Filtered to continual learning · clear filter
This paper develops a new approach to understanding videos by predicting when specific events or evidence occur within the video. Practitioners working on video analysis and AI models might care about this research because it could lead to more accurate and robust video understanding systems.
This paper introduces RAGU, an open-source GraphRAG engine that improves large language models with structured knowledge by separating extraction and consolidation, and trains a compact extractor that outperforms larger models on knowledge-graph construction and GraphRAG tasks. Practitioners might care because RAGU can efficiently generate more accurate and complete context for tasks like factoid-level evidence recall and multi-hop question answering.
This paper creates a platform to help large language models generate code for data pipelines, which can then be edited and used to automate data processing workflows. Practitioners might care about this because it can help reduce the time and cost of developing and maintaining these pipelines.
This paper introduces a framework called DeepSearch-Evolve, which helps train self-improving web agents by iteratively refining their performance using their own experience. Practitioners might care because this approach can lead to more efficient and effective agents that can learn from their own mistakes.
This paper proposes a new pruning method for coding agents that prunes tool outputs directly inside the agent, rather than relying on a separate code classifier, and shows it can save up to 39% of tokens while preserving task quality.
This paper introduces Mage-Flow, a compact model for generating and editing high-resolution images, which can be trained efficiently and deployed on a single GPU. Practitioners might care about the potential applications of this model in interactive image editing and generation tasks.
This paper introduces a vision-language-action model that can perform mobile manipulation tasks in unseen environments with minimal training data, and how it can be scaled up to achieve better performance. Practitioners might care about this model for building robots that can adapt to new tasks with minimal fine-tuning.
This paper develops a large audio language model that can answer questions with specific timestamps, improving its ability to understand long audio recordings. Practitioners in audio and speech recognition may care about this development as it enables more accurate and context-specific information retrieval from audio data.
This paper develops a multilingual speech recognition model for underrepresented languages, addressing the issue of data scarcity and uneven performance, and shows promising results in controlled comparisons.
This paper develops a new framework for evaluating and selecting document sets for AI agents, considering the interactions between documents, and proposes a training-free method that achieves the best downstream generation performance with fewer documents and search rounds.
This paper introduces a new method called REBASE that allows for training-free in-context segmentation, enabling the introduction of new object categories at inference time. Practitioners might care because it eliminates the need for retraining and reduces memory overhead, making it a more efficient approach for real-world applications.
This paper develops a new framework, SciForma, to generate scientific diagrams that accurately represent research logic, which is crucial for scientific communication and methodology validation. Practitioners can benefit from SciForma's ability to ensure structural fidelity in diagram generation, which can improve the accuracy and reliability of scientific research.
This paper introduces Tencent WorkBuddy Bench, a benchmark for coding agents that tests their performance across multiple domains, including code, web, office, and security. Practitioners might care about this benchmark because it provides a standardized way to evaluate and compare the performance of coding agents.
This paper develops a framework called NexForge that helps train more capable artificial agents by automatically generating a large number of tasks and training data, without requiring a lot of manual setup. Practitioners might care because it can improve the performance of their own agent models.
This paper investigates using hypernetworks for large-scale knowledge injection into language models, a technique that can improve their ability to answer factual questions. Practitioners may care because it could lead to more accurate and scalable language models for applications like customer service or question-answering systems.
This paper proposes a new method for fine-tuning vision-language models on robot demonstrations to improve their performance on real-world tasks, by preventing the overwrite of pre-trained representations and aligning language and action predictions. Practitioners may care about this work because it aims to improve the generalizability and robustness of vision-language-action policies in real-world applications.
This paper develops a method to transform video foundation models' representations into compact, reconstruction-capable, and generation-friendly video latents, which can be used in various generative modeling tasks. Practitioners can use VideoRAE to improve the performance of their models by leveraging the semantic and spatio-temporal structure captured by the frozen video foundation encoder.
This paper introduces AutoIndex, a framework that learns to transform raw documents into representations for retrieval systems, allowing for more flexible and effective indexing. Practitioners may care about AutoIndex because it can improve the quality of search results in complex information retrieval tasks.
This paper proposes a new method called Distilled Reinforcement Learning that improves large language model post-training by providing fine-grained guidance to transfer new knowledge from a teacher model to a student model. Practitioners might care because it outperforms standard reinforcement learning and on-policy distillation methods in terms of knowledge transfer and model performance.
This paper introduces Robostral Navigate, a vision-language model that enables robots to navigate using only a single monocular RGB camera, making it more scalable and cost-effective for deployment across various robotic platforms. Practitioners might care about this because it can simplify navigation tasks for robots in real-world environments.
This paper proposes a method to improve on-policy distillation by reusing pre-collected teacher data, allowing for faster training without interacting with the environment. Practitioners may care about this technique because it enables scalable and efficient distillation of complex agent models.
This paper develops a method called Experience Distillation that allows agents to learn from their own interaction histories without needing additional environment interactions, making learning more sample-efficient. Practitioners might care about this because it can improve the performance of agents in complex environments with limited resources.
This paper evaluates language models and deep-research agents at predicting football match outcomes before kickoff, using a dynamic benchmark that can be reused for future leagues. Practitioners can learn from the results to improve their own models' performance in similar tasks.
This paper develops a service-oriented framework, JoyNexus, to efficiently train and deploy Vision-Language-Action models across multiple tenants, improving resource utilization and reducing costs. Practitioners may care about JoyNexus for its potential to streamline the training process and make VLA models more accessible.
This paper proposes a method to improve the explainability of natural-language autoencoders by making it harder for models to manipulate the explanations, and shows that this approach can increase the reliability of activation explanations and improve AI safety.
This paper addresses a common issue in agentic language models where they may over-call tools, leading to suboptimal performance. A practitioner might care about this problem because it can negatively impact the model's ability to provide accurate and efficient responses.
This paper creates a benchmark to test how AI systems respond when an authority figure tries to get an unwilling subordinate to complete a task, and it explores how different levels of authority affect the outcome. Practitioners might care because it helps them understand and manage the complex interactions between AI systems.
This paper introduces UI2App, a benchmark to evaluate the ability of large language models to infer interaction behavior from screenshots of web applications, which is crucial for real-world development workflows. Practitioners might care about this research because it highlights the limitations of current visual-driven approaches and the need for better interaction inference capabilities in web application generation.
This paper explores how to make AI systems understand and generate humor, particularly in visual formats like memes and comics, and why this is important for developing more human-like AI that can understand and create humor. Practitioners might care because humor is a key aspect of human communication and understanding.
This paper develops a mathematical theory for reliability in large language models (LLMs) by analyzing the behavior of verification gates in cascades. It explains why adding more gates can sometimes help, sometimes hurt, or plateau the reliability of LLMs, and provides a way to estimate this reliability using repeated verdicts or other methods. Practitioners might care about this because it can help them design more reliable and robust LLMs.
This paper investigates how the quality of training data affects the performance of text-to-video models, and it provides a controlled testbed to study this relationship. Practitioners may care about the implications of data quality on model performance and training efficiency, especially when developing large-scale text-to-video models.
This paper evaluates the robustness of machine learning models used in plasma diagnostics for fusion devices, which are often affected by sensor failures and data corruption. Practitioners caring about the reliability of these models will be interested in the techniques developed here to improve their performance under challenging conditions.
This paper proposes a new method for dataset distillation that focuses on aligning the final outcome of training, rather than imitating the training process. Practitioners might care about this method because it can improve the accuracy of model training on limited data.