Introduces Composer, a paradigm that generates input-specific parameter adaptations at inference time to enable dynamic per-input model specialization.
March 31, 2026
Original Paper
Test-Time Instance-Specific Parameter Composition: A New Paradigm for Adaptive Generative Modeling
arXiv · 2603.27665
The Takeaway
Moves beyond static pretrained weights by allowing models to inject instance-specific parameters during inference. This provides higher quality and context-aware generation without the need for traditional fine-tuning or retraining.
From the abstract
Existing generative models, such as diffusion and auto-regressive networks, are inherently static, relying on a fixed set of pretrained parameters to handle all inputs. In contrast, humans flexibly adapt their internal generative representations to each perceptual or imaginative context. Inspired by this capability, we introduce Composer, a new paradigm for adaptive generative modeling based on test-time instance-specific parameter composition. Composer generates input-conditioned parameter adap