A new framework lets a vision-language model improve itself without ever updating its weights.
Dynamo, described in a new paper, works by having a frozen model review its own correct and incorrect answers on a small labeled dataset. From that self-audit, it generates two things: reusable reasoning "skills" for problems that trip up the model's logic, and executable visual tools for problems rooted in perception. Both accumulate in a persistent library the model can draw on at inference time. Tested across four visual reasoning benchmarks and five different model backbones, Dynamo lifted accuracy in all 20 model-benchmark combinations, averaging a 5.6-point gain.
The more striking result is how it compares to reinforcement learning. Against task-specific RL approaches like VTool-R1 and DeepEyes, Dynamo closes 65-99% of the performance gap at a fraction of the compute cost. That is a meaningful trade-off: RL fine-tuning is expensive, slow, and locks improvements into a single model checkpoint. A training-free library that stacks on top of any frozen backbone is far easier to maintain and redeploy.
The catch, as with most self-improvement schemes, is that the framework's library is only as good as the mistakes it learns from — and on tasks far outside the training subset, that library may simply not have the right tool for the job.