ICLR 2026 论文列表

Differentially private (DP) transfer learning, i.e., fine-tuning a pretrained model on private data, is the current state-of-the-art approach for training large models under privacy constraints. We focus on two key hyperparameters in this setting: the clipping bound $C$ and batch size $B$. We show a clear mismatch between the current theoretical understanding of how to choose an optimal $C$ (stronger privacy requires smaller $C$) and empirical outcomes (larger $C$ performs better under strong privacy), caused by changes in the gradient distributions. Assuming a limited compute budget (fixed epochs), we demonstrate that the existing heuristics for tuning $B$ do not work, while cumulative DP noise better explains whether smaller or larger batches perform better. We also highlight how the common practice of using a single $(C,B)$ setting across tasks can lead to suboptimal performance. We find that performance drops especially when moving between loose and tight privacy and between plentiful and limited compute, which we explain by analyzing clipping as a form of gradient re-weighting and examining cumulative DP noise.

The growing scale of deep learning models has rendered exhaustive hyperparameter (HP) optimization prohibitively expensive. A promising solution is the use of scale-aware HPs, which can enable direct transfer of optimal settings from small-scale grid searches to large models with minimal performance loss. Such approaches are useful when the optimal settings converge "fast" enough with scale. While approaches like the Maximal Update Parameterization ($\mu$P) have empirically displayed fast transfer when scaling model width, a deeper conceptual understanding of the mechanisms that enable this is still missing. Our work establishes a systematic conceptual framework for analyzing fast HP transfer across different synthetic and practical scenarios. In synthetic settings, we present various quantitative examples where transfer either offers a provable computational advantage or fails even under $\mu$P. We then propose a key property that enables the fast transfer often observed in practice: through a novel decomposition of the optimization trajectory, we identify one component that rapidly converges with model width and determines the optimal HPs, and the other that continues to improve the loss with increased width but has negligible impact on HP choice. We conjecture that this decomposition elucidates the key mechanisms behind fast transfer and empirically validate it in practical settings such as LLM training.

Transferring the optimal learning rate from small to large neural networks can enable efficient training at scales where hyperparameter tuning is otherwise prohibitively expensive. To this end, the Maximal Update Parameterization (µP) proposes a learning rate scaling designed to keep the update dynamics of internal representations stable across different model widths. However, the scaling rules of µP rely on strong assumptions, particularly about the geometric alignment of a layer’s inputs with both its weights and gradient updates. In this large-scale empirical investigation, we show that these assumptions hold only briefly at the start of training in the practical setups where learning rate transfer is most valuable, such as LLM training. For the remainder of training it is weight decay rather than µP that correctly stabilizes the update dynamics of internal representations across widths, facilitating learning rate transfer. This suggests µP's scaling primarily acts as a form of implicit learning rate warmup, allowing us to largely replace it with modified warmup schedules. Together these findings fundamentally challenge prevailing beliefs about learning rate transfer and can explain empirical observations such as why µP requires the independent weight decay variant for good transfer.

We develop a theory of transfer learning in infinitely wide neural networks under gradient flow that quantifies when pretraining on a source task improves generalization on a target task. We analyze both (i) fine-tuning, when the downstream predictor is trained on top of source-induced features and (ii) a jointly rich setting, where both pretraining and downstream tasks can operate in a feature learning regime, but the downstream model is initialized with the features obtained after pre-training. In this setup, the summary statistics of randomly initialized networks after a rich pre-training are adaptive kernels which depend on both source data and labels. For (i), we analyze the performance of a readout for different pretraining data regimes. For (ii), the summary statistics after learning the target task are still adaptive kernels with features from both source and target tasks. We test our theory on linear and polynomial regression tasks as well as real datasets. Our theory allows interpretable conclusions on performance, which depend on the amount of data on both tasks, the alignment between tasks, and the feature learning strength.

Continual learning (CL) aims to continuously learn new tasks from data streams. While most CL research focuses on mitigating catastrophic forgetting, memorizing outdated knowledge can cause negative transfer, where irrelevant prior knowledge interferes with new task learning and impairs adaptability. Inspired by how the human brain selectively unlearns unimportant information to prioritize learning and to recall relevant knowledge, we explore the intuition that effective CL should not only preserve but also selectively unlearn prior knowledge that hinders adaptation. We introduce DEtect, Decide, Unlearn in Continual lEarning (DEDUCE), a novel CL framework that dynamically detects negative transfer and mitigates it by a hybrid unlearning mechanism. Specifically, we investigate two complementary negative transfer detection strategies: transferability bound and gradient conflict analysis. Based on this detection, the model decides whether to activate a Local Unlearning Module (LUM) to filter outdated knowledge before learning new task. Additionally, a Global Unlearning Module (GUM) periodically reclaims model capacity to enhance plasticity. Our experiments demonstrate that DEDUCE effectively mitigates task interference and improves overall accuracy with an average gain of up to 5.5\% over state-of-the-art baselines.

Generalizing policies across environments with altered dynamics remains a key challenge in reinforcement learning, particularly in offline settings where direct interaction or fine-tuning is impractical. We introduce Masked Skill Token Training (MSTT), a fully offline hierarchical RL framework that enables policy transfer using observation-only demonstrations. MSTT constructs a discrete skill space via unsupervised trajectory tokenization and trains a skill-conditioned value function using masked Bellman updates, which simulate dynamics shifts by selectively disabling skills. A diffusion-based trajectory generator, paired with feasibility-based filtering, enables the agent to execute valid, temporally extended actions without requiring action labels or access to the target environment. Our results in both discrete and continuous domains demonstrate the potential of mask-guided planning for robust generalization under dynamics shifts. To our knowledge, MSTT is the first work to explore masking as a mechanism for simulating and generalizing across off-dynamics environments. It marks a promising step toward scalable, structure-aware transfer and opens avenues to explore multi-goal conditioning, and extensions to more complex, real-world scenarios.

Flow matching (FM) based on Ordinary Differential Equations (ODEs) has achieved significant success in generative tasks. However, it faces several inherent limitations, including an inability to model trajectory intersections, capture delay dynamics, and handle transfer between heterogeneous distributions. These limitations often result in a significant mismatch between the modeled transfer process and real-world phenomena, particularly when key coupling or inherent structural information between distributions must be preserved. To address these issues, we propose Delay Flow Matching (DFM), a new FM framework based on Delay Differential Equations (DDEs). Theoretically, we show that DFM possesses universal approximation capability for continuous transfer maps. By incorporating delay terms into the vector field, DFM enables trajectory intersections and better captures delay dynamics. Moreover, by designing appropriate initial functions, DFM ensures accurate transfer between heterogeneous distributions. Consequently, our framework preserves essential coupling relationships and achieves more flexible distribution transfer strategies. We validate DFM's effectiveness across synthetic datasets, single-cell data, and image-generation tasks.

Large Language Models (LLMs) are widely applied in real world scenarios, but fine-tuning them comes with significant computational and storage costs. Parameter-Efficient Fine-Tuning (PEFT) methods such as LoRA mitigate these costs, but the adapted parameters are dependent on the base model and cannot be transferred across different backbones. One way to address this issue is through knowledge distillation, but its effectiveness inherently depends on training data. Recent work such as TransLoRA avoids this by generating synthetic data, but this adds complexity because it requires training an additional discriminator model. In this paper, we propose TiTok, a new framework that enables effective LoRA Transplantation through Token-level knowledge transfer. Specifically, TiTok captures task-relevant information through a token-wise contrastive excess between a source model with and without LoRA. This excess highlights informative tokens and enables selective filtering of synthetic data, all without additional models or overhead. Through experiments on three benchmarks across multiple transfer settings, our experiments show that TiTok is consistently effective, achieving average performance gains of +4–8% compared to baselines overall.