Large language models (LLMs) have demonstrated proficiency in solving complex problems across mathematics, scientific research, and software engineering. Chain-of-thought (CoT) prompting is pivotal in guiding models through intermediate reasoning steps before reaching conclusions. Reinforcement learning (RL) is another essential component that enables structured reasoning, allowing models to recognize and correct errors efficiently. Despite these advancements,…
Recent advancements in LLMs, such as the GPT series and emerging “o1” models, highlight the benefits of scaling training and inference-time computing. While scaling during training—by increasing model size and dataset volume—has been a well-established strategy, recent findings emphasize the advantages of inference-time scaling, where additional computational resources during testing improve output quality and task…
Open-vocabulary object detection (OVD) aims to detect arbitrary objects with user-provided text labels. Although recent progress has enhanced zero-shot detection ability, current techniques handicap themselves with three important challenges. They heavily depend on expensive and large-scale region-level annotations, which are hard to scale. Their captions are typically short and not contextually rich, which makes them…
Developing AI systems that learn from their surroundings during execution involves creating models that adapt dynamically based on new information. In-Context Reinforcement Learning (ICRL) follows this approach by allowing AI agents to learn through trial and error while making decisions. However, this method has significant challenges when applied to complex environments with various tasks. It…
Text-to-speech (TTS) technology has made significant strides in recent years, but challenges remain in creating natural, expressive, and high-fidelity speech synthesis. Many TTS systems struggle to replicate the nuances of human speech, such as intonation, emotion, and accent, often resulting in artificial-sounding voices. Additionally, precise voice cloning remains difficult, limiting the ability to generate personalized…
The International Mathematical Olympiad (IMO) is a globally recognized competition that challenges high school students with complex mathematical problems. Among its four categories, geometry stands out as the most consistent in structure, making it more accessible and well-suited for fundamental reasoning research. Automated geometry problem-solving has traditionally followed two primary approaches: algebraic methods, such as…
Large language models (LLMs) must align with human preferences like helpfulness and harmlessness, but traditional alignment methods require costly retraining and struggle with dynamic or conflicting preferences. Test-time alignment approaches using reward models (RMs) avoid retraining but face inefficiencies due to reliance on trajectory-level rewards, which evaluate full responses rather than guiding token-by-token generation. Existing…
In this tutorial, we demonstrate the workflow for fine-tuning Mistral 7B using QLoRA with Axolotl, showing how to manage limited GPU resources while customizing the model for new tasks. We’ll install Axolotl, create a small example dataset, configure the LoRA-specific hyperparameters, run the fine-tuning process, and test the resulting model’s performance. Step 1: Prepare the…
Large Language Models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks, particularly in mathematical problem-solving and coding applications. Research has shown a strong correlation between the length of reasoning chains and improved accuracy in problem-solving outcomes. However, they face significant challenges: while extended reasoning processes enhance problem-solving capabilities, they often lead to inefficient solutions.…