parameter sharing

This paper addresses the fundamental challenge of balancing catastrophic forgetting and parameter efficiency when large pre-trained models continuously adapt to dynamic task streams. To this end, we propose the first unified theoretical framework for Parameter-Efficient Continual Fine-Tuning (PECFT). Our framework systematically organizes existing approaches along three dimensions: method taxonomy, evaluation metrics, and core challenges—integrating Parameter-Efficient Fine-Tuning (PEFT) techniques (e.g., adapters, LoRA, prompt tuning) with continual learning strategies (e.g., replay, regularization, architecture expansion). Through a comprehensive review of over 100 studies, we identify key trade-offs between performance and efficiency, and pinpoint scalable memory mechanisms and task-aware parameter updates as critical research frontiers. This work bridges a significant gap at the intersection of continual learning and PEFT, providing both theoretical foundations and practical guidelines for efficient, sustainable adaptation of large language models.

Existing learned optimizers (LOs) exhibit limited meta-generalization—particularly to unseen tasks requiring wider, deeper, or longer training trajectories. This work introduces μ-parameterization (μP) theory systematically into two mainstream LO architectures for the first time, proposing a μP-adapted lightweight meta-training paradigm. Methodologically, we derive theoretical scale-invariance conditions for LOs and design a low-overhead meta-training procedure (<250 GPU-hours). Experiments demonstrate that μLO matches or surpasses VeLO’s performance on large-width models—despite VeLO consuming 4,000 TPU-months—while improving meta-generalization in depth by 5× and extending maximal training-step generalization by 25×. This work establishes a rigorous theoretical foundation and an efficient implementation pathway for scalable, highly generalizable learned optimizers.

This paper addresses incremental learning under resource-constrained settings, proposing an efficient, low-parameter method for adapting pretrained models to novel classes. The approach comprises three key contributions: (1) a Neural Mimicry mechanism that reconstructs pretrained weights with high fidelity; (2) a soft parameter-sharing framework based on weight decomposition, enabling layer-wise reparameterization using fewer than 50 task-specific parameters per layer; and (3) a frozen-backbone architecture augmented with learnable scaling coefficients, eliminating the need for fine-tuning or retraining the base model. The method is model-agnostic—compatible with arbitrary architectures and scales—and integrates seamlessly with existing techniques. Evaluated on six benchmark datasets, it outperforms state-of-the-art methods by 3% in accuracy while reducing parameter count by several orders of magnitude compared to LoRA. This substantial parameter efficiency significantly enhances feasibility for deployment on edge and mobile devices.

This work proposes a modular expert recombination framework to address the limitations of existing model fusion approaches, which typically treat task-specific models as monolithic entities and lack component-level granularity and module reusability. The framework constructs a reusable library of component-level experts and employs a lightweight dynamic routing network to adaptively assemble an optimal sub-model at inference time based on the input. The fusion process is formulated as a bi-objective optimization problem, and a surrogate-assisted evolutionary algorithm efficiently searches for Pareto-optimal configurations. Extensive experiments demonstrate that the proposed method consistently outperforms strong baselines across diverse model scales, task types, and fine-tuning strategies, achieving superior generalization, inference efficiency, and storage economy.

This work addresses catastrophic forgetting in fine-tuning pretrained models, where newly acquired knowledge overwrites previously learned information. To mitigate this issue, the authors propose a function-preserving model expansion approach that mathematically duplicates and scales parameters of selected Transformer submodules during initialization. This technique enables stable training and faithful retention of original model capabilities without altering the initial functionality. By circumventing the traditional trade-off between plasticity and stability, the method achieves performance comparable to full fine-tuning while expanding only a minimal number of layers. Consequently, it fully preserves the model’s original knowledge and substantially reduces computational overhead.

This work addresses catastrophic forgetting in continual learning with pre-trained models when access to previous task data is prohibited. The authors propose a structured low-rank adaptation method grounded in geometric redundancy of pre-trained weights. By analyzing the intrinsic geometric structure of the pre-trained weight space, they identify a protected subspace for parameter updates and formulate the update as \( \Delta W = BAQ^\top \), where frozen matrices \( B \) and \( Q \) project trainable low-rank matrix \( A \) exclusively onto redundant directions. This approach is the first to leverage geometric redundancy to explicitly locate plasticity regions, enabling a controllable trade-off between plasticity and stability without requiring data replay. Experimental results demonstrate that the method effectively suppresses functional drift and significantly improves retention of performance on prior tasks, even under worst-case scenarios.

This work addresses the challenge of efficiently fusing independently trained neural network models for capability reuse without access to original training data and with minimal optimization. The authors propose a novel paradigm of direct weight-space fusion: for single-task settings, they introduce the reference-free C²M³ alignment algorithm; for multi-task scenarios, they develop a framework comprising TSV low-rank decomposition, MASS input-adaptive routing, and MERGE³ evolutionary fusion, grounded in gradient-based approximations of task vectors. By innovatively integrating Frank-Wolfe optimization, item response theory for evaluation, and task vector analysis, the method substantially mitigates task interference and computational overhead—reducing evaluation costs by up to 50×—while maintaining strong performance, all without requiring any original training data.

This work addresses the limitation of existing generative models, which rely on fixed pre-trained parameters and lack the ability to dynamically adapt to individual input instances. To overcome this, the authors propose Composer, a framework that enables instance-level adaptation at test time by conditionally generating and injecting lightweight parameters based on the input, without requiring fine-tuning. Composer introduces, for the first time, a test-time mechanism for instance-specific parameter composition, endowing static models with context-awareness and dynamic adaptability. The approach is compatible with both diffusion and autoregressive architectures and supports quantized deployment. Experimental results demonstrate that Composer consistently enhances generation quality across diverse tasks while maintaining low computational and memory overhead, confirming its effectiveness and broad applicability.

This work addresses the inefficiencies in large-scale recommendation systems caused by maintaining separate models for different scenarios and objectives, which hinders development velocity and delays technology adoption. To overcome this, the authors propose the Standardized Model Template (SMT) framework, which leverages composable, standardized machine learning components to enable “design once, deploy everywhere,” uniformly accommodating diverse data distributions and optimization objectives. By decoupling model architecture from scenario-specific configurations, SMT reduces the complexity of technology deployment from O(n·2ᵏ) to O(n+k), breaking away from the conventional “one objective, one model” paradigm. Empirical evaluation on Meta’s ad ranking system demonstrates that SMT improves average cross-entropy by 0.63%, reduces engineering time per model iteration by 92%, and increases the throughput of technology-model pair adoption by 6.3×.

This work challenges the common intuition that more training data invariably accelerates generalization, investigating how dataset scale affects generalization and memorization dynamics in structured output tasks. Through controlled experiments on the Needleman–Wunsch matrix generation task using a small Transformer and a multiplication-based baseline, the study systematically analyzes validation accuracy and convergence behavior across varying data sizes. The findings reveal an optimal “sweet spot” in dataset scale that yields the fastest generalization; beyond this point, although the model still generalizes, it requires substantially more gradient updates, and the learning of underlying rules becomes decoupled from perfect training fit. This phenomenon is absent in the multiplication baseline, highlighting the distinct optimization dynamics inherent to Transformers.