Enterprises deploying large language models for reasoning tasks face a hidden cost: the very training that sharpens mathematical and multimodal reasoning can erode core capabilities such as perception and faithfulness. A new paper on arXiv, authored by Phan, Hoang, Yang, Xianjun, Yao, Yuanshun, Zhang, Jingyu, Bi, Shengjie, Tang, Xiaocheng, Khabsa, Madian, Liu, Lijuan, and Lei, Deren, documents this phenomenon and introduces a mitigation strategy called RECAP.
The Problem: Capability Regression in RLVR
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, in which models forget foundational skills after prolonged training without employing regularization strategies. The authors empirically confirm this concern, observing that open-source reasoning models suffer performance degradation on core capabilities such as perception and faithfulness.
Standard approaches like imposing KL divergence regularization help prevent deviation from the base model, but these terms are computed on the current task and therefore do not guarantee preservation of broader knowledge. Commonly used experience replay across heterogeneous domains makes it nontrivial to decide how much training emphasis each objective should receive.
RECAP: Dynamic Objective Reweighting
To address the forgetting problem, the authors propose RECAP — a replay strategy with dynamic objective reweighting for general knowledge preservation. The reweighting mechanism adapts online using short-horizon signals of convergence and instability, shifting the post-training focus away from saturated objectives and toward underperforming or volatile ones. The method is end-to-end and readily applicable to existing RLVR pipelines without training additional models or heavy tuning.
The method was evaluated on benchmarks using Qwen2.5-VL-3B and Qwen2.5-VL-7B models. According to the paper, extensive experiments demonstrate the effectiveness of the method, which not only preserves general capabilities but also improves reasoning by enabling more flexible trade-offs among in-task rewards.
| Model | Preservation Effect | Reasoning Impact |
|---|---|---|
| Qwen2.5-VL-3B | Core capabilities (perception, faithfulness) retained | Reasoning maintained or improved |
| Qwen2.5-VL-7B | Core capabilities retained | Flexible reward trade-offs achieved |
Implications for Enterprise AI
For CTOs and technology leaders deploying large reasoning models in production, the risk of capability forgetting is critical. Models fine-tuned on domain-specific reasoning tasks may lose general skills like accurate perception or faithful output, which are essential for enterprise applications such as document analysis, compliance checking, or automated reporting. RECAP offers a plug-in approach to mitigate this without requiring additional model training or extensive hyperparameter tuning.
The paper highlights that while RLVR methods boost reasoning, they must be paired with regularization strategies to avoid unintended regressions. RECAP's dynamic reweighting provides an adaptive solution that balances multiple objectives in real time. This could enable enterprises to maintain model generality while achieving specialized reasoning improvements, reducing the need for separate models for different tasks.
As the field moves toward more capable reasoning models, mitigating general-capability forgetting becomes a key concern for reliable AI deployment. The authors' approach provides a practical, end-to-end solution that can be integrated into existing training pipelines, making it directly relevant for organizations leveraging open-source models like Qwen2.5-VL for both reasoning and general-purpose tasks.
