Beyond Reasoning Gains: Mitigating General Capabilities Forgetting in Large Reasoning Models

TL;DR

This paper introduces RECAP, a dynamic replay strategy that reweights objectives during training to prevent forgetting of foundational skills in large reasoning models, improving knowledge retention and reasoning performance.

Contribution

The paper proposes RECAP, an online reweighting method for experience replay that preserves general capabilities in large reasoning models during reinforcement learning with verifiable rewards.

Findings

RECAP effectively preserves core capabilities like perception and faithfulness.

RECAP improves reasoning performance and knowledge retention.

The method is simple, end-to-end, and adaptable to existing pipelines.

Abstract

Reinforcement learning with verifiable rewards (RLVR) has delivered impressive gains in mathematical and multimodal reasoning and has become a standard post-training paradigm for contemporary language and vision-language models. However, the RLVR recipe introduces a significant risk of capability regression, where models forget foundational skills after prolonged training without employing regularization strategies. We empirically confirm this concern, observing that open-source reasoning models suffer performance degradation on core capabilities such as perception and faithfulness. While imposing regularization terms like KL divergence can help prevent deviation from the base model, these terms are calculated on the current task, thus they do not guarantee broader knowledge. Meanwhile, commonly used experience replay across heterogeneous domains makes it nontrivial to decide how much training focus each objective should receive. To address this, we propose RECAP-a replay strategy with dynamic objective reweighting for general knowledge preservation. Our reweighting mechanism adapts in an online manner using short-horizon signals of convergence and instability, shifting the post-training focus away from saturated objectives and toward underperforming or volatile ones. Our method is end-to-end and readily applicable to existing RLVR pipelines without training additional models or heavy tuning. Extensive experiments on benchmarks based on Qwen2.5-VL-3B and Qwen2.5-VL-7B demonstrate the effectiveness of our method, which not only preserves general capabilities but also improves reasoning by enabling more flexible trade-offs among in-task rewards.