This work explores the challenge of building "Machines that Can Remember", framing long-term memory as the problem of efficient ultra-long context modeling. We argue that this requires three key properties: sparsity, random-access flexibility, and length generalization. To address ultra-long-context modeling, we leverage Hierarchical Sparse Attention (HSA), a novel attention mechanism that satisfies all three properties. We integrate HSA into Transformers to build HSA-UltraLong, which is an 8B-parameter MoE model trained on over 8 trillion tokens and is rigorously evaluated on different tasks with in-domain and out-of-domain context lengths to demonstrate its capability in handling ultra-long contexts. Results show that our model performs comparably to full-attention baselines on in-domain lengths while achieving over 90% accuracy on most in-context retrieval tasks with contexts up to 16M. This report outlines our experimental insights and open problems, contributing a foundation for future research in ultra-long context modeling. Figure 1: Despite being pre-trained with an 8K context window and mid-trained up to 32K, HSA-UltraLong achieves near-perfect accuracy on S-NIAH even at a 16M-token context length. The red dashed line at 32K marks the boundary between in-domain (left) and out-of-domain (right).

Despite their impressive capabilities, Large Language Models (LLMs) exhibit unwanted uncertainty, a phenomenon where a model changes a previously correct answer into an incorrect one when re-prompted. This behavior undermines trust and poses serious risks in high-stakes domains. In this work, we investigate the mechanisms that drive this phenomenon. We adapt the Needle-in-a-Haystack retrieval framework and integrate a Flip-style re-evaluation prompt to simulate realistic answer-flipping scenarios. We find that retrieval heads are not primarily responsible for avoiding uncertainty. Instead, we identify a small set of non-retrieval attention heads that disproportionately attend to misleading tokens in uncertain contexts. Masking these heads yields significant improvements, reducing flip behavior by up to 15% without introducing incoherence or overcorrection. However, when tested for downstream tasks, we observe trade-offs with flip behavior. Our findings contribute to the growing field of mechanistic interpretability and present a simple yet effective technique for mitigating uncertainty-driven failure modes in LLMs.

Large language models (LLMs) have achieved impressive performance across natural language processing (NLP) tasks. As real-world applications increasingly demand longer context windows, continued pretraining and supervised fine-tuning (SFT) on long-context data has become a common approach. While the effects of data length in continued pretraining have been extensively studied, their implications for SFT remain unclear. In this work, we systematically investigate how SFT data length influences LLM behavior on short-context tasks. Counterintuitively, we find that long-context SFT improves short-context performance, contrary to the commonly observed degradation from long-context pretraining. To uncover the underlying mechanisms of this phenomenon, we first decouple and analyze two key components, Multi-Head Attention (MHA) and Feed-Forward Network (FFN), and show that both independently benefit from long-context SFT. We further study their interaction and reveal a knowledge preference bias: long-context SFT promotes contextual knowledge, while short-context SFT favors parametric knowledge, making exclusive reliance on long-context SFT suboptimal. Finally, we demonstrate that hybrid training mitigates this bias, offering explainable guidance for fine-tuning LLMs.

Recent work has identified retrieval heads, a subset of attention heads responsible for retrieving salient information in long-context language models (LMs), as measured by their copy-paste behavior in Needlein-a-Haystack tasks. In this paper, we introduce QRHead (Query-Focused Retrieval Head), an improved set of attention heads that enhance retrieval from long context. We identify QRHead by aggregating attention scores with respect to the input query, using a handful of examples from real-world tasks (e.g., long-context QA). We further introduce QRRetriever, an efficient and effective retriever that uses the accumulated attention mass of QRHead as retrieval scores. We use QRRetriever for long-context reasoning by selecting the most relevant parts with the highest retrieval scores. On multi-hop reasoning tasks LongMemEval and CLIPPER, this yields over 10% performance gains over full context and outperforms strong dense retrievers. We also evaluate QRRetriever as a re-ranker on the BEIR benchmark and find that it achieves strong zero-shot performance, outperforming other LLM-based re-rankers such as RankGPT. Further analysis shows that both the query-context attention scoring and task selection are crucial for identifying QRHead with strong downstream utility. Overall, our work contributes a general-purpose retriever and offers interpretability insights into the long-context capabilities of LMs.

Memes are widely used in online social interactions, providing vivid, intuitive, and often humorous means to express intentions and emotions. Existing dialogue datasets are predominantly limited to either manually annotated or pure-text conversations, lacking the expressiveness and contextual nuance that multimodal interactions provide.To address these challenges, we introduce MemeCMD, an automatically generated Chinese Multi-turn Dialogue dataset with contextually retrieved memes. Our dataset combines a large-scale, MLLM-annotated meme library with dialogues auto-generated by dual agents across diverse scenarios. We introduce a retrieval framework and adaptive threshold to ensure contextually relevant, naturally spaced meme usage. Experiments demonstrate the effectiveness of our approach in generating contextually appropriate and diverse meme-incorporated dialogues, offering a scalable and privacy-preserving resource for advancing multimodal conversational AI.

We present TopClustRAG, a retrieval-augmented generation (RAG) system developed for the LiveRAG Challenge, which evaluates end-to-end question answering over large-scale web corpora. Our system employs a hybrid retrieval strategy combining sparse and dense indices, followed by K-Means clustering to group semantically similar passages. Representative passages from each cluster are used to construct cluster-specific prompts for a large language model (LLM), generating intermediate answers that are filtered, reranked, and finally synthesized into a single, comprehensive response. This multi-stage pipeline enhances answer diversity, relevance, and faithfulness to retrieved evidence. Evaluated on the FineWeb Sample-10BT dataset, TopClustRAG ranked 2nd in faithfulness and 7th in correctness on the official leaderboard, demonstrating the effectiveness of clustering-based context filtering and prompt aggregation in large-scale RAG systems.

Large Language Models (LLMs) frequently show distracted attention due to irrelevant information in the input, which severely impairs their long-context capabilities. Inspired by recent studies on the effectiveness of retrieval heads in long-context factutality, we aim at addressing this distraction issue through improving such retrieval heads directly. We propose Multi-Document Attention Focusing (MuDAF), a novel method that explicitly optimizes the attention distribution at the head level through contrastive learning. According to the experimental results, MuDAF can significantly improve the long-context question answering performance of LLMs, especially in multi-document question answering. Extensive evaluations on retrieval scores and attention visualizations show that MuDAF possesses great potential in making attention heads more focused on relevant information and reducing attention distractions.

In this work, we introduce a novel approach called Scaling to Emphasize Attention for Long-context retrieval (SEAL), which enhances the retrieval performance of large language models (LLMs) over extended contexts. Previous studies have shown that each attention head in LLMs has a unique functionality and collectively contributes to the overall behavior of the model. Similarly, we observe that specific heads are closely tied to long-context retrieval, showing positive or negative correlation with retrieval scores. Built on this insight, we propose a learning-based mechanism using zero-shot generated data to emphasize these heads, improving the model's performance in long-context retrieval tasks. By applying SEAL, we can achieve significant improvements in in-domain retrieval performance, including document QA tasks from LongBench, and considerable improvements in outof-domain cases. Additionally, when combined with existing training-free context extension techniques, SEAL extends the context limits of LLMs while maintaining highly reliable outputs, opening new avenues for research in this field. Large Language Models (LLMs) (Brown et al. (2020), Radford et al. (2019), Touvron et al. (2023)) are capable of rapidly generating high-quality answers to a wide range of questions by leveraging the diverse knowledge embedded in their vast number of parameters. However, in-depth analyses have revealed a common issue known as hallucination (Shuster et al. (2021), Lin et al. (2021), Ji et al. (2023)), where the models confidently produce inaccurate answers. Figure 1: Overview of the proposed SEAL and corresponding retrieval score improvements for LongChat-7B-v1.5-32K These approaches have significantly improved the reliability of LLMs by enabling them to reference existing information during generation. However, this trend has also highlighted a key limitation of LLMs: the constraint of their context window length.

Long-context LLMs are increasingly in demand for applications such as retrieval-augmented generation. To defray the cost of pretraining LLMs over long contexts, recent work takes an approach of synthetic context extension: fine-tuning LLMs with synthetically generated long-context data in a post-training stage. However, it remains unclear how and why this synthetic context extension imparts abilities for downstream long-context tasks. In this paper, we investigate fine-tuning on synthetic data for three long-context tasks that require retrieval and reasoning. We vary the realism of "needle" concepts to be retrieved and diversity of the surrounding "haystack" context, from using LLMs to construct synthetic documents to using templated relations and creating symbolic datasets. We find that models trained on synthetic data fall short of the real data, but surprisingly, the mismatch can be interpreted and even predicted in terms of a special set of attention heads that are responsible for retrieval over long context, retrieval heads (Wu et al., 2024). The retrieval heads learned on synthetic data have high overlap with retrieval heads learned on real data, and there is a strong correlation between the recall of heads learned and the downstream performance of a model. Furthermore, with attention knockout and activation patching, we mechanistically show that retrieval heads are necessary and explain model performance, although they are not totally sufficient. Our results shed light on how to interpret synthetic data fine-tuning performance and how to approach creating better data for learning real-world capabilities over long contexts.

The performance of Retrieval-Augmented Generation (RAG) systems in information retrieval is significantly influenced by the characteristics of the documents being processed. In this study, the structured nature of textbooks, the conciseness of articles, and the narrative complexity of novels are shown to require distinct retrieval strategies. A comparative evaluation of multiple documentsplitting methods reveals that the Recursive Character Splitter outperforms the Token-based Splitter in preserving contextual integrity. A novel evaluation technique is introduced, utilizing an open-source model to generate a comprehensive dataset of question-and-answer pairs, simulating realistic retrieval scenarios to enhance testing efficiency and metric reliability. The evaluation employs weighted scoring metrics, including SequenceMatcher, BLEU, METEOR, and BERT Score, to assess the system's accuracy and relevance. This approach establishes a refined standard for evaluating the precision of RAG systems, with future research focusing on optimizing chunk and overlap sizes to improve retrieval accuracy and efficiency. 2