We present video-SALMONN 2, a family of audio-visual large language models that set new state-of-the-art (SOTA) results in video description and question answering (QA). Our core contribution is multi-round direct preference optimisation (MrDPO), paired with a caption-quality objective that jointly rewards completeness and factual accuracy. Unlike standard DPO with a fixed reference policy, MrDPO periodically refreshes the reference by bootstrapping from a newly re-initialised lightweight adapter trained on the latest preferences, avoiding reference staleness and enabling continual improvement. This strategy produces captions that are consistently more detailed and accurate than those from proprietary systems such as GPT-4o and Gemini-1.5 Pro. We further distil these gains by using our model to generate a high-quality video-caption corpus for supervised fine-tuning of new models, transferring benefits beyond captioning to strong performance on complex video-QA tasks. Across widely used audio-visual and visual-only understanding benchmarks (including Video-MME, WorldSense, AVUT, Video-Holmes, DailyOmni, MLVU, and LVBench), our 3B and 7B models achieve SOTA results at comparable scales, while the 72B model surpasses all other open-source systems. Our source code, models, and data are released at \href{https://github.com/bytedance/video-SALMONN-2}{https://github.com/bytedance/video-SALMONN-2}.

Latent-space monitors aim to detect undesirable behaviours in Large Language Models by leveraging their internal representations rather than relying solely on black-box outputs. These methods have shown promise in identifying behaviours such as deception and unsafe completions. However, these monitors may themselves become training signals, for example, by using problematic samples found in deployment to retrain models. This raises an important question: can models learn to evade such monitors? To evaluate this capability, we introduce RL-Obfuscation, in which LLMs are finetuned via reinforcement learning to evade latent-space monitors while maintaining their blackbox behaviour. We apply RL-Obfuscation to Language Models ranging from 7B to 14B parameters and evaluate their Evasion Success Rate against a suite of monitors. We find that token-level monitors are highly vulnerable to this attack while more holistic monitors, such as max-pooling or attention-based probes, remain robust. Moreover, for these vulnerable monitors, models trained to evade a single static monitor can generalise to evade other unseen monitors. We also find that the models can be trained to conditionally bypass latent-space monitors on only certain inputs. Finally, we study how the models bypass these monitors and find that the model can learn to repurpose tokens to have different internal representations.

Recent advances in large language models (LLMs), particularly those enhanced through reinforced post-training, have demonstrated impressive reasoning capabilities, as exemplified by models such as OpenAI o1 and DeepSeek-R1. However, these capabilities are predominantly benchmarked on domains like mathematical problem solving and code generation, leaving open the question of whether such reasoning skills generalize to complex real-world scenarios. In this paper, we introduce LocationReasoner, a benchmark designed to evaluate LLMs' reasoning abilities in the context of real-world site selection, where models must identify feasible locations by reasoning over diverse and complicated spatial, environmental, and logistic constraints. The benchmark covers carefully crafted queries of varying difficulty levels and is supported by a sandbox environment with in-house tools for constraint-based location search. Automated verification further guarantees the scalability of the benchmark, enabling the addition of arbitrary number of queries. Extensive evaluations on real-world site selection data from Boston, New York, and Tampa reveal that state-of-the-art reasoning models offer limited improvement over their non-reasoning predecessors in real-world contexts, with even the latest OpenAI o4 model failing on 30% of site selection tasks. Moreover, agentic strategies such as ReAct and Reflexion often suffer from over-reasoning, leading to worse outcomes than direct prompting. With key limitations of LLMs in holistic and non-linear reasoning highlighted, we release LocationReasoner to foster the development of LLMs and agents capable of robust, grounded reasoning in real-world decision-making tasks. Codes and data for our benchmark are available at https://github.com/miho-koda/LocationReasoner.

Training large language models typically demands extensive GPU memory and substantial financial investment, which poses a barrier for many small- to medium-sized teams. In this paper, we propose a full-parameter pre-training and fine-tuning framework based on block coordinate descent (BCD), enhanced with engineering optimizations, to enable efficient training of large-scale models on cost-effective RTX 4090, A100 and A800 GPU clusters. Under identical hardware configurations, we reduce the training cost of a 7B model to 33% on A100/A800 and only 2.6% on RTX 4090, compared to standard full-parameter training. It also enables large models previously restricted to A100 clusters to be trained on RTX 4090 without degrading performance. BCD achieves comparable or better accuracy than full-parameter and fine-tuning methods at most cases, with lower GPU consumption and improved hardware utilization.

The rapid adoption of Large Language Models(LLMs) for code generation has transformed software development, yet little attention has been given to how security vulnerabilities evolve through iterative LLM feedback. This paper analyzes security degradation in AI-generated code through a controlled experiment with 400 code samples across 40 rounds of "improvements" using four distinct prompting strategies. Our findings show a 37.6% increase in critical vulnerabilities after just five iterations, with distinct vulnerability patterns emerging across different prompting approaches. This evidence challenges the assumption that iterative LLM refinement improves code security and highlights the essential role of human expertise in the loop. We propose practical guidelines for developers to mitigate these risks, emphasizing the need for robust human validation between LLM iterations to prevent the paradoxical introduction of new security issues during supposedly beneficial code "improvements".

The "Reason-Then-Respond" paradigm, enhanced by Reinforcement Learning, has shown great promise in advancing Multimodal Large Language Models. However, its application to the video domain has led to specialized models that excel at either question answering (QA) or captioning tasks, but struggle to master both. Naively combining reward signals from these tasks results in mutual performance degradation, which we attribute to a conflict between their opposing task natures. To address this challenge, we propose a novel training framework built upon two intermediate proxy tasks: DarkEventInfer, which presents videos with masked event segments, requiring models to infer the obscured content based on contextual video cues; and MixVidQA, which presents interleaved video sequences composed of two distinct clips, challenging models to isolate and reason about one while disregarding the other. These proxy tasks compel the model to simultaneously develop both holistic, divergent understanding and precise, convergent reasoning capabilities. Extensive experiments show that VidBridge-R1 achieves significant performance gains on both QA and captioning within one model, demonstrating the efficacy of our approach in fostering more generaliz-able and powerful video understanding models. The release of OpenAI o1/o3 (Jaech et al., 2024) and DeepSeek-R1 (Guo et al., 2025) has introduced a novel Reason-Then-Respond paradigm to the development of large language models (LLMs), which significantly enhances model performance through test-time scaling. Inspired by this approach, a growing body of research (Team et al., 2025; Chen et al., 2025a; Shen et al., 2025; Deng et al., 2025; Xia et al., 2025; Y ao et al., 2025) has extended this paradigm to multimodal large language models (MLLMs). By leveraging reinforcement learning (RL), particularly the Group Relative Policy Optimization (GRPO) algorithm (Shao et al., 2024), these studies have achieved promising results in image-based reasoning tasks. Recently, several studies (Feng et al., 2025; Zhang et al., 2025b; Chen et al., 2025e;f) have begun to explore the application of the Reason-Then-Respond paradigm in the video modality. Some efforts focus on enhancing question answering (QA) capabilities in general or reasoning scenarios (Li et al., 2025b; Dang et al., 2025), while some other works concentrate solely on improving video captioning performance (Li et al., 2025a; Meng et al., 2025a). This work was conducted during the author's internship at Kling Team, Kuaishou Technology A key advantage of MLLMs lies in their versatility, enabling strong performance across diverse tasks simultaneously. It is therefore undesirable to enhance reasoning capabilities at the expense of generalizability by over-specializing the model in a single task. To preserve generality in both QA and captioning tasks, an intuitive approach is to combine the reward signals from them during training.

Vision-Language-Action (VLA) models have advanced autonomous driving, but existing benchmarks still lack scenario diversity, reliable action-level annotation, and evaluation protocols aligned with human preferences. To address these limitations, we introduce DriveAction, the first action-driven benchmark specifically designed for VLA models, comprising 16,185 QA pairs generated from 2,610 driving scenarios. DriveAction leverages real-world driving data proactively collected by drivers of autonomous vehicles to ensure broad and representative scenario coverage, offers high-level discrete action labels collected directly from drivers' actual driving operations, and implements an action-rooted tree-structured evaluation framework that explicitly links vision, language, and action tasks, supporting both comprehensive and task-specific assessment. Our experiments demonstrate that state-of-the-art vision-language models (VLMs) require both vision and language guidance for accurate action prediction: on average, accuracy drops by 3.3% without vision input, by 4.1% without language input, and by 8.0% without either. Our evaluation supports precise identification of model bottlenecks with robust and consistent results, thus providing new insights and a rigorous foundation for advancing human-like decisions in autonomous driving.

Long-term forecasting of chaotic systems remains a fundamental challenge due to the intrinsic sensitivity to initial conditions and the complex geometry of strange attractors. Conventional approaches, such as reservoir computing, typically require training data that incorporates long-term continuous dynamical behavior to comprehensively capture system dynamics. While advanced deep sequence models can capture transient dynamics within the training data, they often struggle to maintain predictive stability and dynamical coherence over extended horizons. Here, we propose PhyxMamba, a framework that integrates a Mamba-based state-space model with physics-informed principles to forecast long-term behavior of chaotic systems given short-term historical observations on their state evolution. We first reconstruct the attractor manifold with time-delay embeddings to extract global dynamical features. After that, we introduce a generative training scheme that enables Mamba to replicate the physical process. It is further augmented by multi-patch prediction and attractor geometry regularization for physical constraints, enhancing predictive accuracy and preserving key statistical properties of systems. Extensive experiments on simulated and real-world chaotic systems demonstrate that PhyxMamba delivers superior forecasting accuracy and faithfully captures essential statistics from short-term historical observations.

In this work, we present the first study to explore inference-time scaling on table reasoning tasks. We develop and evaluate two post-training strategies to enable inference-time scaling: distillation from frontier model reasoning traces and reinforcement learning with verifiable rewards (RLVR). For distillation, we introduce a large-scale dataset of reasoning traces generated by DeepSeek-R1, which we use to fine-tune LLMs into the Table-R1-SFT model. For RLVR, we propose task-specific verifiable reward functions and apply the GRPO algorithm to obtain the Table-R1-Zero model. We evaluate our Table-R1-series models across diverse table reasoning tasks, including short-form QA, fact verification, and free-form QA. Notably, the Table-R1-Zero model matches or exceeds the performance of GPT-4.1 and DeepSeek-R1, while using only a 7B-parameter LLM. It also demonstrates strong generalization to out-of-domain datasets. Extensive ablation and qualitative analyses reveal the benefits of instruction tuning, model architecture choices, and cross-task generalization, as well as emergence of essential table reasoning skills during RL training.

While Ukrainian NLP has seen progress in many texts processing tasks, emotion classification remains an underexplored area with no publicly available benchmark to date. In this work, we introduce EmoBench-UA, the first annotated dataset for emotion detection in Ukrainian texts. Our annotation schema is adapted from the previous English-centric works on emotion detection (Mohammad et al., 2018; Mohammad, 2022) guidelines. The dataset was created through crowdsourcing using the Toloka.ai platform ensuring high-quality of the annotation process. Then, we evaluate a range of approaches on the collected dataset, starting from linguistic-based baselines, synthetic data translated from English, to large language models (LLMs). Our findings highlight the challenges of emotion classification in non-mainstream languages like Ukrainian and emphasize the need for further development of Ukrainian-specific models and training resources.