Self-rewarding correction for mathematical reasoning (2502.19613v1)

Abstract: We study self-rewarding reasoning LLMs, which can simultaneously generate step-by-step reasoning and evaluate the correctness of their outputs during the inference time-without external feedback. This integrated approach allows a single model to independently guide its reasoning process, offering computational advantages for model deployment. We particularly focus on the representative task of self-correction, where models autonomously detect errors in their responses, revise outputs, and decide when to terminate iterative refinement loops. To enable this, we propose a two-staged algorithmic framework for constructing self-rewarding reasoning models using only self-generated data. In the first stage, we employ sequential rejection sampling to synthesize long chain-of-thought trajectories that incorporate both self-rewarding and self-correction mechanisms. Fine-tuning models on these curated data allows them to learn the patterns of self-rewarding and self-correction. In the second stage, we further enhance the models' ability to assess response accuracy and refine outputs through reinforcement learning with rule-based signals. Experiments with Llama-3 and Qwen-2.5 demonstrate that our approach surpasses intrinsic self-correction capabilities and achieves performance comparable to systems that rely on external reward models.

The paper introduces a novel self-rewarding reasoning framework for LLMs that integrates generation and evaluation into a single model. This framework aims to enhance the self-correction capabilities of LLMs in mathematical reasoning tasks, reducing computational overhead compared to approaches relying on external reward models.

The key contributions of the paper are:

• A self-rewarding reasoning framework integrating the generator and reward model into a single LLM, enabling autonomous reasoning, evaluation, and correction.
• A two-stage algorithmic framework for self-correction in mathematical reasoning, relying only on self-generated data. The first stage uses sequential rejection sampling to construct long chain-of-thought (CoT) trajectories encoding self-rewarding and self-correction behaviors. The second stage enhances these behaviors through reinforcement learning with rule-based signals.
• Empirical validation demonstrating that self-rewarding correction significantly outperforms intrinsic self-correction.

The self-rewarding reasoning process is formulated as a multi-turn Markov Decision Process (MDP). An LLM generates an initial reasoning attempt $a_1 \sim \pi_1(\cdot | s_1)$ given a prompt $s_1 = x \in \mathcal{X}$ from a distribution $\mathcal{D}_0$, where $\pi$ is the LLM. It then self-rewards its response by generating an evaluation $y_1 \sim \pi_1(\cdot | s_1, a_1)$. If the model assesses its answer as correct ($y_1$ = [VERIFY] correct), the generation stops. Otherwise, the LLM generates a refined response and evaluation $(a_2, y_2) \sim \pi_2(\cdot | s_2)$, conditioned on the updated state $s_2 = (s_1, a_1, y_1)$. The self-refinement continues until the model produces a self-evaluation $y_h$ assessing the answer as correct.

$y_1 \sim \pi_1(\cdot | s_1, a_1)$

$s_2 = (s_1, a_1, y_1)$

where:

• $s_1$ is the initial prompt.
• $a_1$ is the initial reasoning attempt.
• $\pi_1$ is the LLM.
• $y_1$ is the self-rewarding evaluation.

The two-stage training framework consists of:

1. Self-rewarding instruction-following fine-tuning (IFT): An initial LLM $\pi_0$ is fine-tuned using demonstration data collected via sequential rejection sampling, resulting in an improved model $\pi_{\text{ref}}$ integrating self-rewarding reasoning abilities.
2. Reinforcement learning (RL) optimization: $\pi_{\text{ref}}$ is further refined using RL, leveraging it as the reference model. This stage enhances the model's ability to assess correctness and refine responses.

The self-rewarding signal is trained by token prediction, where models include reasoning in their evaluations and output specific tokens to indicate their evaluation results, such as "[VERIFY] correct" and "[VERIFY] wrong". Data collection uses a rejection sampling approach, generating self-correction trajectories and preserving desired ones. The process includes generating initial reasoning responses, sampling self-rewarding signals, and correction sampling. The LLMs are fine-tuned using a standard SFT pipeline to maximize:

$$\mathbb{E}_{\mathcal{D}_{IFT}} [\log P(y_1|x, a_1) + \log P(a_2|x, a_1, y_1)] + \mathbb{E}_{\mathcal{D}_{IFT}} [\log P(a_2|x, a_1, y_1)] + \mathbb{E}_{\mathcal{D}_{IFT}} [\log P(y_1|x, a_1)]$$

where:

• $x$ is the initial prompt.
• $a_1$ is the initial reasoning attempt.
• $y_1$ is the self-rewarding evaluation of the first turn.
• $a_2$ is the revised reasoning attempt.

For the RL stage, the paper considers both deep RL methods and direct alignment algorithms. A trajectory-wise reward function $u^*(\tau)$ is used for trajectory $\tau = (x, a_1, y_1, \dots, a_H, y_H)$, where $H$ is the horizon. The oracle reward $u^*(\tau) = r^*(x, a_H)$ is used, where $r^*$ is the ground-truth verifier. The KL-regularized objective is:

$$\max_{\pi} \mathbb{E}_{x \sim \mathcal{D}_0, a_1 \sim \pi_0(\cdot | x)} \mathbb{E}_{\tau \sim \pi(\cdot | x, a_1)} \sum_{h=1}^{H} r^*(\tau) - \eta D_{KL}(\pi_h (\cdot | s_h), \pi_{\text{ref}} (\cdot | s_h))$$

where:

• $\pi$ is the policy being optimized.
• $\pi_0$ is the initial LLM.
• $\pi_{\text{ref}}$ is the reference model.
• $r^*(\tau)$ is the trajectory-wise reward.
• $D_{KL}$ is the Kullback-Leibler divergence.
• $\eta$ is a regularization coefficient.

The paper also adopts Direct Preference Optimization (DPO) to solve the equation, using the multi-turn DPO (M-DPO) framework. The loss function $C_{\text{M-DPO}}(\theta)$ is:

$$- \mathbb{E}_{(\tau^w, \tau^l) \sim \mathcal{D}} \log \sigma \Big( \eta \big[ \log \frac{T_\theta(y_1^w | x, a_1)}{T_{\text{ref}}(y_1^w | x, a_1)} - \log \frac{T_\theta(y_1^l | x, a_1)}{T_{\text{ref}}(y_1^l | x, a_1)} + \sum_{h=1}^{H} \log \frac{T_\theta(a_h^w, y_h^w | s_h)}{T_{\text{ref}}(a_h^w, y_h^w | s_h)} - \log \frac{T_\theta(a_h^l, y_h^l | s_h)}{T_{\text{ref}}(a_h^l, y_h^l | s_h)} \big] \Big)$$

where:

• $\tau^w$ is the winning trajectory.
• $\tau^l$ is the losing trajectory.
• $T_\theta$ is the policy being optimized.
• $T_{\text{ref}}$ is the reference policy.
• $\sigma$ is the sigmoid function.
• $\eta$ is a regularization coefficient.

The models are evaluated on mathematical reasoning abilities using benchmarks including MATH500, OlympiadBench, and Minerva Math. Evaluation metrics include turn 1 accuracy, final accuracy, improvement in accuracy from the first attempt to the final answer ($\Delta(t_1, t_2)$), fraction of problems changed from incorrect to correct ($\Delta_{i \rightarrow c}(t_1, t_2)$), and fraction of problems changed from correct to incorrect ($\Delta_{c \rightarrow i}(t_1, t_2)$).

The main results demonstrate that intrinsic self-correction with prompting generally fails, while self-rewarding reasoning models significantly outperform existing baselines. For instance, on MATH500, self-rewarding IFT achieves $\Delta_{i \rightarrow c}$ = 5.0% and $\Delta_{c \rightarrow i}$ = 0.4%. Self-rewarding reasoning models also improve final accuracy compared to single-turn baselines. The paper finds that deep RL algorithms outperform direct alignment algorithms.

Further experiments were performed using a simplified two-turn conversation framework and Llama models, confirming the generality of the proposed framework. Ablation studies on data distribution show that the data composition in self-rewarding IFT influences the outcome supervised reward model (ORM) accuracy. The paper also investigates additional rule designs in RL training.

The paper concludes by highlighting the effectiveness of the self-rewarding reasoning framework in enhancing self-correction capabilities and computational efficiency. Future research directions include addressing the lower reward model accuracy compared to external ORMs, incorporating multi-turn RL methods, and extending the framework to step-wise correction.