Large language models (LLMs) can adapt to new tasks via in-context learning (ICL) without parameter updates, making them powerful learning engines for fast adaptation. While extensive research has examined ICL as a few-shot learner, whether it can achieve long-term retention and cross-task knowledge accumulation when multitasks arrive sequentially remains underexplored. Motivated by human memory studies, we investigate the retention characteristics of ICL in multitask settings and extend it to in-context continual learning (ICCL), where continual learning ability emerges through task scheduling and prompt rearrangement. Experiments on Markov-Chain benchmarks demonstrate that, for specific large-language models, ICCL benefits from distributed practice (DP) in a manner analogous to humans, consistently revealing a spacing "sweet spot" for retention. Beyond retention performance, we propose a human-retention similarity metric to quantify how closely a continual-learning (CL) method aligns with human retention dynamics. Using this metric, we show that linear-attention models such as MAMBA and RWKV exhibit particularly human-like retention patterns, despite their retention performance lagging behind that of Transformer-based LLMs. Overall, our results establish ICCL as both cognitively plausible and practically effective, providing an inference-only CL paradigm that mitigates catastrophic forgetting and addresses the stability-plasticity dilemma in conventional CL methods.

Multi-Domain Continual Learning (MDCL) acquires knowledge from sequential tasks with shifting class sets and distribution. Despite the Parameter-Efficient Fine-Tuning (PEFT) methods can adapt for this dual heterogeneity, they still suffer from catastrophic forgetting and forward forgetting. To address these challenges, we propose a Two-Level Routing Grouped Mixture-of-Experts (TRGE) method. Firstly, TRGE dynamically expands the pre-trained CLIP model, assigning specific expert group for each task to mitigate catastrophic forgetting. With the number of experts continually grows in this process, TRGE maintains the static experts count within the group and introduces the intra-group router to alleviate routing overfitting caused by the increasing routing complexity. Meanwhile, we design an inter-group routing policy based on task identifiers and task prototype distance, which dynamically selects relevant expert groups and combines their outputs to enhance inter-task collaboration. Secondly, to get the correct task identifiers, we leverage Multi-modal Large Language Models (MLLMs) which own powerful multimodal comprehension capabilities to generate semantic task descriptions and recognize the correct task identifier. Finally, to mitigate forward forgetting, we dynamically fuse outputs for unseen samples from the frozen CLIP model and TRGE adapter based on training progress, leveraging both pre-trained and learned knowledge. Through extensive experiments across various settings, our method outperforms other advanced methods with fewer trainable parameters.

The foundation model enables general-purpose problem-solving and enjoys desirable rapid adaptation due to its adopted cross-task generalization paradigms, e.g., pretraining, meta-training, and finetuning. Recent advances in these paradigms show the crucial role of challenging tasks' prioritized sampling in enhancing adaptation robustness. However, ranking task difficulties exhausts massive task queries to evaluate, thus computation and annotation intensive, which is typically unaffordable in practice. This work underscores the criticality of both adaptation robustness and learning efficiency, especially in scenarios where tasks are risky or costly to evaluate, e.g., policy evaluations in Markov decision processes (MDPs) or inference with large models. To this end, we present Model Predictive Task Sampling (MPTS) to establish connections between the task space and adaptation risk landscape to form a theoretical guideline in robust active task sampling. MPTS characterizes the task episodic information with a generative model and directly predicts task-specific adaptation risk values from posterior inference. The developed risk learner can amortize expensive evaluation and provably approximately rank task difficulties in the pursuit of task robust adaptation. MPTS can be seamlessly integrated into zero-shot, few-shot, and many-shot learning paradigms. Extensive experimental results are conducted to exhibit the superiority of the proposed framework, remarkably increasing task adaptation robustness and retaining learning efficiency in contrast to existing state-of-the-art (SOTA) methods. The code is available at the project site https://github.com/thu-rllab/MPTS.

The purpose of offline multi-task reinforcement learning (MTRL) is to develop a unified policy applicable to diverse tasks without the need for online environmental interaction. Recent advancements approach this through sequence modeling, leveraging the Transformer architecture's scalability and the benefits of parameter sharing to exploit task similarities. However, variations in task content and complexity pose significant challenges in policy formulation, necessitating judicious parameter sharing and management of conflicting gradients for optimal policy performance. Furthermore, identifying the optimal parameter subspace for each task often necessitates prior knowledge of the task identifier during inference, limiting applicability in real-world scenarios with variable task content and unknown current tasks. In this work, we introduce the Harmony Multi-Task Decision Transformer (HarmoDT), a novel solution designed to identify an optimal harmony subspace of parameters for each task. We formulate this as a bi-level optimization problem within a meta-learning framework, where the upper level learns masks to define the harmony subspace, while the inner level focuses on updating parameters to improve the overall performance of the unified policy. To eliminate the need for task identifiers, we further design a group-wise variant (G-HarmoDT) that clusters tasks into coherent groups based on gradient information, and utilizes a gating network to determine task identifiers during inference. Empirical evaluations across various benchmarks highlight the superiority of our approach, demonstrating its effectiveness in the multi-task context with specific improvements of 8% gain in task-provided settings, 5% in task-agnostic settings, and 10% in unseen settings.

Large Language Models (LLMs) have demonstrated remarkable in-context learning (ICL) capabilities. In this study, we explore a surprising phenomenon related to ICL: LLMs can perform multiple, computationally distinct ICL tasks simultaneously, during a single inference call, a capability we term "task superposition". We provide empirical evidence of this phenomenon across various LLM families and scales and show that this phenomenon emerges even if we train the model to in-context learn one task at a time. We offer theoretical explanations that this capability is well within the expressive power of transformers. We also explore how LLMs internally compose task vectors during superposition. Furthermore, we show that larger models can solve more ICL tasks in parallel, and better calibrate their output distribution. Our findings offer insights into the latent capabilities of LLMs, further substantiate the perspective of "LLMs as superposition of simulators", and raise questions about the mechanisms enabling simultaneous task execution.

Meta-learning is a practical learning paradigm to transfer skills across tasks from a few examples. Nevertheless, the existence of task distribution shifts tends to weaken meta-learners' generalization capability, particularly when the task distribution is naively hand-crafted or based on simple priors that fail to cover typical scenarios sufficiently. Here, we consider explicitly generative modeling task distributions placed over task identifiers and propose robustifying fast adaptation from adversarial training. Our approach, which can be interpreted as a model of a Stackelberg game, not only uncovers the task structure during problem-solving from an explicit generative model but also theoretically increases the adaptation robustness in worst cases. This work has practical implications, particularly in dealing with task distribution shifts in meta-learning, and contributes to theoretical insights in the field. Our method demonstrates its robustness in the presence of task subpopulation shifts and improved performance over SOTA baselines in extensive experiments. The project is available at https://sites.google.com/view/ar-metalearn.

Dealing with representation shift is one of the main problems in online continual learning. Current methods mainly solve this by reducing representation shift, but leave the classifier on top of the representation to slowly adapt, in many update steps, to the remaining representation shift, increasing forgetting. We propose DeepCCG, an empirical Bayesian approach to solve this problem. DeepCCG works by updating the posterior of a class conditional Gaussian classifier such that the classifier adapts instantly to representation shift. The use of a class conditional Gaussian classifier also enables DeepCCG to use a log conditional marginal likelihood loss to update the representation, which can be seen as a new type of replay. To perform the update to the classifier and representation, DeepCCG maintains a fixed number of examples in memory and so a key part of DeepCCG is selecting what examples to store, choosing the subset that minimises the KL divergence between the true posterior and the posterior induced by the subset. We demonstrate the performance of DeepCCG on a range of settings, including those with overlapping tasks which thus far have been under-explored. In the experiments, DeepCCG outperforms all other methods, evidencing its potential.

Animals thrive in a constantly changing environment and leverage the temporal structure to learn well-factorized causal representations. In contrast, traditional neural networks suffer from forgetting in changing environments and many methods have been proposed to limit forgetting with different trade-offs. Inspired by the brain thalamocortical circuit, we introduce a simple algorithm that uses optimization at inference time to generate internal representations of the current task dynamically. The algorithm alternates between updating the model weights and a latent task embedding, allowing the agent to parse the stream of temporal experience into discrete events and organize learning about them. On a continual learning benchmark, it achieves competitive end average accuracy by mitigating forgetting, but importantly, by requiring the model to adapt through latent updates, it organizes knowledge into flexible structures with a cognitive interface to control them. Tasks later in the sequence can be solved through knowledge transfer as they become reachable within the well-factorized latent space. The algorithm meets many of the desiderata of an ideal continually learning agent in open-ended environments, and its simplicity suggests fundamental computations in circuits with abundant feedback control loops such as the thalamocortical circuits in the brain.