We show that language models' activations linearly encode when information was learned during training. Our setup involves creating a model with a known training order by sequentially fine-tuning Llama-3.2-1B on six disjoint but otherwise similar datasets about named entities. We find that the average activations of test samples corresponding to the six training datasets encode the training order: when projected into a 2D subspace, these centroids are arranged exactly in the order of training and lie on a straight line. Further, we show that linear probes can accurately (~90%) distinguish "early" vs. "late" entities, generalizing to entities unseen during the probes' own training. The model can also be fine-tuned to explicitly report an unseen entity's training stage (~80% accuracy). Interestingly, the training-order encoding does not seem attributable to simple differences in activation magnitudes, losses, or model confidence. Our paper demonstrates that models are capable of differentiating information by its acquisition time, and carries significant implications for how they might manage conflicting data and respond to knowledge modifications.

Recent advances in Multimodal Large Language Models (MLLMs) have enhanced their versatility as they integrate a growing number of modalities. Considering the heavy cost of training MLLMs, it is necessary to reuse the existing ones and further extend them to more modalities through Modality-incremental Continual Learning (MCL). However, this often comes with a performance degradation in the previously learned modalities. In this work, we revisit the MCL and investigate a more severe issue it faces in contrast to traditional continual learning, that its degradation comes not only from catastrophic forgetting but also from the misalignment between the modality-agnostic and modality-specific components. To address this problem, we propose an elegantly simple MCL paradigm called "MErge then ReAlign" (MERA). Our method avoids introducing heavy training overhead or modifying the model architecture, hence is easy to deploy and highly reusable in the MLLM community. Extensive experiments demonstrate that, despite the simplicity of MERA, it shows impressive performance, holding up to a 99.84% Backward Relative Gain when extending to four modalities, achieving a nearly lossless MCL performance.

In multi-domain learning, a single model is trained on diverse data domains to leverage shared knowledge and improve generalization. The order in which the data from these domains is used for training can significantly affect the model's performance on each domain. However, this dependence is under-studied. In this paper, we investigate the influence of training order (or data mixing) in multi-domain learning using the concept of Lie bracket of gradient vector fields. By analyzing the infinitesimal effects of changing the training order, we identify regions in the parameter space where altering the order between two training domains can benefit the target loss. We validate the predictions of our theoretical framework on the influence of training order (or data mixing) both on a toy example and bilingual LLM pre-training.

Data debugging is to find a subset of the training data such that the model obtained by retraining on the subset has a better accuracy. A bunch of heuristic approaches are proposed, however, none of them are guaranteed to solve this problem effectively. This leaves an open issue whether there exists an efficient algorithm to find the subset such that the model obtained by retraining on it has a better accuracy. To answer this open question and provide theoretical basis for further study on developing better algorithms for data debugging, we investigate the computational complexity of the problem named Debuggable. Given a machine learning model $\mathcal{M}$ obtained by training on dataset $D$ and a test instance $(\mathbf{x}_\text{test},y_\text{test})$ where $\mathcal{M}(\mathbf{x}_\text{test})\neq y_\text{test}$, Debuggable is to determine whether there exists a subset $D^\prime$ of $D$ such that the model $\mathcal{M}^\prime$ obtained by retraining on $D^\prime$ satisfies $\mathcal{M}^\prime(\mathbf{x}_\text{test})=y_\text{test}$. To cover a wide range of commonly used models, we take SGD-trained linear classifier as the model and derive the following main results. (1) If the loss function and the dimension of the model are not fixed, Debuggable is NP-complete regardless of the training order in which all the training samples are processed during SGD. (2) For hinge-like loss functions, a comprehensive analysis on the computational complexity of Debuggable is provided; (3) If the loss function is a linear function, Debuggable can be solved in linear time, that is, data debugging can be solved easily in this case. These results not only highlight the limitations of current approaches but also offer new insights into data debugging.

Identifying latent interactions within complex systems is key to unlocking deeper insights into their operational dynamics, including how their elements affect each other and contribute to the overall system behavior. For instance, in neuroscience, discovering neuron-to-neuron interactions is essential for understanding brain function; in ecology, recognizing the interactions among populations is key for understanding complex ecosystems. Such systems, often modeled as dynamical systems, typically exhibit noisy high-dimensional and non-stationary temporal behavior that renders their identification challenging. Existing dynamical system identification methods often yield operators that accurately capture short-term behavior but fail to predict long-term trends, suggesting an incomplete capture of the underlying process. Methods that consider extended forecasts (e.g., recurrent neural networks) lack explicit representations of element-wise interactions and require substantial training data, thereby failing to capture interpretable network operators. Here we introduce Lookahead-driven Inference of Networked Operators for Continuous Stability (LINOCS), a robust learning procedure for identifying hidden dynamical interactions in noisy time-series data. LINOCS integrates several multi-step predictions with adaptive weights during training to recover dynamical operators that can yield accurate long-term predictions. We demonstrate LINOCS' ability to recover the ground truth dynamical operators underlying synthetic time-series data for multiple dynamical systems models (including linear, piece-wise linear, time-changing linear systems' decomposition, and regularized linear time-varying systems) as well as its capability to produce meaningful operators with robust reconstructions through various real-world examples.

Instruction tuning effectively optimizes Large Language Models (LLMs) for downstream tasks. Due to the changing environment in real-life applications, LLMs necessitate continual task-specific adaptation without catastrophic forgetting. Considering the heavy computational cost, replay-based Continual Learning (CL) methods are the simplest and most widely used for LLMs to address the forgetting issue. However, traditional replay-based methods do not fully utilize instructions to customize the replay strategy. In this work, we propose a novel paradigm called Instruction-based Continual Learning (InsCL). InsCL dynamically replays previous data based on task similarity, calculated by Wasserstein Distance with instructions. Moreover, we further introduce an Instruction Information Metric (InsInfo) to quantify the complexity and diversity of instructions. According to InsInfo, InsCL guides the replay process more inclined to high-quality data. We conduct extensive experiments over 16 tasks with different training orders, observing consistent performance improvements of InsCL. When all tasks have been trained, InsCL achieves performance gains of 3.0 Relative Gain compared with Random Replay, and 27.96 Relative Gain compared with No Replay.

This paper studies the evolving domain of Continual Learning (CL) in large language models (LLMs), with a focus on developing strategies for efficient and sustainable training. Our primary emphasis is on continual domain-adaptive pretraining, a process designed to equip LLMs with the ability to integrate new information from various domains while retaining previously learned knowledge and enhancing cross-domain knowledge transfer without relying on domain-specific identification. Unlike previous studies, which mostly concentrate on a limited selection of tasks or domains and primarily aim to address the issue of forgetting, our research evaluates the adaptability and capabilities of LLMs to changing data landscapes in practical scenarios. To this end, we introduce a new benchmark designed to measure the adaptability of LLMs to these evolving data environments, offering a comprehensive framework for evaluation. We examine the impact of model size on learning efficacy and forgetting, as well as how the progression and similarity of emerging domains affect the knowledge transfer within these models. Our findings uncover several key insights: (i) when the sequence of domains shows semantic similarity, continual pretraining enables LLMs to better specialize in the current domain compared to stand-alone fine-tuning, (ii) training across a diverse range of domains enhances both backward and forward knowledge transfer, and (iii) smaller models are particularly sensitive to continual pretraining, showing the most significant rates of both forgetting and learning. We posit that our research marks a shift towards establishing a more realistic benchmark for investigating CL in LLMs, and has the potential to play a key role in guiding the direction of future research in the field.

Spiking Neural Networks (SNNs) have a greater potential for modeling time series data than Artificial Neural Networks (ANNs), due to their inherent neuron dynamics and low energy consumption. However, it is difficult to demonstrate their superiority in classification accuracy, because current efforts mainly focus on designing better network structures. In this work, enlighten by brain-inspired science, we find that, not only the structure but also the learning process should be human-like. To achieve this, we investigate the power of Curriculum Learning (CL) on SNNs by designing a novel method named CSNN with two theoretically guaranteed mechanisms: The active-to-dormant training order makes the curriculum similar to that of human learning and suitable for spiking neurons; The value-based regional encoding makes the neuron activity to mimic the brain memory when learning sequential data. Experiments on multiple time series sources including simulated, sensor, motion, and healthcare demonstrate that CL has a more positive effect on SNNs than ANNs with about twice the accuracy change, and CSNN can increase about 3% SNNs' accuracy by improving network sparsity, neuron firing status, anti-noise ability, and convergence speed.