Many applications of large language models (LLMs) require long-context understanding, but models continue to struggle with such tasks. We hypothesize that conventional next-token prediction training could contribute to this, because each token is assigned equal weight. Yet, intuitively, the amount of context needed to predict the next token accurately varies greatly across different data. To reflect this, we propose various novel token-weighting schemes that assign different weights to each training token in the loss, thereby generalizing existing works. For this, we categorize token-weighting methods using a two-step framework which compares the confidences of a long-context and short-context model to score tokens. We evaluate all methods on multiple long-context understanding tasks and show that non-uniform loss weights are helpful to improve the long-context abilities of LLMs. Different short-context models can be used effectively for token scoring, including models that are much smaller than the long-context model that is trained. All in all, this work contributes to a better understanding of the trade-offs long-context language modeling faces and provides guidelines for model steering via loss-weighting based on empirical evidence. The code can be found on Github.

Long-context modeling is one of the critical capabilities of language AI for digesting and reasoning over complex information pieces. In practice, long-context capabilities are typically built into a pre-trained language model (LM) through a carefully designed context extension stage, with the goal of producing generalist long-context capabilities. In our preliminary experiments, however, we discovered that the current open-weight generalist long-context models are still lacking in practical long-context processing tasks. While this means perfectly effective long-context modeling demands task-specific data, the cost can be prohibitive. In this paper, we draw inspiration from how humans process a large body of information: a lossy retrieval stage ranks a large set of documents while the reader ends up reading deeply only the top candidates. We build an automatic data synthesis pipeline that mimics this process using short-context LMs. The short-context LMs are further tuned using these self-generated data to obtain task-specific longcontext capabilities. Similar to how pre-training learns from imperfect data, we hypothesize and further demonstrate that the short-context model can bootstrap over the synthetic data, outperforming not only long-context generalist models but also the retrieval and read pipeline used to synthesize the training data in realworld tasks such as long-context retrieval augmented generation. The field of Artificial Intelligence (AI) and Natural Language Processing (NLP) have made substantial progress in building and teaching neural language models (LMs) to understand and generate language (Radford et al., 2019; Brown et al., 2020; OpenAI, 2023; Anthropic, 2023; 2024; Touvron et al., 2023a;b; MetaAI et al., 2024). Large-scale deep learning has enabled large LMs to learn from massive amounts of human-generated text (Radford et al., 2019; Brown et al., 2020).

Despite the recent progress in long-context language models, it remains elusive how transformer-based models exhibit the capability to retrieve relevant information from arbitrary locations within the long context. This paper aims to address this question. Our systematic investigation across a wide spectrum of models reveals that a special type of attention heads are largely responsible for retrieving information, which we dub retrieval heads. We identify intriguing properties of retrieval heads:(1) universal: all the explored models with long-context capability have a set of retrieval heads; (2) sparse: only a small portion (less than 5\%) of the attention heads are retrieval. (3) intrinsic: retrieval heads already exist in models pretrained with short context. When extending the context length by continual pretraining, it is still the same set of heads that perform information retrieval. (4) dynamically activated: take Llama-2 7B for example, 12 retrieval heads always attend to the required information no matter how the context is changed. The rest of the retrieval heads are activated in different contexts. (5) causal: completely pruning retrieval heads leads to failure in retrieving relevant information and results in hallucination, while pruning random non-retrieval heads does not affect the model's retrieval ability. We further show that retrieval heads strongly influence chain-of-thought (CoT) reasoning, where the model needs to frequently refer back the question and previously-generated context. Conversely, tasks where the model directly generates the answer using its intrinsic knowledge are less impacted by masking out retrieval heads. These observations collectively explain which internal part of the model seeks information from the input tokens. We believe our insights will foster future research on reducing hallucination, improving reasoning, and compressing the KV cache.