 100 = 800k**) and selecting 4 diverse and correct responses per prompt from them. On the other hand, for per-step DPO we only need one positive and one negative trace. Even accounting for the advantage estimation run on every step in the negative trace, we only needed to draw **578k samples** from the SFT policy in our experiments (indicating that our calculation in the previous response was already overestimating the _actual_ cost of per-step DPO). ___ **Theoretical justification of the cost comparison given same set of prompts:** In general, when $p$ is the accuracy of the SFT policy, the number of samples needed to draw a single correct sample is a geometric distribution with mean $1/p$. Note, that we actually need $k$, way more than a single correct sample per prompt, to ensure there is enough diversity of solutions per prompt in RFT data (from these $k$ correct ones we selected 4 correct and diverse ones, as done in Yuan et al. [68]). From tail bounds on geometric distribution, to obtain at least $k$ positive samples per prompt, if we sample $\frac{k}{p} \cdot (1+\sqrt{\log(k/\delta)})$ samples from the SFT policy, then with probability at least $1- \delta$, we will have $k$ positive traces. On the other hand, per-step DPO requires only one positive and negative sample, obtaining which requires at most $\frac{2}{p} \cdot (1+ \sqrt{\log(2/\delta)})$ many samples with probability $\geq 1 - \delta$, when $p < 0.5$. For per-step DPO, we additionally require Monte Carlo advantage estimation, which requires sampling $M$ samples per step (with at most $L$ steps in the trace). Thus, the total number of samples (from SFT policy) needed to run per-step DPO is at most $\frac{2}{p} \cdot (1+ \sqrt{\log(2/\delta)}) + M \times L$. Note that for RFT we need about $k/p$, and since we follow Yuan et al. [68], we get the best performance of RFT when $k \approx 40$ at least, which puts the $k/p$ term in RFT around **70 to 100** for MATH and GSM8k. On the other hand the $M \times L$ in per-step DPO is **at most $50$** for both of these datasets in the experiments. Thus, even in the worst-case per-step DPO (which achieves 8x scaling) requires fewer samples than the best performing variant of RFT (which achieves 2x scaling).

Reviewer e4554/10 · confidence 3/52024-07-13

Summary

This paper investigates the impact of synthetic data on improving the reasoning capabilities of large language models (LLMs). The authors conduct an empirical study followed by a theoretical formalization to understand when and how synthetic data helps or hurts model performance on reasoning tasks. Key findings include:- (1) Finetuning on synthetic correct/positive problem-solution pairs offers modest gains, but sampling more correct solutions from the finetuned learner doubles sample efficiency. (2) Training solely on model-generated positives can amplify spurious correlations, leading to flat or inverse scaling trends as data increases. (3) Utilizing negative responses (incorrect model-generated responses) alongside positives, with appropriate credit assignment for intermediate steps, yields consistent gains over positive-only data. (4) The proposed per-step negative data approach is equivalent to advantage-weighted reinforcement learning (RL) and helps unlearn spurious correlations. The authors demonstrate that their per-step negative data approach improves sample efficiency by 8x compared to standard finetuning on positives. They also provide theoretical and empirical analyses to explain why credit assignment improves model generalization.

Strengths

- The work reveals important nuances in using synthetic data, particularly the value of leveraging negative examples and the importance of credit assignment for intermediate steps. - The findings offer concrete guidance for improving LLM training with synthetic data, potentially addressing data scarcity issues in language model development. - The authors conduct extensive experiments, including scaling studies and comparisons with various baselines, strengthening the validity of their claims.

Weaknesses

- While the focus on reasoning tasks is well-justified, it's unclear how well the findings generalize to other types of language tasks or domains. - The study primarily uses 7B parameter models (DeepSeek-Math-7B and LLama2-7B). It would be valuable to see if the results hold for larger or smaller model sizes. - The paper doesn't discuss the computational costs of their approach, which could be a significant factor in its practical application, especially for larger models or datasets. - While the authors compare their method to several baselines, it would be beneficial to see comparisons with more recent techniques in synthetic data generation or model finetuning. - The study doesn't address potential long-term consequences of training on synthetic data, such as potential drift or accumulation of errors over multiple generations of synthetic data.

Questions

None.

Rating

4

Confidence

3

Soundness

2

Presentation

2

Contribution

2

Limitations

Yes.

Reviewer jvk25/10 · confidence 4/52024-07-15

Summary

This paper studies the effect of synthetic data on improving the reasoning abilities of LLMs. The authors have compared with multiple approaches including SFT, RFT, DPO, per-step DPO, etc., and characterized the model performance w.r.t. different scales of synthetic data under each training regime. Several practical findings about improving sample efficiency are presented, including leveraging the finetuned model to sample for positive samples, utilizing per-step negative examples to alleviate the spurious correlations that might appear in the synthetic positive data, etc.

Strengths

- The paper studies an important problem. - The conceptual model of negative sample construction covered in Section 6.1 is an interesting read, as the authors formally express the synthetic data problem in the context of RL.

Weaknesses

- It seems that the authors wanted to cram lots of conclusions into a single paper. While I appreciate the incentive, the paper would make a much more academic read if more experimental results are clearly presented in tables. Currently, it seems that there are few quantitative results in the paper. This also led many specific concerns (see below). - In Lines 242-245, it reads `But, we find that even when RFT data is restricted to one solution per question, LLM trained on it outperforms SFT consistently by > 1%. Since verification is cheap, we can sample more solutions and also benefit from coverage.` Where is the evidence to this claim? In this claim, does the size of D_{syn} equal that of D_{SFT}^{+}? - It makes me confused when I compare Lines 224-244 and Lines 245-263. In Lines 224-244, the authors suggest that RFT outperforms SFT in terms of sample efficiency. But in Lines 245-263, the authors suggest that RFT data contain spurious correlations, and when manually raising the percentage of spurious correlations in RFT data, SFT would outperform RFT. Especially, in Lines 248-249, `This is unlike scaling of problems and solutions in SFT data (in Figure 2(a,b)).` Since RFT might even suffer from negative scaling because of spurious correlations, I don't think that RFT could be claimed to be `more sample-efficient`. - Also about the artificial spurious correlation amplification experiments in Lines 257-263: Does Figure 4 show the experimental results (there's no reference in the paragraph)? Does the `D_{π}^{+} spurious` correspond to the construction of `for each question in D_{syn} we sample “spurious steps” from π_{sft}`? Does this "100% spurious correlation injection" represent the authentic "spurious steps" ratio in RFT data (and what is the authentic "spurious steps" ratio in RFT data?)? Does D_{syn} contain no spurious correlation steps? As the RFT data scaling leads to negative performance, will those RFT data with a less percentage of spurious correlation steps lead to a not-so-negative performance? If this is true, can we just use the advantage defined in Eq.(3) as a stronger filtering rule and apply on RFT data? A more in-depth quantitative study is required here, so that the importance of the use of negative examples in Section 6 is better revealed. - Is the `per-step DPO` method used in the paper proposed in the cited [23] paper? If it is true, is it also true that no new method is proposed in Section 6? Note that it is fine to focus on analysis instead of proposing new techniques, but the author should make it more clear about this in the paper. It would also be helpful if the authors briefly introduce the per-step DPO method used in the paper in a background/methodology section. - Are the advantage in Eq.(3) and the per-step DPO method applicable to other types of reasoning tasks in addition to MATH and GSM8K? For example, ARC requires LLM reasoning (LLM often conducts this task by writing codes). Are the methods and the analysis applicable to ARC? - Lines 343-344: `As such, we needed to tune β in Equation 1 for DPO but could not fully avoid 344 performance degradation.` Where is the evidence to this claim? Experimental results with a sweep of β would be supportive. - Similarly, Lines 377-386 also require more experimental evidence. It would be helpful if quantitative observation is added to the claim like `When the initial advantage of a spurious step is incorrectly over-estimated, negative data algorithms up-weight the likelihood further. This only leads to further memorization. `, `leads to test-time model collapse`, and `On the other hand, when \tilde{π} = BoK(π_{sft}) for a higher value of K, the Monte-Carlo advantage estimator has a lower variance (and error). This discussion also justifies the choice of K=5`. - The paper lacks a Conclusion section, which I believe is because of the length limit of a submitted paper. But it is still better to add such a section and properly refactor the previous contents. I encourage the authors to address my concerns, and I would consider raising my score provided that the issues I raised are resolved properly. Others: Line 322: "instantation" --> "instantiation"

Questions

See the Weaknesses part.

Rating

5

Confidence

4

Soundness

2

Presentation

2

Contribution

3

Limitations

The authors have addressed the limitations.

Authorsrebuttal2024-08-10

Request for discussion

Dear reviewers, Apologies for bothering you! Since we are getting close to the end of the discussion period, we would be grateful and would sincerely appreciate if you could respond to our rebuttal, leaving us enough time to address any remaining questions. Thanks, Authors

Reviewer iuJa2024-08-11

Thanks for the really comprehensive response! I'll keep my official score at 8 - while I think the work is excellent, the description for a score of 9 requires the work to be "groundbreaking", which I think is a very high bar. However, I should emphasise that my main concerns have been addressed, and if it were possible I'd raise my score by 0.5 points to reflect this.

Authorsrebuttal2024-08-12

Thank you for your appreciation

Thank you immensely for your kind appreciation and very positive assessment of our work! To further improve our paper, we will incorporate the clarifications and discussion we provide in our responses. -Authors

Reviewer jvk22024-08-12

I appreciate the authors for adding the experiments. I request the authors to add these experiments into the main paper to support the claims, and also improve the presentation of the paper. I've raised the score.

Authorsrebuttal2024-08-12

Thank you for increasing your score! We will definitely add in these experiments and discussion, to the main paper and improve the presentation per the discussion above. Since there is still one more day, we are also wondering if there would be some other discussion or evidence that we can provide in this period to help improve your evaluation of our paper further. Please let us know. Thanks a lot!

Authorsrebuttal2024-08-13

Dear Reviewer, We apologize for bothering you, but since there is only one day left in the discussion period, we wanted to check in with you to see if our rebuttal addresses all outstanding concerns and have a chance to address any new ones. Thanks, Authors

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC