Search AI Products and News

In the field of artificial intelligence, reward models are a critical component for aligning large language models (LLMs) with human feedback, but existing models face the issue of "reward hacking." These models often focus on superficial features, such as the length or format of responses, rather than identifying genuine quality metrics, such as factual accuracy and relevance. The root cause lies in standard training objectives failing to distinguish between spurious associations and true causal drivers present in the training data. This failure leads to fragile reward models (RMs), which generate misaligned policies.

Recently, the Qwen VLo multimodal large model was officially released, achieving significant advancements in image content understanding and generation, offering users a brand-new visual creation experience. According to the introduction, Qwen VLo has been comprehensively upgraded based on the advantages of the original Qwen-VL series models. The model not only can accurately understand the "world", but also can perform high-quality re-creation based on understanding, truly achieving the transition from perception to generation. Users can now access Qwen Chat (chat.qwen.ai)

Recently, an open-source RAG (Retrieval-Augmented Generation) engine called RAGFlow has attracted significant attention in the industry. This enterprise-level AI tool, based on deep document understanding, provides companies with a brand-new solution for processing complex documents and achieving precise question answering through its powerful multimodal data processing capabilities and efficient workflows. RAGFlow: The Pioneer of Deep Document Understanding RAGFlow is a fully open-source RAG engine that focuses on deep document understanding, aiming to help businesses and individuals extract valuable information from massive amounts of unstructured data.

The Meta AI research team has made another breakthrough in the artificial intelligence field, officially releasing the new video understanding model V-JEPA2 (Video Joint Embedding Predictive Architecture2) on June 11, 2025. Led by Meta's chief AI scientist Yann LeCun, this model has opened up new possibilities in video understanding and physical world modeling with its innovative self-supervised learning technology and zero-shot robot control capabilities.

In recent years, the field of artificial intelligence has undergone tremendous changes, especially with large language models (LLMs) making remarkable progress in multi-modal tasks. These models have demonstrated strong potential in their ability to understand and generate language, but most current multi-modal models still adopt auto-regressive (AR) architectures, which limit inference processes to be rather monotonous and lacking flexibility. To address this limitation, a research team from The University of Hong Kong and Huawei Noah's Ark Lab has proposed a brand new model – FUDOKI, aiming to break these constraints. The core innovation of FUDOKI is

In the recent Fiction.Live benchmark test, Gemini2.5Pro performed excellently in understanding and reproducing complex stories and backgrounds, leading ahead of OpenAI's o3 model. This test goes far beyond traditional "needle-in-a-haystack" tasks, focusing on a model's ability to handle deep semantics and context-dependent information within large contexts. According to the test data, when the context window length reaches 192,000 tokens (approximately 144,000 words), the performance of the o3 model drops sharply, while

Recently, the Beijing Academy of Artificial Intelligence (BAEI), in collaboration with Shanghai Jiao Tong University and other institutions, officially released a new generation of ultra-long video understanding model - Video-XL-2. The introduction of this model represents a significant breakthrough in the field of open-source ultra-long video understanding technology, injecting new vitality into the development of multimodal large models for understanding long video content. In terms of technical architecture, Video-XL-2 mainly consists of three core components: a visual encoder, the Dynamic Token Synthesis (DTS) module, and a large language model (LLM). The model adopts

Recently, Hume AI officially released its third-generation voice interaction model, EVI3. This new voice AI has drawn extensive attention in the industry due to its outstanding emotion understanding capabilities and personalized interaction experience. EVI3 can accurately identify emotions in users' voices and generate specific styles of sound and personality based on user preferences, marking a significant breakthrough in the field of emotional interaction and natural communication. Below is the latest information and in-depth analysis about EVI3 from AIbase. Experience it at: https://demo.

A recent paper jointly published by Tsinghua University and Shanghai Jiao Tong University challenges the widely held belief that pure reinforcement learning (RL) enhances large model reasoning capabilities. The research found that models incorporating reinforcement learning performed worse than their original counterparts in certain tasks.

ByteDance Research has announced the open-sourcing of ChatTS-14B, a 14-billion parameter large language model (LLM) specifically designed for understanding and reasoning with time series data. Released under the Apache2.0 license, ChatTS-14B's open-source release has garnered significant attention within the AI community, marking a substantial advancement in the intersection of time series analysis and generative AI. ChatTS-14B: An Intelligent Conversational Engine for Time Series. ChatTS-14B is based on Qwen2.5-1...

At the recently held Volcano Engine FORCE LINK AI Innovation Roadshow, Volcano Engine President Tan Dai officially announced the Doubao 1.5 Deep Thinking Model, marking the official launch of this innovative technology for enterprise services. Doubao 1.5 not only upgrades its basic functions but also introduces more intelligent application scenarios to help enterprises achieve more efficient digital transformation. Tan Dai stated at the launch that the Doubao 1.5 Deep Thinking Model can combine visual understanding to provide more innovative functions. For example, the model can generate...

OpenAI recently unveiled its latest AI model, o3, marking a significant advancement in AI's ability to understand and analyze images, particularly low-quality sketches and diagrams. Alongside o3, a smaller version, o4-mini, was also released, expanding OpenAI's offerings. The core functionality of the o3 model lies in its ability to 'think with images,' allowing users to upload diverse images, such as whiteboard sketches and complex charts, for in-depth AI analysis and discussion.

In the field of artificial intelligence, 3D vision and spatial understanding are becoming crucial for advancements in embodied AI, autonomous navigation, and virtual reality applications. In March 2025, Hangzhou Qunhe Technology announced the open-sourcing of its self-developed 3D vision large language model, SpatialLM, at the GTC2025 global conference, generating significant industry interest. This model, with its powerful spatial cognition capabilities and cost-effective data processing, offers revolutionary breakthroughs for robot training, architectural design, and AR/VR applications. AIbase is based on the latest information...

In geoscience, geological maps are crucial tools for understanding the Earth's surface and subsurface structures. However, interpreting these complex diagrams requires specialized knowledge and extensive experience. To enhance intelligence in this field, Microsoft Research Asia recently introduced GeoMap-Bench, a new benchmark dataset for evaluating the performance of multimodal large language models (MLLMs) in understanding geological maps. The launch of GeoMap-Bench marks a significant step forward in AI applications for geological map interpretation. Microsoft researchers, in collaboration with...

Hangzhou-based tech company, Qunhe Technology, has again garnered industry attention for its open-sourced spatial understanding model, SpatialLM, which was acknowledged by Google in a recent research paper. This model's innovation lies in its ability to enable robots to understand the geometrical relationships of the physical world through ordinary videos, marking a significant breakthrough in robot training. SpatialLM's core function is to convert videos captured by a phone into three-dimensional spatial layout information. Users simply need to record their home layout using their phone, and SpatialLM will...