Beyond Reasoning Gains: Mitigating General-Capability Forgetting in Large Reasoning Models

Researchers have proposed a new training strategy called RECAP to prevent large reasoning models from losing foundational skills while they improve at tasks like math and multimodal reasoning, according to a paper posted on arXiv [1]. Reinforcement learning with verifiable rewards, or RLVR, has become a standard post-training method for language and vision-language models, delivering strong gains in reasoning [1]. But the technique carries a risk: models can suffer “capability regression,” forgetting core abilities such as perception and faithfulness after extended training without safeguards [1]. The authors empirically confirmed the problem in open-source reasoning models [1]. Existing countermeasures include regularization terms like KL divergence, which help keep a model from drifting too far from its base version. Those terms are computed on the current task, however, and do not guarantee that broader knowledge is preserved [1]. Experience replay across different domains is another common tactic, but deciding how much weight to give each objective is difficult because of cross-domain heterogeneity [1]. The new method, called RECAP — short for Replay-Enhanced CApability Preservation — mixes general-capability data back into the RLVR objective and then dynamically reweights objectives based on their convergence rate and instability [4]. The mechanism operates online, using short-horizon signals to shift training emphasis away from saturated objectives and toward underperforming or volatile ones [1]. The approach is end-to-end, requires no auxiliary models, and can be dropped into existing RLVR pipelines without heavy tuning [4]. The researchers tested RECAP on Qwen2.5-VL-3B and Qwen2.5-VL-7B, two vision-language models developed by Alibaba Cloud and released under open-source licenses [1][11]. The Qwen2.5-VL family, which also includes a 72-billion-parameter version, was introduced as a successor to Qwen2-VL and is available on Hugging Face and ModelScope [10]. In one experiment, training solely on reasoning rewards degraded perception performance on the 7B model by 7 percent [5]. The RECAP strategy not only preserved general capabilities but also improved reasoning by allowing more flexible trade-offs among in-task rewards [1]. The paper was submitted on 24 October 2025 and revised on 18 June 2026 [1]. A version of the work also appears in the ACL Anthology under the title “Beyond Reasoning Gains: Mitigating General-Capability Forgetting in Large Reasoning Models” [3].