In-car conversational AI is becoming increasingly critical as autonomous vehicles and smart assistants gain widespread adoption. Yet, existing datasets fail to capture the spontaneous disfluencies such as hesitations, false starts, repetitions, and self-corrections that characterize real driver-AI dialogs. To address this, we introduce DiscoDrive, a synthetic corpus of 3500 multi-turn dialogs across seven automotive domains, generated using a two-stage, prompt-driven pipeline that dynamically integrates disfluencies during synthesis. We show that DiscoDrive is effective both as a training resource, enabling DialoGPT-Medium and T5-Base to match or exceed KVRET-trained models on the MultiWOZ 2.2 and Schema-Guided Dialogue (SGD) relevant test sets (BLEU-4 improvements of 0.26 to 0.61; METEOR +2.10; ROUGE-L +3.48; BERTScore F1 improvements of 1.35 to 3.48), and as a data augmentation resource in low-resource scenarios, delivering additional gains of up to BLEU-4 +0.38, METEOR +1.95, ROUGE-L +2.87, and BERTScore F1 +4.00 when combined with 10 percent of KVRET. Human evaluations further confirm that dialogs sampled from DiscoDrive are rated higher than KVRET's human-collected dialogs in naturalness (3.8 vs 3.6) and coherence (4.1 vs 4.0), and are perceived as more context-appropriate than leading post-hoc methods (such as LARD), without compromising clarity. DiscoDrive fills a critical gap in existing resources and serves as a versatile corpus for both training and augmenting conversational AI, enabling robust handling of real-world, disfluent in-car interactions.

Large-scale reinforcement learning with verifiable rewards (RLVR) has demonstrated its effectiveness in harnessing the potential of large language models (LLMs) for single-turn reasoning tasks. In realistic reasoning scenarios, LLMs can often utilize external tools to assist in task-solving processes. However, current RL algorithms inadequately balance the models' intrinsic long-horizon reasoning capabilities and their proficiency in multi-turn tool interactions. To bridge this gap, we propose Agentic Reinforced Policy Optimization (ARPO), a novel agentic RL algorithm tailored for training multi-turn LLM-based agents. Through preliminary experiments, we observe that LLMs tend to exhibit highly uncertain behavior, characterized by an increase in the entropy distribution of generated tokens, immediately following interactions with external tools. Motivated by this observation, ARPO incorporates an entropy-based adaptive rollout mechanism, dynamically balancing global trajectory sampling and step-level sampling, thereby promoting exploration at steps with high uncertainty after tool usage. By integrating an advantage attribution estimation, ARPO enables LLMs to internalize advantage differences in stepwise tool-use interactions. Our experiments across 13 challenging benchmarks in computational reasoning, knowledge reasoning, and deep search domains demonstrate ARPO's superiority over trajectory-level RL algorithms. Remarkably, ARPO achieves improved performance using only half of the tool-use budget required by existing methods, offering a scalable solution for aligning LLM-based agents with real-time dynamic environments. Our code and datasets are released at https://github.com/dongguanting/ARPO

Contrastive Language-Image Pretraining has demonstrated remarkable zero-shot generalization by aligning visual and textual modalities in a shared embedding space. However, when continuously fine-tuned on diverse tasks, CLIP suffers from catastrophic forgetting and degradation of its embedding alignment, undermining its zero-shot capabilities. In this work, we propose Gradient Null Space Projection (GNSP), an efficient continual learning method that projects task-specific gradients onto the null space of previously learned knowledge. This orthogonal projection mathematically prevents interference with previous tasks without relying on rehearsal or architectural modification. Furthermore, to preserve the inherent generalization property of CLIP, we introduce knowledge distillation and combine it with a modality alignment preservation loss inspired by CLIP pre-training to stabilize the structure of the multimodal embedding space during fine-tuning. On the MTIL benchmark consisting of 11 tasks, our method achieved SOTA performance on both the Average and Last key metrics. More importantly, experiments show that our method successfully maintains the original modality gap and cross-modal retrieval performance of CLIP, confirming its effectiveness in maintaining a robust visual-language space throughout the continual learning process.

Processing long-context inputs with large language models presents a significant challenge due to the enormous memory requirements of the Key-Value (KV) cache during inference. Existing KV cache compression methods exhibit noticeable performance degradation when memory is reduced by more than 85%. Additionally, strategies that leverage GPU-CPU collaboration for approximate attention remain underexplored in this setting. We propose HCAttention, a heterogeneous attention computation framework that integrates key quantization, value offloading, and dynamic KV eviction to enable efficient inference under extreme memory constraints. The method is compatible with existing transformer architectures and does not require model fine-tuning. Experimental results on the LongBench benchmark demonstrate that our approach preserves the accuracy of full-attention model while shrinking the KV cache memory footprint to 25% of its original size. Remarkably, it stays competitive with only 12.5% of the cache, setting a new state-of-the-art in LLM KV cache compression. To the best of our knowledge, HCAttention is the first to extend the Llama-3-8B model to process 4 million tokens on a single A100 GPU with 80GB memory.

Figure 1: Ag2x2 enables zero-shot acquisition of bimanual manipulation skills without relying on expert demonstrations or engineered rewards. The framework operates in two key stages: (left) learning coordination-aware visual representations directly from human manipulation videos (shown in sequential frames of cooking with highlighted hand) while preserving critical hand position data despite domain differences; and (right) leveraging these representations to acquire diverse bimanual manipulation skills in simulation autonomously, demonstrated through multiple Franka robot arms performing sequential steps of various tasks including cabinet opening (top row), door manipulation (middle row), and rope handling (bottom row). Abstract -- Bimanual manipulation, fundamental to human daily activities, remains a challenging task due to its inherent complexity of coordinated control. Recent advances have enabled zero-shot learning of single-arm manipulation skills through agent-agnostic visual representations derived from human videos; however, these methods overlook crucial agent-specific information necessary for bimanual coordination, such as end-effector positions. We propose Ag2x2, a computational framework for bimanual manipulation through coordination-aware visual representations that jointly encode object states and hand motion patterns while maintaining agent-agnosticism. Extensive experiments demonstrate that Ag2x2 achieves a 73.5% success rate across 13 diverse bimanual tasks from Bi-DexHands and PerAct This performance outperforms baseline methods and even surpasses the success rate of policies trained with expert-engineered rewards. Furthermore, we show that representations learned through Ag2x2 can be effectively leveraged for imitation learning, establishing a scalable pipeline for skill acquisition without expert supervision. Ziyin Xiong and Yinghan Chen contributed equally to this work.

Aberrant protein-protein interactions (PPIs) underpin a plethora of human diseases, and disruption of these harmful interactions constitute a compelling treatment avenue. Advances in computational approaches to PPI prediction have closely followed progress in deep learning and natural language processing. In this review, we outline the state-of-the-art for sequence-based PPI prediction methods and explore their impact on target identification and drug discovery. We begin with an overview of commonly used training data sources and techniques used to curate these data to enhance the quality of the training set. Subsequently, we survey various PPI predictor types, including traditional similarity-based approaches, and deep learning-based approaches with a particular emphasis on the transformer architecture. Finally, we provide examples of PPI prediction in systems-level proteomics analyses, target identification, and design of therapeutic peptides and antibodies. We also take the opportunity to showcase the potential of PPI-aware drug discovery models in accelerating therapeutic development.

Diffusion-based deep generative models have emerged as powerful tools for inverse materials design. Yet, many existing approaches overlook essential chemical constraints such as oxidation state balance, which can lead to chemically invalid structures. Here we introduce CrysVCD (Crystal generator with Valence-Constrained Design), a modular framework that integrates chemical rules directly into the generative process. CrysVCD first employs a transformer-based elemental language model to generate valence-balanced compositions, followed by a diffusion model to generate crystal structures. The valence constraint enables orders-of-magnitude more efficient chemical valence checking, compared to pure data-driven approaches with post-screening. When fine-tuned on stability metrics, CrysVCD achieves 85% thermodynamic stability and 68% phonon stability. Moreover, CrysVCD supports conditional generation of functional materials, enabling discovery of candidates such as high thermal conductivity semiconductors and high-$κ$ dielectric compounds. Designed as a general-purpose plugin, CrysVCD can be integrated into diverse generative pipeline to promote chemical validity, offering a reliable, scientifically grounded path for materials discovery.

Structural drawings are widely used in many fields, e.g., mechanical engineering, civil engineering, etc. In civil engineering, structural drawings serve as the main communication tool between architects, engineers, and builders to avoid conflicts, act as legal documentation, and provide a reference for future maintenance or evaluation needs. They are often organized using key elements such as title/subtitle blocks, scales, plan views, elevation view, sections, and detailed sections, which are annotated with standardized symbols and line types for interpretation by engineers and contractors. Despite advances in software capabilities, the task of generating a structural drawing remains labor-intensive and time-consuming for structural engineers. Here we introduce a novel generative AI-based method for generating structural drawings employing a large language model (LLM) agent. The method incorporates a retrieval-augmented generation (RAG) technique using externally-sourced facts to enhance the accuracy and reliability of the language model. This method is capable of understanding varied natural language descriptions, processing these to extract necessary information, and generating code to produce the desired structural drawing in AutoCAD. The approach developed, demonstrated and evaluated herein enables the efficient and direct conversion of a structural drawing's natural language description into an AutoCAD drawing, significantly reducing the workload compared to current working process associated with manual drawing production, facilitating the typical iterative process of engineers for expressing design ideas in a simplified way.

Recent advances in large language models (LLMs) have highlighted the potential of reinforcement learning with verifiable rewards (RLVR) to enhance reasoning capabilities through extended output sequences. However, traditional RL frameworks face inefficiencies when handling ultra-long outputs due to long-tail sequence distributions and entropy collapse during training. To address these challenges, we propose an Ultra-Long Output Reinforcement Learning (UloRL) approach for advancing large language models' reasoning abilities. Specifically, we divide ultra long output decoding into short segments, enabling efficient training by mitigating delays caused by long-tail samples. Additionally, we introduce dynamic masking of well-Mastered Positive Tokens (MPTs) to prevent entropy collapse. Experimental results demonstrate the effectiveness of our approach. On the Qwen3-30B-A3B model, RL with segment rollout achieved 2.06x increase in training speed, while RL training with 128k-token outputs improves the model's performance on AIME2025 from 70.9\% to 85.1\% and on BeyondAIME from 50.7\% to 61.9\%, even surpassing Qwen3-235B-A22B with remarkable gains. These findings underscore the potential of our methods to advance the reasoning capabilities of LLMs with ultra-long sequence generation. We will release our code and model for further use by the community.

Answer Set Programming (ASP) is a powerful paradigm for non-monotonic reasoning. Recently, large language models (LLMs) have demonstrated promising capabilities in logical reasoning. Despite this potential, current evaluations of LLM capabilities in ASP are often limited. Existing works normally employ overly simplified ASP programs, do not support negation, disjunction, or multiple answer sets. Furthermore, there is a lack of benchmarks that introduce tasks specifically designed for ASP solving. To bridge this gap, we introduce ASPBench, a comprehensive ASP benchmark, including three ASP specific tasks: ASP entailment, answer set verification, and answer set computation. Our extensive evaluations on ASPBench reveal that while 14 state-of-the-art LLMs, including \emph{deepseek-r1}, \emph{o4-mini}, and \emph{gemini-2.5-flash-thinking}, perform relatively well on the first two simpler tasks, they struggle with answer set computation, which is the core of ASP solving. These findings offer insights into the current limitations of LLMs in ASP solving. This highlights the need for new approaches that integrate symbolic reasoning capabilities more effectively. The code and dataset are available at https://github.com/HomuraT/ASPBench.