loss function design

Current AI research often treats models as static artifacts, overlooking the fundamental influence of training dynamics on critical properties such as capability, bias, robustness, and safety. This work proposes shifting the focus toward the training process itself to establish a science of AI centered on training dynamics. By analyzing the interactions among data, objectives, architectures, and optimizers, the paper develops a theoretical framework that is predictive, intervenable, and design-oriented. Integrating approaches from mechanistic interpretability, fairness, memory mechanisms, and simplicity biases, it uncovers causal links between early-training signals and final model behavior. The study systematically outlines key challenges and open problems, offering both theoretical pathways and practical foundations for extending scaling laws beyond performance to encompass multidimensional model attributes.

Current post-training of language models relies on abstract scalar rewards, which lack transparency regarding the instructional content of preference data and can lead models to learn spurious correlations, resulting in undesirable behaviors such as excessive stylization or sycophancy. This work proposes a data-centric post-training framework that, for the first time, leverages interpretability methods to explicitly model latent conceptual signals within preference data. By analyzing and identifying key features that distinguish preferred from non-preferred responses prior to optimization, the approach integrates interpretability protocols, statistical hypothesis testing, and fine-grained interventions at both feature and data levels. This enables effective diagnosis and suppression of harmful learning signals, significantly reducing off-target behaviors across multiple benchmarks while enhancing model safety and controllable personality.

Existing activation intervention methods struggle to disentangle angular and norm components in concept representations, leading to unstable control effects and opaque mechanisms. This work proposes an angle-norm decomposition framework and, through controlled experiments and evaluations across seven language models, empirically demonstrates that conceptual information is primarily encoded in the angular structure, while the norm plays a critical role in intervention stability. The study elucidates the geometric origins underlying the performance differences among various intervention approaches and advocates for parameterizing interventions via interpretable angular and radial components. This provides both theoretical grounding and practical guidance for achieving more stable and interpretable control over language model behaviors.

This work addresses the long-standing isolation among research domains such as alignment training, model organisms, and toy models, which has hindered empirical cross-pollination and led to redundant exploration and inefficiency. For the first time, it systematically transfers supervised fine-tuning (SFT) practices across these domains by integrating cross-model output training, mixed-strategy data, and benign fine-tuning to rigorously evaluate the portability of key findings. The study demonstrates three successful transfer effects: enhanced behavioral generalization, mitigation of capability degradation, and the critical insight that preserving capabilities alone is insufficient to ensure robustness in subsequent training phases. These results underscore both the efficacy and limitations of reusing methodologies across domains, thereby fostering more synergistic development across disparate research areas.

Inconsistent and nonstandard naming conventions for pre-trained models (PTMs) severely impede model discovery and reliable reuse, yet systematic empirical studies on PTM naming practices remain absent. Method: This paper presents the first large-scale analysis of PTM naming on Hugging Face, grounded in a survey of 108 engineers revealing significant mismatches between current naming practices and engineering requirements. We propose DARA, an automated framework that innovatively integrates deep neural network architectural metadata—such as layer count and number of attention heads—to quantify naming credibility. DARA further employs naming pattern mining and anomaly detection to identify structural errors and semantically misleading naming patterns. Contribution/Results: We introduce the first reusable, empirically grounded framework for assessing PTM naming quality, enabling improved model retrieval accuracy and cross-project reusability. The framework is open-sourced, providing both actionable insights and practical tooling for the ML community.

This work addresses the challenge of selecting effective fine-tuning strategies for encoder-decoder pre-trained language models in generation and question-answering tasks. It proposes the Match Task to Objective (MTO) framework, which establishes the first systematic alignment mechanism between downstream tasks and pre-training objectives. MTO automatically constructs training data and prompt templates that are consistent with the original pre-training objective and extends this alignment to soft prompt tuning, thereby enabling precise task–objective matching. Experimental results demonstrate that MTO achieves over 120% performance improvement under few-shot settings compared to existing methods, significantly outperforms strong baselines in full-data scenarios, and substantially enhances the effectiveness of prompt tuning.

Clinical time series analysis faces significant challenges including limited sample sizes, data heterogeneity, and protocol drift, necessitating generalizable representation learning approaches that jointly support classification and regression tasks. This work proposes PathoFM, a framework that systematically investigates the impact of inductive biases on representation transfer using gait data from spinal cord injury patients. The approach introduces a multi-objective self-supervised pretraining strategy that integrates local structural reconstruction, causal temporal continuity, and individual-specific contextual conditioning, all jointly optimized within a Transformer encoder. Experimental results demonstrate that this dynamics-driven hybrid objective substantially outperforms single-objective methods, achieving superior generalization performance across both cross-task and cross-subject settings.

This work addresses the challenge in aerodynamic inverse design, where high-dimensional geometry and computationally expensive simulations hinder the simultaneous optimization of performance and geometric plausibility. To overcome this, the authors propose a unified framework that integrates optimal design points with design distributions by combining optimization and guided generative modeling. Key innovations include a novel loss function for cost predictor training, a density gradient-based optimization strategy, and an efficient approximate conditional covariance estimation algorithm that enables a guidance generation framework without additional training. The approach is implemented with OpenFOAM simulations and offline reinforcement learning, and validated through 3D-printed wind tunnel experiments. It demonstrates significant performance improvements on both 2D control tasks and high-fidelity 3D benchmarks for automotive and aerospace applications, showcasing both effectiveness and practicality.

Existing control methods for large language models—such as fine-tuning, LoRA, and activation interventions—lack a unified theoretical framework, hindering systematic comparison and mechanistic understanding. This work proposes a preference–utility analysis framework that unifies diverse interventions as dynamic weight updates driven by control signals. By leveraging polarity-contrastive examples on the log-odds scale, the framework jointly quantifies a model’s preference (bias toward a target concept) and utility (coherence of generated text). Drawing on activation manifold theory, we reveal a pervasive trade-off wherein increased preference typically degrades utility, stemming from representational drift away from the effective generative manifold. Guided by this insight, we introduce the SPLIT steering algorithm, which enhances target preferences while more effectively preserving generative utility, thereby demonstrating the universality and controllability of this trade-off.

This work investigates how predictive representation learning objectives often discard exogenous features that, while unpredictable, are relevant for control due to their inherent bias toward predictability. Through a 2×2 controlled experimental design that independently manipulates feature controllability and control relevance, the study systematically evaluates six representation learning objectives—including JEPA, action-conditional JEPA, inverse dynamics models, and their reward- or controllability-augmented variants—on their ability to preserve such features. The analysis reveals, for the first time, the precise failure mechanism by which predictive objectives neglect control-relevant information. To address this, the authors propose Reward-Anchored JEPA, which leverages as little as 2% reward-labeled data. Empirical results demonstrate that all purely predictive, reward-free objectives fail to retain the feature (performing near chance), whereas the proposed method robustly recovers its representation across multiple environments and latent space dimensions.