MaRK: Contextual Modulation of Recurrent Parameters in Pretrained SSMs
An approach is proposed that adapts the recurrent operator of a frozen state-space model to the context through low-rank parameter changes, affecting the input–output kernel at each generation step. A comparison of three adapter geometries shows that Chebyshev polynomials yield lower validation error, while the constrained parameterization allows for analytical stability verification.

Purpose and essence of MaRK
MaRK (Markov-adapted Recurrent Kernels) is a framework for dynamically conditioning operators. It maps context vectors to bounded modulations of recurrence operators.
- In the LPV-SSM representation, MaRK defines a context-indexed family of Markov parameter sequences.
- The approach is designed for pretrained SSMs.
- Standard conditioning mechanisms leave the underlying temporal dynamics unchanged. These include feeding information into the input and modulating activations.
Practical criterion: MaRK is needed when the temporal dynamics of an SSM must change, rather than only its inputs or activations.
Which SSM parameters are modulated
MaRK modulates five groups of parameters in a frozen SSM:
- A — recurrence;
- B — input read-in;
- C — output readout;
- D — skip;
- Δ — discretization.
Practical criterion: MaRK changes A, B, C, D, and Δ. If a task requires changing these parameters, the approach covers them.
Adapters and frozen-base training
MaRK uses a frozen Hydra SSM base model with 111 million parameters. The study covers three adapter geometries:
- Hypernet;
- Chebyshev polynomials;
- Discrete Cosine Transform (DCT) kernels.
The adapters change the Markov parameter sequence through auxiliary low-rank mappings. The base model remains frozen.
Switching from a bidirectional objective to an iterative diffusion setting requires 6.3–11 million trainable auxiliary parameters. Practical criterion: only the auxiliary mappings need to be trained, while the base SSM remains frozen.
Experimental results and choice of geometry
In synthetic LPV recovery experiments and Markov operator diagnostics, MaRK recovers coordinate-invariant temporal operators under matched assumptions. The method produces distinct, time-step-conditioned stable memory profiles.
| Adapter geometry | Validation loss |
|---|---|
| Chebyshev | 2.55 |
| DCT | 2.59 |
| Hypernet | 3.77 |
Chebyshev achieved the best validation loss among the three options, at 2.55. DCT scored 2.59, and Hypernet 3.77. Practical criterion: based on validation loss in the described experiments, Chebyshev scored 2.55, DCT 2.59, and Hypernet 3.77. The choice of adapter geometry affects the result.
Stability and limitations
A bounded parameterization of recurrence provides an analytical Affine Quadratic Stability certificate for the modulated recurrence. MaRK is intended for iterative generation with pretrained SSMs. Each diffusion step can change the model’s input–output memory kernel.
Practical criterion: MaRK is appropriate when an analytical stability certificate is needed, along with a change to the memory kernel at each diffusion step.
Publication status and version
- Paper authors: Syed Ibrahim Omer, Ginny Y. Wong, Xiangyu Zhao.
- The arXiv paper 2610.09092 contains version v1 dated 6 Oct 2026.
- The latest revision is dated 18 Oct 2026 (version v3).
- The journal reference points to NeurIPS 2026 Submission.
Practical criterion: an accurate reference to the results includes version v3 dated 18 Oct 2026 and the status NeurIPS 2026 Submission.
Frequently asked questions
Related materials
All materials
B-PINN and Collider Bias: Causes, Diagnosis, and a Chain-Based Alternative
The authors show that the widely used B-PINN framework, which links independent prior parameters through virtual likelihoods of differential equation residuals, can systematically bias posterior estimates of physical quantities. They propose instead a hierarchical chain model that discretizes stochastic dynamics and uses particle MCMC sampling, and provide criteria for diagnosing the reliability of the standard approach.
14 November 2026
Fine-Tuning Generative Image Models: Jointly Accounting for Preferences and Categories
The article describes a post-training method for open-domain text-to-image models that supplements human preference signals with rubric-based checks for prompt alignment and desired properties. The results show that weighted reward averaging underperforms the proposed approach, while the Arena snapshot shows notable Elo gains for fine-tuned Flux2dev and Ideogram-4.
14 November 2026
Zero2Repo Benchmark: The Limits of Autonomous Project Building from an Empty Directory
The study proposes evaluating coding agents on a full repository build: the input is a specification, an interface contract, and a blank environment; validation uses hidden tests without an LLM judge. Even leading agents rarely pass every check, and most failures stem from individual missed requirements.
13 November 2026