Open coding agents are commonly trained through supervised fine-tuning on interaction trajectories whose final patches pass the benchmark test suite. However, this outcome-level filtering criterion ignores process quality: trajectories that eventually succeed may still contain inefficient interaction patterns such as repetitive file inspection, unstable tool usage, or delayed modification behavior, all of which become part of the imitation target. We propose (Trajectory Quality Score), a structural scoring framework for coding-agent trajectories that operates directly on trajectory structure, requires no LLM judge, and runs in linear time. characterizes trajectory quality through structural signals organized around structured convergence: information focus, intent–action alignment, and policy dynamics. Using 155,144 successful trajectories from a public open-license corpus, we rank trajectories by and fine-tune coding agents on filtered subsets. Models trained on top-ranked trajectories consistently outperform those trained on random or bottom-ranked subsets on . Our results suggest that process-level trajectory quality is an important factor in coding-agent supervision and that structural trajectory filtering provides an effective and scalable alternative to LLM-based data selection.

While reinforcement learning with verifiable reward (RLVR) enhances reasoning in large language models, its efficacy in long-horizon multimodal tasks remains underexplored. Visual Sudoku puzzles pose a representative challenge, requiring multi-step deductive reasoning over both local constraints and global dependency chains. To this end, we introduce FG-Sudoku, a fine-grained benchmark categorized by grid size and required deductive steps. Evaluations reveal that frontier MLLMs fail on complex long-range dependencies (e.g., 9×9 puzzles). During post-training, this near-zero success rate causes severe reward sparsity, halting standard RLVR. We propose Adaptive Curriculum Group Relative Policy Optimization (AdaC-GRPO) to overcome this. Unlike rigid or prior-based schedules, AdaC-GRPO dynamically constructs training batches by tracking the model’s "Zone of Proximal Development" via online rollout accuracy. Experiments demonstrate that AdaC-GRPO enables Qwen2.5-VL-7B to approach the performance of Gemini-2.5-Flash while bridging 62.2% of the overall accuracy gap to Gemini-2.5-Pro, with its generalizability further validated across InternVL-3.5 and Qwen3-VL.

Hybrid attention dominates frontier LLMs, yet Vision Transformers (ViTs) in multimodal LLMs lack a satisfactory hybrid design, with no consensus on why certain attention patterns work better. To fill this gap, we study ViT attention heads and find they differentiate into object- and background-specialist roles, a pattern most pronounced under full attention; we call this Semantic Head Specialization (SHS). We propose SHS-Index to quantify this specialization, show that it distinguishes full-attention from chunk-window ViTs, and find that it strongly tracks downstream benchmark performance. We then identify three structural factors that shape SHS—window interaction, token serialization, and local softmax allocation—and use them as design principles for hybrid attention. Guided by these factors, we design Ariadne Attention, a hybrid that matches full attention on 22 image and video tasks at 6.5 × less attention compute. Our findings establish head specialization as a measurable property for diagnosing and designing principled hybrid ViT attention at the multimodal-LLM scale.

Supervised fine-tuning (SFT) enables large language models to align with training data for better performance in many aspects. Nevertheless, the gap between the distribution of current datasets from human annotations or model generations and the real-world data distribution heavily limits the capacities and potentials of models. As a result, we propose a new SFT technique, ATLANTIS, to bridge the gap. We adopt importance sampling to estimate the optimal data distribution in the real world from existing training datasets because the former is hard to sample from. Furthermore, we introduce an extra small model and reference model to estimate the sampling ratio through the probability gap between them. We evaluate our method with benchmarks in knowledge & understanding and preference aspects. The experiment results prove that ATLANTIS can bring consistent and significant improvements to models’ performance. What’s more, our method can be flexibly transferred among models with different structures. Our analyses demonstrate that our method is well-compatible with other SFT techniques to further enhance models’ capacities and has great potential to be combined with existing training frameworks.

Multimodal punchlines, which involve humor or sarcasm conveyed in image-caption pairs, are a popular way of communication on online multimedia platforms. With the rapid development of multimodal large language models (MLLMs), it is essential to assess their ability to effectively comprehend these punchlines. However, existing benchmarks on punchline comprehension suffer from three major limitations: 1) language shortcuts that allow models to solely rely on text, 2) lack of question diversity, and 3) narrow focus on a specific domain of multimodal content (e.g., cartoon). To address these limitations, we introduce a multimodal Punchline comprehension Benchmark, named PunchBench, which is tailored for accurate and comprehensive evaluation of punchline comprehension. To enhance the evaluation accuracy, we generate synonymous and antonymous captions by modifying original captions, which mitigates the impact of shortcuts in the captions. To provide a comprehensive evaluation, PunchBench incorporates diverse question formats and image-captions from various domains. On this basis, we conduct extensive evaluations and reveal a significant gap between state-of-the-art MLLMs and humans in punchline comprehension. To improve punchline comprehension, we propose Simple-to-Complex Chain-of-Question (SC-CoQ) strategy, enabling the models to incrementally address complicated questions by first mastering simple ones. SC-CoQ effectively enhances the performance of various MLLMs on PunchBench, surpassing in-context learning and chain-of-thought.

Financial prediction from Monetary Policy Conference (MPC) calls is a new yet challenging task, which targets at predicting the price movement and volatility for specific financial assets by analyzing multimodal information including text, video, and audio. Although the existing work has achieved great success using cross-modal transformer blocks, it overlooks the potential external financial knowledge, the varying contributions of different modalities to financial prediction, as well as the innate relations among different financial assets. To tackle these limitations, we propose a novel Modal-Adaptive kNowledge-enhAnced Graph-basEd financial pRediction scheme, named MANAGER. Specifically, MANAGER resorts to FinDKG to obtain the external related knowledge for the input text. Meanwhile, MANAGER adopts BEiT-3 and Hidden-unit BERT (HuBERT) to extract the video and audio features, respectively. Thereafter, MANAGER introduces a novel knowledge-enhanced cross-modal graph that fully characterizes the semantic relations among text, external knowledge, video and audio, to adaptively utilize the information in different modalities, with ChatGLM2 as the backbone. Extensive experiments on a publicly available dataset Monopoly verify the superiority of our model over cutting-edge methods.

Recently, there is a surge in interest surrounding video large language models (Video LLMs). However, existing benchmarks fail to provide a comprehensive feedback on the temporal perception ability of Video LLMs. On the one hand, most of them are unable to distinguish between different temporal aspects (e.g., speed, direction) and thus cannot reflect the nuanced performance on these specific aspects. On the other hand, they are limited in the diversity of task formats (e.g., only multi-choice QA), which hinders the understanding of how temporal perception performance may vary across different types of tasks. Motivated by these two problems, we propose the TempCompass benchmark, which introduces a diversity of temporal aspects and task formats. To collect high-quality test data, we devise two novel strategies: (1) In video collection, we construct conflicting videos that share the same static content but differ in a specific temporal aspect, which prevents Video LLMs from leveraging single-frame bias or language priors. (2) To collect the task instructions, we propose a paradigm where humans first annotate meta-information for a video and then an LLM generates the instruction. We also design an LLM-based approach to automatically and accurately evaluate the responses from Video LLMs. Based on TempCompass, we comprehensively evaluate 9 state-of-the-art (SOTA) Video LLMs and 3 Image LLMs, and reveal the discerning fact that these models exhibit notably poor temporal perception ability.

Large-scale video-language pre-training has made remarkable strides in advancing video-language understanding tasks. However, the heavy computational burden of video encoding remains a formidable efficiency bottleneck, particularly for long-form videos. These videos contain massive visual tokens due to their inherent 3D properties and spatiotemporal redundancy, making it challenging to capture complex temporal and spatial relationships. To tackle this issue, we propose an efficient method called TEmporal-Spatial Token Aggregation (TESTA). TESTA condenses video semantics by adaptively aggregating similar frames, as well as similar patches within each frame. TESTA can reduce the number of visual tokens by 75% and thus accelerate video encoding. Building upon TESTA, we introduce a pre-trained video-language model equipped with a divided space-time token aggregation module in each video encoder block. We evaluate our model on five datasets for paragraph-to-video retrieval and long-form VideoQA tasks. Experimental results show that TESTA improves computing efficiency by 1.7 times, and achieves significant performance gains from its scalability in processing longer input frames, e.g., +13.7 R@1 on QuerYD and +6.5 R@1 on Condensed Movie.

Previous work has demonstrated the viability of applying deep learning techniques in the financial area. Recently, the task of stock embedding learning has been drawing attention from the research community, which aims to represent the characteristics of stocks with distributed vectors that can be used in various financial analysis scenarios. Existing approaches for learning stock embeddings either require expert knowledge, or mainly focus on the textual part of information corresponding to individual temporal movements. In this paper, we propose to model stock properties as the combination of internal attributes and relational attributes, which takes into consideration both the time-invariant properties of individual stocks and their movement patterns in relation to the market. To learn the two types of attributes from financial news and transaction data, we design several training objectives based on contrastive learning to extract and separate the long-term and temporary information in the data that are able to counter the inherent randomness of the stock market. Experiments and further analyses on portfolio optimization reveal the effectiveness of our method in extracting comprehensive stock information from various data sources.

Pre-trained self-supervised models such as BERT have achieved striking success in learning sequence representations, especially for natural language processing. These models typically corrupt the given sequences with certain types of noise, such as masking, shuffling, or substitution, and then try to recover the original input. However, such pre-training approaches are prone to learning representations that are covariant with the noise, leading to the discrepancy between the pre-training and fine-tuning stage. To remedy this, we present ContrAstive Pre-Training (CAPT) to learn noise invariant sequence representations. The proposed CAPT encourages the consistency between representations of the original sequence and its corrupted version via unsupervised instance-wise training signals. In this way, it not only alleviates the pretrain-finetune discrepancy induced by the noise of pre-training, but also aids the pre-trained model in better capturing global semantics of the input via more effective sentence-level supervision. Different from most prior work that focuses on a particular modality, comprehensive empirical evidence on 11 natural language understanding and cross-modal tasks illustrates that CAPT is applicable for both language and vision-language tasks, and obtains surprisingly consistent improvement, including 0.6% absolute gain on GLUE benchmarks and 0.8% absolute increment on NLVR2.