Beyond Reasoning Gains: Mitigating General-Capability Forgetting in Large Reasoning Models
Researchers propose RECAP, a dynamic reweighting strategy that preserves general AI capabilities while improving reasoning performance in large language models trained with reinforcement learning. The method addresses a critical problem where models forget foundational skills like perception and faithfulness during post-training optimization on reasoning tasks.
Reinforcement learning with verifiable rewards has become the dominant paradigm for training advanced reasoning models, enabling significant improvements in mathematical and multimodal reasoning capabilities. However, this specialized training introduces a fundamental trade-off: models optimizing intensively on narrow reasoning tasks experience measurable degradation in foundational capabilities including perception, faithfulness, and general knowledge retention. This capability forgetting represents a significant architectural vulnerability in modern AI systems, where narrow optimization directly undermines broader competence.
The RECAP framework addresses this through intelligent dynamic reweighting of training objectives. Rather than relying on static regularization terms like KL divergence—which preserve deviation from base models but don't guarantee broader knowledge retention—RECAP monitors short-horizon convergence signals across diverse domains. The system automatically shifts training emphasis away from saturated objectives toward underperforming or volatile ones, creating more balanced capability preservation. This approach requires no additional models or extensive hyperparameter tuning, making it readily deployable in existing RLVR pipelines.
For the AI development industry, this research has substantial implications. AI labs building reasoning-focused models now have evidence that naive post-training approaches create capability regressions, creating pressure to adopt preservation strategies before deploying systems. The accessibility of RECAP—being end-to-end implementable without additional computational overhead—may accelerate adoption across research teams and commercial vendors. Developers building on top of these models benefit from systems that maintain broader competence alongside specialized reasoning improvements.
The research suggests future post-training paradigms will increasingly prioritize multi-objective optimization rather than single-task specialization. This shift could influence how AI companies structure their training pipelines and evaluate model quality, moving beyond narrow benchmark scores toward more comprehensive capability assessments.
- →RLVR training causes measurable forgetting of foundational capabilities like perception and faithfulness in reasoning models
- →RECAP's dynamic reweighting mechanism automatically balances training emphasis across objectives without additional models
- →The method improves both reasoning performance and general capability preservation simultaneously
- →Implementation requires no heavy tuning and integrates directly into existing post-training pipelines
- →Findings suggest future AI training will prioritize multi-objective optimization over single-task specialization