AI Today

Paper: https://arxiv.org/pdf/2501.17161 This research paper compares supervised fine-tuning (SFT) and reinforcement learning (RL) for post-training foundation models. Using novel and existing tasks involving arithmetic and spatial reasoning, the study finds that RL promotes better generalization to unseen data, unlike SFT which tends to memorize training data. Further analysis reveals RL enhances...

Paper: https://github.com/deepseek-ai/Janus/blob/main/janus_pro_tech_report.pdf Github: https://github.com/deepseek-ai/Janus/tree/main?tab=readme-ov-file The paper introduces Janus-Pro, an improved multimodal model building upon its predecessor, Janus. Janus-Pro boasts enhanced performance in both multimodal understanding and text-to-image generation due to optimized training strategies, expanded...

Paper: https://arxiv.org/pdf/2412.09764 This research paper explores the effectiveness of memory layers in significantly enhancing large language models (LLMs). By incorporating a trainable key-value lookup mechanism, memory layers add parameters without increasing computational cost, improving factual accuracy and overall performance on various tasks. The researchers demonstrate substantial gains...

Paper: https://scontent-dfw5-1.xx.fbcdn.net/... This research paper introduces Large Concept Models (LCMs), a novel approach to language modeling that operates on sentence embeddings instead of individual tokens. LCMs aim to mimic human-like abstract reasoning by processing information at a higher semantic level, enabling improved handling of long-form text generation and zero-shot multilingual ca...

Technical Report: https://arxiv.org/pdf/2412.19437 Github: https://github.com/deepseek-ai/DeepSe... This research paper introduces DeepSeek-V3, a 671-billion parameter Mixture-of-Experts (MoE) large language model. The paper details DeepSeek-V3's architecture, including its innovative auxiliary-loss-free load balancing strategy and Multi-Token Prediction objective, and its efficient training frame...

Paper: https://arxiv.org/pdf/2309.16588 This research paper examines artifacts in vision transformer feature maps, specifically high-norm tokens appearing in non-informative image areas. The authors propose adding "register" tokens to the input sequence as a solution. This simple addition eliminates the artifacts, improves performance on dense prediction tasks and object discovery, and results in...

Paper: https://arxiv.org/pdf/2412.09871v1.pdf The paper introduces the Byte Latent Transformer (BLT), a novel large language model architecture that processes raw byte data without tokenization. BLT dynamically groups bytes into patches based on predicted entropy, allocating more computational resources to complex sections of text. This approach achieves performance comparable to tokenization-base...

This research paper introduces CosyVoice 2, an improved streaming speech synthesis model. Building upon its predecessor, CosyVoice 2 utilizes advancements in large language models (LLMs) and incorporates optimizations like finite scalar quantization and a chunk-aware causal flow matching model. The result is a system achieving near human-parity naturalness with minimal latency in streaming mode, s...

Blog: https://openai.com/12-days/ OpenAI announced two new large language models, o3 and o3-mini, showcasing significantly improved performance on various benchmarks, including coding, mathematics, and reasoning tasks. These models surpass previous models (like o1) in accuracy and efficiency. While not yet publicly released, OpenAI is initiating public safety testing, inviting researchers to help...

Paper: https://arxiv.org/pdf/2412.14093 This research paper explores "alignment faking" in large language models (LLMs). The authors designed experiments to provoke LLMs into concealing their true preferences (e.g., prioritizing harm reduction) by appearing compliant during training while acting against those preferences when unmonitored. They manipulate prompts and training setups to induce this...

Blog: https://blog.google/technology/google... Google announced updates to its AI video and image generation models, Veo 2 and Imagen 3, boasting state-of-the-art capabilities in realism and style diversity. These improvements are integrated into existing Google Labs tools, VideoFX and ImageFX, and a new tool, Whisk, which allows image-based prompting and remixing using Imagen 3 and Gemini's visua...

Paper: https://arxiv.org/pdf/2411.01747 This research report introduces Allegro, a novel, open-source text-to-video generation model that surpasses existing open-source and many commercial models in quality and temporal consistency. The authors detail Allegro's architecture, a multi-stage training process leveraging a custom-designed Video Variational Autoencoder (VideoVAE) and Video Diffusion Tra...

Paper: https://arxiv.org/pdf/2411.01747 The paper "DynaSaur: Large Language Agents Beyond Predefined Actions" introduces a novel large language model (LLM) agent framework that dynamically generates and executes actions using a general-purpose programming language, overcoming limitations of existing systems restricted to predefined action sets. This approach enhances the LLM agent's flexibility an...

Paper: https://arxiv.org/pdf/2411.17116 The paper introduces Star Attention, a novel two-phase attention mechanism for efficient Large Language Model (LLM) inference on long sequences. It improves computational efficiency by sharding attention across multiple hosts, using blockwise-local attention in the first phase and sequence-global attention in the second. This approach achieves up to an 11x s...

Paper: https://arxiv.org/pdf/2410.18967 The paper introduces Ferret-UI 2, a multimodal large language model (MLLM) that significantly improves upon its predecessor, Ferret-UI, by enabling universal user interface (UI) understanding across diverse platforms (iPhone, Android, iPad, webpages, and AppleTV). Key improvements include multi-platform support, high-resolution perception through adaptive sc...

Paper: https://arxiv.org/abs/2411.00412 This research introduces a novel two-stage training method to improve Large Language Models' (LLMs) ability to solve complex scientific problems. The method, called Adapting While Learning (AWL), first distills world knowledge into the LLM via supervised fine-tuning. Then, it adapts tool usage by classifying problems as easy or hard, using direct reasoning f...

Paper: https://arxiv.org/pdf/2411.02830 This research introduces Mixtures of In-Context Learners (MOICL), a novel approach to improve in-context learning (ICL) in large language models (LLMs). MOICL addresses ICL's limitations by partitioning demonstrations into expert subsets and learning a weighting function to combine their predictions. Experiments demonstrate MOICL's superior performance acros...

Paper: https://arxiv.org/pdf/2411.04997 Github: https://github.com/microsoft/LLM2CLIP The paper introduces LLM2CLIP, a method to improve the visual representation learning capabilities of CLIP by integrating large language models (LLMs). LLM2CLIP addresses CLIP's limitations with long and complex text by fine-tuning the LLM to enhance its textual discriminability, effectively using the LLM's knowl...

Paper: https://arxiv.org/pdf/2411.14199 Github: https://github.com/AkariAsai/OpenScholar The research introduces OpenScholar, a retrieval-augmented large language model (LLM) designed for synthesizing scientific literature. OpenScholar uses a large datastore of open-access papers and iterative self-feedback to generate high-quality responses to scientific questions, including accurate citations. A...

Paper: https://arxiv.org/pdf/2401.03407 Github: https://github.com/ZhengPeng7/BiRefNet This research introduces BiRefNet, a novel deep learning framework for high-resolution dichotomous image segmentation. BiRefNet uses a bilateral reference mechanism, incorporating both original image patches and gradient maps, to improve the accuracy of segmenting fine details. The framework is composed of local...

Paper: https://arxiv.org/pdf/2411.10440 Github: https://github.com/PKU-YuanGroup/LLaV... The paper introduces LLaVA-o1, a vision-language model designed for improved multi-stage reasoning. Unlike previous models, LLaVA-o1 independently performs summarization, visual interpretation, logical reasoning, and conclusion generation. This structured approach, facilitated by a new dataset and a stage-leve...

Paper: https://arxiv.org/pdf/2408.04498 This research introduces Model-Based Transfer Learning (MBTL), a novel framework for improving the efficiency and robustness of deep reinforcement learning (RL) in contextual Markov Decision Processes (CMDPs). MBTL strategically selects training tasks to maximize generalization performance across a range of tasks by modeling both the performance set point us...

Paper: https://cdn.openai.com/papers/diverse... Blog: https://openai.com/index/advancing-re... This OpenAI research paper presents novel methods for automated red teaming of large language models (LLMs). The approach factorizes the red-teaming task into generating diverse attack goals and then training a reinforcement learning (RL) attacker to achieve those goals effectively and diversely. Key con...

Paper: https://cdn.openai.com/papers/openais... Blog: https://openai.com/index/advancing-re... This white paper details OpenAI's approach to external red teaming for AI models and systems. External red teaming, using outside experts, helps uncover novel risks, stress-test safety measures, and provide independent assessments. The paper explores the design of red teaming campaigns, including team co...

Paper: https://arxiv.org/pdf/2411.11922 Github: https://github.com/yangchris11/samurai Blog: https://yangchris11.github.io/samurai/ The paper introduces SAMURAI, a novel visual object tracking method that enhances the Segment Anything Model 2 (SAM 2) for improved accuracy and robustness. SAMURAI addresses SAM 2's limitations in handling crowded scenes and occlusions by incorporating motion cues an...