multitask learning

This work addresses the semantic gap between natural language instructions and robotic physical actions to enhance the naturalness and reliability of human-robot collaboration. We propose the first four-dimensional taxonomy for language-conditioned robotic manipulation—comprising reward shaping, policy learning, neurosymbolic AI, and foundation model–driven approaches—and systematically analyze their fundamental limitations in generalization and safety. Integrating large language models (LLMs), vision-language models (VLMs), neurosymbolic reasoning, and multimodal semantic parsing, we develop a unified analytical framework spanning semantic extraction, environmental assessment, and auxiliary task design. Our analysis rigorously characterizes the performance boundaries of each paradigm for the first time, establishing theoretical foundations and concrete technical pathways toward safe, generalizable, and interpretable language-driven robotic systems.

This work addresses the challenge of limited generalization across diverse gaits in legged robot locomotion policies, stemming from inadequate goal representation. We propose a unified policy framework conditioned on future foot-landing sequences—discrete contact patterns—rather than explicit gait labels or kinematic targets. This is the first approach to encode contact sequences as policy conditions, leveraging shared underlying dynamics across gaits to enable end-to-end generalization to walking, trotting, bounding, and other gaits within a single policy. Our method employs imitation learning, using a model predictive controller (MPC) as the expert teacher, and jointly trains a contact-state encoder with a neural policy network. Evaluated on bipedal and quadrupedal simulation platforms, our approach achieves over 40% higher out-of-distribution gait transfer success rates compared to baseline methods, demonstrating significantly improved robustness and cross-gait generalization capability.

To address the challenges of manual customization, labor-intensive workflows, and poor reusability of single-task models in multi-task modeling, this paper proposes a no-retraining automated fusion framework. Methodologically, it introduces Adaptive Knowledge Fusion (AKF), the first approach leveraging model decomposition and Transformer-based attention to decouple heterogeneous single-task models into composable components, enabling plug-and-play, cross-architecture and cross-task integration via adaptive attention mechanisms. Crucially, AKF eliminates reliance on explicit task relationship modeling or joint training. Experiments across three benchmark datasets demonstrate that the framework achieves multi-task performance comparable to end-to-end trained models, while substantially reducing modeling complexity and improving inference efficiency by 37%.

This paper addresses three key challenges in cooperative multi-agent reinforcement learning: (1) heavy reliance on human priors for task decomposition, (2) low sample efficiency, and (3) opaque credit assignment. We propose the first end-to-end, model-agnostic framework for learning task decomposition. Our method automatically discovers an optimal symbolic task decomposition—formalized as a Reward Machine—directly from environment interactions, dynamically partitioning the global task into assignable subtasks while jointly optimizing agent policies. Key contributions include: (1) fully automated discovery of task structure without manual specification of decomposition hierarchies; (2) integration of task-conditioned neural architectures with the formal semantics of Reward Machines, thereby ensuring both policy generalizability and interpretable, semantically grounded credit assignment; and (3) substantial improvements in sample efficiency and convergence stability across multiple benchmarks featuring strongly coupled agent dynamics.

Large pretrained models suffer from concept entanglement in task vector editing, leading to uncontrollable behavioral modulation. To address this, we propose an interpretable task vector decomposition framework that, for the first time, disentangles a task vector into two orthogonal components: a shared subspace component encoding general-purpose knowledge and a task-specific component capturing exclusive concepts. Decomposition is achieved via parameter-space projection and invariant subspace identification, enforced by an orthogonality constraint to ensure clean separation. Our method is broadly applicable across modalities—demonstrated on image classification, diffusion models, and large language models (LLMs). It improves multi-task fusion accuracy by 5% in vision tasks, preserves generation quality during style mixing in diffusion models, and reduces toxicity by 47% in LLMs without degrading general capabilities. This work overcomes the fundamental limitation of conventional vector arithmetic—its lack of concept-level controllability.

Multi-task reinforcement learning (RL) faces challenges including high-dimensional state spaces, sparse rewards, and poor policy robustness. To address these, this work introduces category theory as a foundational framework for RL—marking the first systematic application of categorical principles to model the structure and composability of Markov decision processes (MDPs) axiomatically, revealing their functorial nature and natural transformation mechanisms. We propose a functional RL framework grounded in universal properties, enabling provably sound skill abstraction, structure-preserving policy transfer, and composable task decomposition, reuse, and reconstruction. Evaluated on complex robotic manipulation tasks, our approach significantly improves cross-task generalization and sample efficiency, mitigates the curse of dimensionality, and enhances policy robustness. This work establishes a novel theoretical foundation and practical methodology for verifiable, composable agent learning.

This work addresses the performance degradation commonly observed in merged multi-task models due to parameter interference, which often results in inferior performance compared to single-task experts. Existing dynamic routing approaches typically require additional training or prior knowledge of task identities, limiting their practicality. To overcome these limitations, the authors propose a training-free, task-ID-agnostic dynamic routing mechanism that leverages a few task-specific support samples to construct low-rank task manifolds via singular value decomposition (SVD). Routing decisions are made by evaluating the projection residuals of test samples onto these manifolds. The method seamlessly integrates with lightweight subspace- or mask-based merging strategies and demonstrates consistent performance gains across multiple computer vision and natural language processing benchmarks, effectively narrowing the gap with single-task expert models even when task identities are unknown at inference time.

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.

This study presents the first systematic and fair comparison between model merging and joint multi-task reinforcement learning, investigating whether the former can effectively substitute for the latter. Leveraging the AppWorld benchmark, the authors train Qwen3-8B expert models of varying difficulty using the LOOP framework and apply merging techniques such as TIES and RAM+. Through analyses involving task vector cosine similarity, support set overlap, and upper/lower-bound calibration, they find that merged models achieve task completion rates comparable to jointly trained models, with no significant performance differences across merging methods. Further geometric analysis reveals that this parity stems from task vectors exhibiting substantial parameter support overlap (~65%) yet near-orthogonal directions, as evidenced by low cosine similarities ranging from 0.06 to 0.10.

This work addresses the lack of a clear theoretical understanding of the relationship between goal-conditioned reinforcement learning (GCRL) and unsupervised skill learning (MISL). It unifies these two paradigms through a control-maximization perspective, revealing that while three common formulations of GCRL are not equivalent, they all reflect a policy’s sensitivity to future trajectories with respect to goal instructions. Furthermore, the study establishes a rigorous theoretical bound linking the diversity objective in MISL to downstream GCRL performance. Through information-theoretic analysis and mutual information optimization, this work provides a principled foundation for unsupervised pre-training, offering practical guidance for selecting pre-training objectives aligned with specific downstream tasks to enhance goal-reaching performance.

Existing training and evaluation frameworks lack controllable shared latent structures, making it difficult to systematically analyze how agents leverage cross-task experience to improve decision-making. This work proposes LatentGym—the first benchmark suite grounded in real, controllable latent variables—that decouples exploration (acquiring latent knowledge) from exploitation (applying learned knowledge), thereby enabling fine-grained assessment of cross-task adaptation mechanisms. Experiments demonstrate that the platform can uncover the root causes of large language models’ failures in cross-task generalization, validate the efficacy of post-training on task sequences, and elucidate how design choices such as inter-task feedback critically shape learning dynamics and generalization performance.