Vision-language-action models have gained significant attention for their ability to model multimodal sequences in embodied instruction following tasks. However, most existing models rely on causal attention, which we find suboptimal for processing sequences composed of interleaved segments from different modalities. In this paper, we introduce Astra, a novel Transformer architecture featuring trajectory attention and learnable action queries, designed to efficiently process segmented multimodal trajectories and predict actions for imitation learning. Furthermore, we propose a contrastive dynamics learning objective to enhance the model’s understanding of environment dynamics and multimodal alignment, complementing the primary behavior cloning objective. Through extensive experiments on three large-scale robot manipulation benchmarks, Astra demonstrates substantial performance improvements over previous models.

Existing vision-language planning methods perform well on short-horizon tasks but struggle with long-horizon reasoning in dynamic environments due to the difficulty of training models to generate high-quality reasoning processes. To address this, we propose Structured Preference Optimization (SPO), a framework that enhances reasoning and action selection for long-horizon task planning through structured evaluation and optimized training. SPO introduces: 1) Structured Preference Evaluation and Optimization, which evaluates reasoning chains across task relevance, historical consistency (as part of textual coherence), and image awareness (alignment with visual observations) to construct high-quality preference pairs; and 2) Curriculum-Guided Progressive Learning, enabling the model to adapt from simple to complex tasks, thereby improving generalization and robustness. To advance research in vision-language long-horizon task planning, we introduce ExtendaBench, a comprehensive benchmark covering 1,509 tasks across VirtualHome and Habitat 2.0, categorized into ultra-short, short, medium, and long tasks. Experimental results demonstrate that SPO significantly improves reasoning quality and final decision accuracy, outperforming prior methods on long-horizon tasks and underscoring the effectiveness of preference-driven optimization in vision-language task planning. Specifically, SPO achieves a +5.98% GCR and +4.68% SR improvement in VirtualHome and a +3.30% GCR and +2.11% SR improvement in Habitat over the best-performing baselines.

The natural language generation domain has witnessed great success thanks to Transformer models. Although they have achieved state-of-the-art generative quality, they often neglect generative diversity. Prior attempts to tackle this issue suffer from either low model capacity or over-complicated architectures. Some recent methods employ the VAE framework to enhance diversity, but their latent variables fully depend on the input context, restricting exploration of the latent space. In this paper, we introduce VOLTA, a framework that elevates generative diversity by bridging Transformer with VAE via a more effective cross-attention-based connection, departing from conventional embedding concatenation or summation. Additionally, we propose integrating InfoGAN-style latent codes to enable input-independent variability, further diversifying the generation. Moreover, our framework accommodates discrete inputs alongside its existing support for continuous inputs. We perform comprehensive experiments with two types of Transformers on six datasets from three different NLG tasks to show that our approach can significantly improve generative diversity while maintaining generative quality.

Reinforcement learning (RL) has emerged as a promising approach to fine-tune offline pretrained GPT-2 model in task-oriented dialogue (TOD) systems. In order to obtain human-like online interactions while extending the usage of RL, building pretrained user simulators (US) along with dialogue systems (DS) and facilitating jointly fine-tuning via RL becomes prevalent. However, joint training brings distributional shift problem caused by compounding exposure bias. Existing methods usually iterative update US and DS to ameliorate the ensued non-stationarity problem, which could lead to sub-optimal policy and less sample efficiency. To take a step further for tackling the problem, we introduce an Offline-to-oNline Co-Evolutional (ONCE) framework, which enables bias-aware concurrent joint update for RL-based fine-tuning whilst takes advantages from GPT-2 based end-to-end modeling on US and DS. Extensive experiments demonstrate that ONCE builds high-quality loops of policy learning and dialogues data collection, and achieves state-of-the-art online and offline evaluation results on MultiWOZ2.1 dataset. Opensourced code will be implemented with Mindspore (MS, 2022) and released on our homepage.