LLMs struggle with performance in Retrieval Augmented Generation (RAG) when many documents are retrieved. Prior studies did not isolate if this is due to document quantity versus context length. This paper investigates how the number of documents affects LLM performance, while keeping context length constant. → The paper creates custom datasets from a multi-hop Question Answering dataset. → Document count is varied while context length and relevant information positions are kept constant. → Distractor documents are removed, and remaining documents are extended to maintain fixed context length. → Evaluations using Llama-3.1, Qwen2, and Gemma2 models show that increasing document count often degrades performance by up to 10 percent. → Qwen2 model is less impacted by document count. → Adding random, unrelated documents improves performance, unlike adding related but distracting documents. ---------------------------- Paper - arxiv. org/abs/2503.04388v1 Paper Title: "More Documents, Same Length: Isolating the Challenge of Multiple Documents in RAG"

LLMs hallucinate, mixing truthful and false information. This makes factuality alignment noisy during training because response-level preference learning cannot isolate factual errors. This paper introduces Mask-DPO. It uses sentence-level factuality masks during Direct Preference Optimization. Mask-DPO learns only from factually correct parts of preferred responses. It avoids penalizing factual content in dispreferred responses. 📌 Sentence-level masks in Mask-DPO refine preference learning. This method reduces noise from mixed factuality responses. 📌 Mask-DPO's fine-grained DPO significantly boosts factuality and generalization. It achieves this by reducing noisy feedback. ---------- Methods Explored in this Paper 🔧: → Mask-DPO uses sentence-level factuality annotations as masks. → It constructs preference pairs by ranking responses based on their sentence-level factuality scores. → During training, Mask-DPO applies masks to the DPO loss function. → These masks ignore incorrect sentences in preferred responses and correct sentences in dispreferred responses. → This fine-grained approach focuses learning on factual correctness at the sentence level. ---------------------------- Paper - arxiv. org/abs/2503.02846 Paper Title: "Mask-DPO: Generalizable Fine-grained Factuality Alignment of LLMs"