A new adapter technique lets researchers repurpose a frozen AI model for an entirely different job without retraining it from scratch.
Researchers have introduced MaRK (Markov-adapted Recurrent Kernels), a method for adjusting state space models (SSMs) - an efficient alternative to Transformers - without touching their core weights. Instead of feeding conditioning signals in from the outside, as most adapter methods do, MaRK reaches directly into the model's internal recurrence parameters, modulating how it reads, writes, and remembers information at each step. The team tested the approach on a frozen 111M-parameter Hydra SSM, trying three adapter designs: a Hypernet, a Chebyshev polynomial, and a Discrete Cosine Transform kernel. All three needed only 6.3 to 11 million extra trainable parameters to convert the model from a general-purpose objective into an iterative diffusion setup, with the Chebyshev version performing best at a validation loss of 2.55.
This matters because SSMs are gaining traction as a cheaper, faster alternative to Transformer-based models, but repurposing them for new tasks has mostly meant bolting on external conditioning rather than touching their internal dynamics. MaRK shows that parameter-efficient fine-tuning - the kind of lightweight adaptation popularized by LoRA for Transformers - can emerge naturally from how you structure the adapter, not just from freezing most of the network.
The results so far come from synthetic recovery tests and a single backbone, so the real test is whether this holds up on models and tasks nobody built the benchmark for.