Unsupervised Layer-Wise Dynamic Test Time Adaptation for LLMs

Test-time adaptation (TTA) for large language models (LLMs) updates model parameters at inference time using signals available at deployment. This paper focuses on a common yet under-explored regime: unsupervised, sample-specific TTA, where the model adapts independently for each prompt using only the prompt itself, without gold answers or external supervision. Although appealing, naive unsupervised TTA with a fixed, handcrafted learning rate can be unstable: updates may overfit to prompt-specific statistics, drift from the desired answer distribution, and ultimately degrade generation quality. This failure mode is not surprising, as in this case TTA must adapt to a single prompt within only a few gradient steps, unlike standard training that averages updates over large datasets and long optimization horizons. Therefore, we propose layer-wise dynamic test-time adaptation, a framework which explicitly modulates TTA strength as a function of prompt representation, LLM structure and adaptation step. In our setting, TTA updates only LoRA parameters, and a lightweight hypernetwork predicts per-layer, per-step learning-rate multipliers, enabling fine-grained control. Experiments across various datasets and LLMs consistently show that our method substantially strengthens TTA by learning effective scaling patterns over adaptation steps and transformer layer projections, improving stability while delivering better performance.

AsyncDiff: Asynchronous Timestep Conditioning for Enhanced Text-to-Image Diffusion Inference

Text-to-image diffusion inference typically follows synchronized schedules, where the numerical integrator advances the latent state to the same timestep at which the denoiser is conditioned. We propose an asynchronous inference mechanism that decouples these two, allowing the denoiser to be conditioned at a different, learned timestep while keeping image update schedule unchanged. A lightweight timestep prediction module (TPM), trained with Group Relative Policy Optimization (GRPO), selects a more feasible conditioning timestep based on the current state, effectively choosing a desired noise level to control image detail and textural richness. At deployment, a scaling hyper-parameter can be used to interpolate between the original and de-synchronized timesteps, enabling conservative or aggressive adjustments. To keep the study computationally affordable, we cap the inference at 15 steps for SD3.5 and 10 steps for Flux. Evaluated on Stable Diffusion 3.5 Medium and Flux.1-dev across MS-COCO 2014 and T2I-CompBench datasets, our method optimizes a composite reward that averages Image Reward, HPSv2, CLIP Score and Pick Score, and shows consistent improvement.