A new arXiv preprint introduces a method called Homeostatic Continual Learning that aims to let an AI agent keep learning as its environment changes without losing previously acquired knowledge. The approach targets catastrophic forgetting, a long-standing problem in continual learning research. The abstract provides only a brief description of the method's core mechanism.
A revised arXiv preprint in machine learning looks at "intruder dimensions" that can appear during LoRA fine-tuning: new leading singular vectors of the updated weight matrix W+BA that are nearly orthogonal to the pretrained model's singular vectors, a phenomenon previously tied to catastrophic forgetting. The abstract notes that no theory has explained these dimensions since they were first identified, and frames the new work as a spectral law addressing that gap. The excerpt provided is truncated, so the full results and claims are not yet visible.
A new study examines how autoregressive speech recognition systems that output timestamps as decoded tokens can gradually lose alignment during long stretches without speech. The authors propose a replay-based distribution editing approach that corrects this drift while limiting forgetting of previously learned behavior. The work targets timestamped transcription without relying on frame-level aligners or inference-time fixes.
A revised arXiv preprint describes a continual learning approach that leverages pre-trained models to help systems keep earlier knowledge while picking up new tasks. The work targets catastrophic forgetting, where performance on previously learned tasks degrades as new ones are acquired. The paper is listed as a replacement submission, with no peer-reviewed venue indicated.
A new arXiv paper addresses how domain-specific fine-tuning can erode the general abilities an LLM originally had. The authors propose a counteraction-aware extension of multi-teacher on-policy distillation, which trains on student-generated text under multiple teachers to restore lost capabilities while keeping domain performance. The method targets the trade-off between specialization and retaining broad competence.
A revised arXiv preprint examines how small language models deployed on edge devices can be personalized over time without losing prior knowledge. The work focuses on monitoring stability during sequential adaptation, a known risk when models are updated repeatedly. It is a research contribution rather than a product or model release.
A new arXiv preprint describes a decentralized learning setup where each agent keeps one private data sample alongside a shared model, and samples are processed one after another. Each update is designed to keep the model's predictions at already-learned samples unchanged, which the authors present as a way to limit forgetting while protecting privacy. The work sits at the intersection of gossip-style distributed training and privacy-preserving machine learning.
A new arXiv paper examines ways to keep machine learning models from forgetting previously learned classes when trained incrementally on sound event classification. It evaluates architectural and regularization strategies using the FSD50K dataset and related benchmarks. The work is a research investigation rather than a released product.
A new arXiv paper studies how synthetic data affects the ability of LLM agents to select the right external skill from large repositories. The authors describe a deployed skill router covering 34,396 skills and run a large-scale evaluation of retrieval under limited data conditions. The findings point to catastrophic forgetting as a risk when synthetic data is used for training these routers.
A new arXiv paper examines where and how much neural networks forget when trained sequentially on gynecological medical imaging tasks. The authors analyze forgetting layer by layer to identify which parts of a segmentation model should be preserved and which can be adapted. The work targets scenarios where consecutive clinical tasks vary in imaging modality and anatomy.
A new arXiv paper presents a method that distills image prototypes to guide test-time adaptation of models facing distribution shifts. The technique addresses two common failure modes in this setting: error buildup from unreliable pseudo-labels and degradation of knowledge learned during pretraining. By anchoring adaptation to distilled prototypes, the authors aim to make inference-time model updates more stable.