深度学习 · transfer learning

Artificially generated speech is increasingly embedded in everyday life. Voice cloning in particular enables applications where identity preservation is important, such as completing a recording, dubbing in a new language, or preserving the voices of individuals with speech loss. However, in our work, we find that despite the term, voice cloning does not faithfully "clone" an individual's voice. Instead, we find that widely-used voice cloning models systematically apply style transfer to source voices. As rated by human annotators, cloned voices are perceived as more authoritative, warm, customer-service-like, and human-like compared to their sources. Human annotators also report greater trust in cloned voices than source voices, and a greater willingness to disclose sensitive personal information to them. Our work furthermore shows that voice cloning leads to homogenization of speaker characteristics, as measured by reduced variance in accent, speaking rate, and the audio embedding space. Together, our results highlight a new set of limitations and risks of voice cloning technology and their potential impact on human behavior.

We study transfer learning for a linear regression task using several least-squares pretrained models that can be overparameterized. We formulate the target learning task as optimization that minimizes squared errors on the target dataset with penalty on the distance of the learned model from the pretrained models. We analytically formulate the test error of the learned target model and provide the corresponding empirical evaluations. Our results elucidate when using more pretrained models can improve transfer learning. Specifically, if the pretrained models are overparameterized, using sufficiently many of them is important for beneficial transfer learning. However, the learning may be compromised by overparameterization bias of pretrained models, i.e., the minimum L2-norm solution's restriction to a small subspace spanned by the training examples in the high-dimensional parameter space. We propose a simple debiasing via multiplicative correction factor that can reduce the overparameterization bias and leverage more pretrained models to learn a target predictor.

Scaling laws and hyperparameter tuning remain central challenges in training large-scale models, particularly under differential privacy (DP), where optimization is further complicated by noise injection and per-sample gradient clipping. While Maximal Update Parametrization ($\mu$P) allows for zero-shot hyperparameter transfer in standard training, we show that it fails under DP constraints due to the distinct spectral properties of high-dimensional noise versus low-rank gradients. In this work, we introduce DP-$\mu$P, a theoretically grounded framework that extends $\mu$P to the privacy-preserving setting. By analyzing the interaction between gradient clipping, noise injection, and feature learning, we derive a Signal-to-Noise Ratio (SNR) condition that governs layer-wise learning rate scaling. Combining this with our proposed spectral scaling formulation, we provide specific scaling rules for DP-SGD, DP-Adam, and DP-Muon. Empirically, we demonstrate that DP-$\mu$P enables effective zero-shot transfer of learning rates across model widths on diverse architectures (MLP, ViT, GPT-2), matching or exceeding the performance of individually tuned baselines while significantly reducing computational costs.

Compound AI systems promise capabilities beyond those of individual models, yet their success depends critically on effective orchestration. Existing routing approaches face two limitations: (1) input-level routers make coarse query-level decisions that ignore evolving task requirements; (2) RL-trained orchestrators are expensive to adapt and often suffer from routing collapse, repeatedly invoking one option in multi-turn scenarios. We introduce SkillOrchestra, a framework for skill-aware orchestration. Instead of directly learning a routing policy end-to-end, SkillOrchestra learns fine-grained skills from execution experience and models agent-specific competence and cost under those skills. At deployment, the orchestrator infers the skill demands of the current interaction and selects agents that best satisfy them under an explicit performance-cost trade-off. Extensive experiments across ten benchmarks demonstrate that SkillOrchestra outperforms SoTA RL-based orchestrators by up to 27.5% with 700x and 300x learning cost reduction compared to Router-R1 and ToolOrchestra, respectively. These results show that explicit skill modeling enables scalable, interpretable, and sample-efficient orchestration, offering a principled alternative to data-intensive RL-based approaches.

Representational Engineering (RepE) aims to expose and steer high-level concept directions in language model activations, and recent work shows that hidden states can be transferred across heterogeneous architectures via trained neural adapters. Whether specific concepts survive such transfer remains an open question. We study one such concept, \textit{truth}, by mapping source-model activations through autoencoder adapters into target-model representation space, varying injection strength $\alpha \in [0, 1]$, and reading the result with the target's native truth probe across 21 cross-family trajectories from four model families, with comparison against a closed-form rigid alignment baseline on a representative subset. We find that ordinal truth ranking is preserved through partial injection (median AUROC near the native ceiling through $\alpha \in [0, 0.7]$) but collapses sharply at full injection ($\alpha{=}1$), with parametric separability dropping $\sim$$14\times$ while rank order partially survives. The apparent accuracy overshoot under partial injection decomposes cleanly into two effects: a large recalibration component dominated by native probe miscalibration ($R^2 = 0.75$), and a small but universal genuine separability gain present in all 21 trajectories. These findings characterize cross-family truth structure as partially compatible, ordinally stable, but non-isomorphic: preserved enough under partial injection that learned and rigid alignment methods both surface meaningful truth signal, but with full replacement representing a hard failure mode that practitioners should treat as a deployment boundary.

We present a data poisoning attack---Phantom Transfer---with the property that, even if you know precisely how the poison was placed into an otherwise benign dataset, you cannot filter it out. We achieve this by modifying subliminal learning to work in real-world contexts and demonstrate that the attack works regardless of which model produces the data, which model is trained on the data or what the attack target is. The attack survives 11 tested data-level defences, including one where every sample is paraphrased by a different model. We characterise when this attack works best and show that it can be used to plant password-triggered behaviours into models while still beating defences. In short, we provide an existence proof that maximum-affordance defences can fail to stop sophisticated data poisoning attacks. We suggest that future defences should supplement by include white-box methods and post-training model audits.

We study issues which may arise when using transformer models to formally reason about problems stated informally using natural language. In particular, we focus on transfer learning between tasks that are semantically equivalent but phrased differently. We assume that examples with a particular target label are only available for a subset of tasks, and aim to transfer zero-shot the ability to predict this label to other tasks. Using controlled experiments on propositional logic problems, we show that transformers are capable of such transfer, but that it can be impeded by the presence of task-specific tokens, due to the model picking up on correlations between these tokens and the label early in the training. As a simple solution, we show that introducing the target label later in the training dramatically improves transfer.

Large Language Models (LLMs) are trained to support an increasing number of languages, yet their predefined tokenizers remain a bottleneck for adapting models to lower-resource or distinct-script languages. Existing tokenizer transfer methods typically rely on semantic heuristics to initialize new embeddings, ignoring higher-layer model dynamics and limiting transfer quality. We propose Model-Aware Tokenizer Transfer (MATT), a method that incorporates model internals into the tokenizer transfer process. MATT introduces an Attention Influence Modeling (AIM) objective that distills inter-token communication patterns from a source model into a target model with a new tokenizer, providing an efficient warm-up before standard language modeling. Unlike approaches that focus solely on embedding similarity, MATT leverages attention behavior to guide embedding initialization and adaptation. Experiments across diverse linguistic settings show that MATT recovers a large fraction of the original model’s performance within a few GPU hours, outperforming heuristic baselines. These results demonstrate that incorporating model-level signals offers a practical and effective path toward robust tokenizer transfer in multilingual LLMs.

Transfer learning in reinforcement learning (RL) has shown strong empirical success. In this work, we take a more principled perspective by studying when and how transferring knowledge between MDPs (a source and a target) can be provably beneficial. Specifically, we consider the case where there exists an undo map such that applying this map to the target’s state space recovers the source exactly. We propose an algorithm that learns this map by matching state feature statistics gathered from both MDPs, and then uses it to transfer the source policy. We theoretically justify the algorithm by analyzing the setting when the undo map is linear and the source is linearly-$Q^\star$ realizable, where our approach has strictly better sample complexity than tabula rasa RL in the target MDP. Empirically, we demonstrate that these benefits extend beyond this regime: on challenging continuous control tasks and Atari games, our method achieves significantly better sample efficiency. Overall, our results highlight how shared structure between tasks can be leveraged for efficient transfer of policies across environments.

The Normalized Transformer, or nGPT (Loshchilov et al., 2025) achieves impressive training speedups and does not require weight decay or learning rate warmup. However, despite having hyperparameters that explicitly scale with model size, we observe that nGPT does not exhibit learning rate transfer across model dimension and token horizon. To rectify this, we combine numerical experiments with a principled use of alignment exponents (Everett et al., 2024) to revisit and modify the $\mu$P approach to yperparameter transfer (Yang and Hu, 2021). The result is a novel nGPT parameterization we call $\nu$GPT. Through extensive empirical validation, we find $\nu$GPT exhibits learning rate transfer across width, depth, and token horizon.

We study learning rate (LR) scaling for hybrid architectures combining Transformer and State-Space Model (SSM) blocks, a class adopted by several recent production language models. In particular, we focus on the gap between the theoretical scaling rules derived for SSMs under zero-order-hold (ZOH) discretization at infinite width with proportional state size, and the field-standard practical implementations using simplified-ZOH Mamba at fixed state size. Surprisingly, in this practical regime hybrid architectures achieve a near-zero LR transfer gap across widths 256-2048 and depths 4-32 at sub-billion-parameter scale using only the original $\mu$P prescription, even though SSM operations fall outside its Tensor Programs representability conditions and every parameterization we test fails the standard coordinate-check diagnostic of $\mu$P correctness. We attribute this to a two-condition decomposition of LR transfer in hybrid architectures: a global update-to-weight invariance, enforced by $\mu$P's initialization and LR scaling; and a local per-component balance, provided by AdamW's $g/\sqrt{v}$ per-parameter normalization. Our observations show that the optimal LR is invariant to width up to $8\times$, as well as to depth, sequence length, batch size, and Transformer-to-SSM ratio, and transfers to Nemotron-H, a production hybrid outside our custom architecture set. We hope these findings fill the gap between theoretical scaling rules and practical hybrid implementations, and stimulate further research toward bridging it.

LLMs exhibit uneven performance across programming languages, yet benchmark gaps often reflect narrow task regimes and evaluation-specific confounds. This makes it hard to tell whether poor performance reflects a broader weakness with a given language, a broader difficulty on a given task, or a limitation specific to that task-language combination. We adapt Information Parity, a measure of relative cross-language representational efficiency developed for natural language, to code. We define Code Information Parity (Code IP) on a controlled parallel corpus of 22 algorithms across 20 languages, audited to preserve core semantics while minimizing boilerplate and excluding library shortcuts. The results are regime-specific: Code IP correlates strongly with multilingual benchmark performance on short-horizon generation ($\rho$ from $-0.55$ to $-0.81$), remains negative though moderated on execution prediction ($\rho=-0.429$), and disappears on whole-program synthesis ($\rho=+0.105$). Rather than a universal metric of multilingual coding ability, Code IP isolates a reliable cross-language signal on short-horizon generation and delineates where that signal no longer separates cleanly from broader task and evaluation effects.

Dense Associative Memory (DenseAM) is a promising family of AI architectures that is represented by a neural network performing temporal dynamics on an energy landscape. While hyperparameter transfer methods are well-studied for feed-forward networks, these methods have not been developed for settings in which weights are shared across layers and within the layer, which is common in DenseAMs. Additionally, DenseAMs utilize rapidly peaking activation functions that are rarely used in feed-forward architectures. The confluence of these aspects makes DenseAM a challenging framework for using existing methods for hyperparameter transfer. Our work initiates the development of hyperparameter transfer methods for this class of models. We derive explicit prescriptions for how the hyperparameters tuned on small models can be transferred to models trained at scale. We demonstrate excellent agreement between these theoretical findings and empirical results.

Multimodal learning has advanced visual understanding through powerful multimodal LLMs (MLLMs). In visual generation, however, these models are often used only as global text or context encoders for diffusion generators, limiting their ability to provide structured spatial and temporal guidance. This creates an interface gap: MLLMs can parse complex layouts, attributes, and knowledge-intensive scenes, yet current generation pipelines often struggle to transfer such understanding into images and videos with precise, controllable structure. We propose MetaCanvas, a lightweight framework that organizes MLLM representations into spatially indexed canvas tokens and interfaces them with diffusion generators through patch-wise residual fusion. This provides structured spatial and spatiotemporal conditioning rather than relying only on a single global embedding or 1D query sequence. We implement MetaCanvas on three diffusion backbones and evaluate it across six generation and editing tasks requiring precise layouts, robust attribute binding, and fine-grained multimodal control. Under matched settings, MetaCanvas consistently improves over global-conditioning baselines, narrows the gap to specialized closed-source visual generation and editing systems, and unlocks new editing and in-context capabilities for video diffusion backbones. These results suggest that spatially aligned MLLM-derived canvas tokens provide a promising latent interface for transferring multimodal understanding into diffusion-based generation.

Fine-tuning LLMs on narrow harmful datasets can induce Emergent Misalignment (EM), where models exhibit misaligned behavior far beyond the fine-tuning distribution. We argue that emergent misalignment can be better understood as a data-mediated transfer phenomenon: harmful fine-tuning examples do not induce uniform behavioral spillover, but interact with the structural properties of the dataset and the difficulty of the tasks relative to the model. Across our experiments, we find that misalignment appears more readily when fine-tuning and evaluation prompts share similar underlying functional structure, when prompts leave more room for coherent harmful completions, and when the target behavior has been more reliably learned by the model. The training pipeline itself also matters: pretraining composition shapes later misalignment. We further study Subliminal Learning (SL), where misalignment is transmitted by fine-tuning on seemingly benign data generated by a harmful teacher. Moving beyond the standard SFT setting, we for the first time compare this transfer under off-policy and on-policy distillation as well, allowing us to separate the roles of the teacher guidance and the training data distribution in transmitting misalignment. Together, these results argue for a data-centric view: Emergent/subliminal misalignment should not be treated as a simple consequence of isolated harmful fine-tuning examples, but as the result of interactions between fine-tuning data structure, pretraining distributions, and training channels.

Production deployments often swap between different-sized models in a family for cost-quality cascading, mid-conversation switching, and routing, and each swap forces the receiver to repay the prefill from scratch. We propose cross-model KV cache transfer, where the receiver reuses the source’s KV cache, skipping prefill. We find that cross-model KV has substantial linear structure across matched-KV pairs, where source and target share KV head count and per-head dimension. On Qwen3 14B→32B, one source layer explains 56% of variance in the target’s keys and 32% in values, rising to 79% and 65% with multiple source layers. Building on this, we design a closed-form ridge mapper that operates per head and proceeds in three steps. First, for each target layer we select the top-kmost predictive source layers and concatenate their KV as input. Second, we strip RoPE from the keys before mapping, so the fit is position-free and reusable across context lengths. Third, we fit ridge regression on a small calibration set of 500 FineWeb-Edu sequences of 1,024 tokens each. Surprisingly, across six pairs in three families, this linear mapper retains 71–98% of the receiver’s standalone-prefill accuracy on four pairs, while two degrade sharply. A nonlinear MLP recovers up to +22 pp HellaSwag retention on the failures. The mapper runs 3–31×faster than re-prefill and remains stable across multi-turn handoff, making cross-model KV cache transfer practical.

Policies achieving strong performance in simulators or learned world models can fail when deployed into the real environment. While there must be environmental differences that account for these failures, not all differences are equally responsible. A benign error of irrelevant visual details may co-exist with a failure to predict a single critical transition causing an inevitable catastrophic failure. In this paper, we introduce \emph{counterfactual debugging} as an approach for \emph{world model transfer gap attribution} to identify the time steps whose transition or reward errors are \emph{causally responsible} for a performance degradation exhibited in a real trajectory, rather than only visually or statistically different. We develop a scalable algorithm for computing these attributions that exploits the sparsity of causal errors by recursively divide-and-conquer, significantly reducing computational costs and achieving an exponential speedup. The resulting ranked attribution report can better explain the performance gap in the real trajectory. Experiments across several distinct environments with a variety of injected world-model failures (observation corruption, reward misprediction, and physics violations) demonstrate that counterfactual debugging correctly identifies the errors that are responsible for performance gap, providing theoretically grounded, actionable insights for model improvement.

Style transfer aims to render target content in the style of a reference image, but existing methods often suffer from content leakage, where objects, layouts, or semantics from the style reference appear in the generated output. Although prior data-driven and training-free methods can reduce leakage, they often face a leakage--degradation dilemma: stronger content suppression may weaken style fidelity, while richer style preservation may reintroduce unwanted reference content. We identify this dilemma across the full style-transfer pipeline, including feature separation, feature-space grounding, and diffusion generation. To address these issues, we propose CLeaR, a training-free framework for content-leakage-resistant style transfer. CLeaR first uses Orthogonal Subspace Projection to define content-reduced style targets in each vision foundation model (VFM) feature space. It then performs Ensemble Inversion, which optimizes a shared pixel-space style anchor satisfying style constraints across multiple VFMs. Finally, Energy-Guided Calibration maintains style alignment during diffusion sampling by steering the denoising trajectory toward the ensemble-defined style manifold. We further provide a theoretical analysis showing that the style-anchor estimation error decreases with the number of VFMs. Experiments on StyleBench demonstrate that CLeaR improves style alignment, reduces content leakage, and achieves better LLM-as-Judge evaluation compared with existing methods.