●builderIf your fine-tuning workloads involve long-context summarization or QA, HRM offers a drop-in LoRA alternative with significant benchmark gains at identical parameter counts.
●researcherStrong empirical evidence that MLP blocks are better SSM adapter injection sites than attention projectors for sequential state-accumulation tasks — worth replicating on other base models.