Training Large Language Models (LLMs) efficiently at scale presents a formidable challenge, driven by their ever-increasing computational demands and the need for enhanced performance. In this work, we introduce Liger-Kernel, an open-sourced set of Triton kernels developed specifically for LLM training. With kernel optimization techniques like kernel operation fusing and input chunking, our kernels achieve on average a 20% increase in training throughput and a 60% reduction in GPU memory usage for popular LLMs compared to HuggingFace implementations. In addition, Liger-Kernel is designed with modularity, accessibility, and adaptability in mind, catering to both casual and expert users. Comprehensive benchmarks and integration tests are built in to ensure compatibility, performance, correctness, and convergence across diverse computing environments and model architectures. The source code is available under a permissive license at: github.com/linkedin/Liger-Kernel.

Each year, the lack of efficient data standardization and interoperability in cancer care contributes to the severe lack of timely and effective diagnosis, while constantly adding to the burden of cost, with cancer costs nationally reaching over $208 billion in 2023 alone. Traditional methods regarding clinical trial enrollment and clinical care in oncology are often manual, time-consuming, and lack a data-driven approach. This paper presents a novel framework to streamline standardization, interoperability, and exchange of cancer domains and enhance the integration of oncology-based EHRs across disparate healthcare systems. This paper utilizes advanced LLMs and Computer Engineering to streamline cancer clinical trials and discovery. By utilizing FHIR's resource-based approach and LLM-generated mCODE profiles, we ensure timely, accurate, and efficient sharing of patient information across disparate healthcare systems. Our methodology involves transforming unstructured patient treatment data, PDFs, free-text information, and progress notes into enriched mCODE profiles, facilitating seamless integration with our novel AI and ML-based clinical trial matching engine. The results of this study show a significant improvement in data standardization, with accuracy rates of our trained LLM peaking at over 92% with datasets consisting of thousands of patient data. Additionally, our LLM demonstrated an accuracy rate of 87% for SNOMED-CT, 90% for LOINC, and 84% for RxNorm codes. This trumps the current status quo, with LLMs such as GPT-4 and Claude's 3.5 peaking at an average of 77%. This paper successfully underscores the potential of our standardization and interoperability framework, paving the way for more efficient and personalized cancer treatment.

Large Language Models (LLMs) acquire extensive knowledge and remarkable abilities from extensive text corpora, making them powerful tools for various applications. To make LLMs more usable, aligning them with human preferences is essential. Existing alignment techniques, such as Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO), typically embed predefined preferences directly within the model's parameters. These methods, however, often result in a static alignment that can not account for the diversity of human preferences in practical applications. In response to this challenge, we propose an effective method, \textbf{MetaAlign}, which aims to help LLMs dynamically align with various explicit or implicit preferences specified at inference time. Experimental results show that LLMs optimized on our meticulously constructed MetaAlign Dataset can effectively align with any preferences specified at the inference stage, validating the feasibility of MetaAlign. We hope that our work can provide some insights into the alignment of language models.

We conducted experiments comparing the effectiveness of using simpler versus more complex dataset in different stages of the post-training process to better understand the optimal post-training strategy for large language models. Here we conduct comparison experiment on two kinds of instructions: simple instructions and specialized instructions, denoted as type 1 and type 2. As showen in Table 10, we observe that performing SFT on simpler instructions helps the model to establish a foundational level of instruction-following ability. This is reflected in moderate performance on AlpacaEval 2 (LC 16.25%, WR 17.62%) but lower performance on the more challenging Arena-Hard benchmark (WR 10.7%). When the model is fine-tuned on more specialized and complex data, there is a marginal improvement (LC 14.70%, WR 16.01%, Arena-Hard WR 14.7%), and the significant performance gains are achieved when DPO is applied after SFT. For example, SFT followed by DPO with complex, specialized instructions yields substantial improvements (LC 21.64%, WR 30.06%,

Tables are ubiquitous across various domains for concisely representing structured information. Empowering large language models (LLMs) to reason over tabular data represents an actively explored direction. However, since typical LLMs only support one-dimensional~(1D) inputs, existing methods often flatten the two-dimensional~(2D) table structure into a sequence of tokens, which can severely disrupt the spatial relationships and result in an inevitable loss of vital contextual information. In this paper, we first empirically demonstrate the detrimental impact of such flattening operations on the performance of LLMs in capturing the spatial information of tables through two elaborate proxy tasks. Subsequently, we introduce a simple yet effective positional encoding method, termed ``2D-TPE'' (Two-Dimensional Table Positional Encoding), to address this challenge. 2D-TPE enables each attention head to dynamically select a permutation order of tokens within the context for attending to them, where each permutation represents a distinct traversal mode for the table, such as column-wise or row-wise traversal. 2D-TPE effectively mitigates the risk of losing essential spatial information while preserving computational efficiency, thus better preserving the table structure. Extensive experiments across five benchmarks demonstrate that 2D-TPE outperforms strong baselines, underscoring the importance of preserving the table structure for accurate table comprehension. Comprehensive analysis further reveals the substantially better scalability of 2D-TPE to large tables than baselines.

Large Language Models (LLMs) have demonstrated remarkable capabilities in generating coherent text but remain limited by the static nature of their training data. Retrieval Augmented Generation (RAG) addresses this issue by combining LLMs with up-to-date information retrieval, but also expand the attack surface of the system. This paper investigates prompt injection attacks on RAG, focusing on malicious objectives beyond misinformation, such as inserting harmful links, promoting unauthorized services, and initiating denial-of-service behaviors. We build upon existing corpus poisoning techniques and propose a novel backdoor attack aimed at the fine-tuning process of the dense retriever component. Our experiments reveal that corpus poisoning can achieve significant attack success rates through the injection of a small number of compromised documents into the retriever's corpus. In contrast, backdoor attacks demonstrate even higher success rates but necessitate a more complex setup, as the victim must fine-tune the retriever using the attacker's poisoned dataset.

As language technologies become widespread, it is important to understand how variations in language affect reader perceptions -- formalized as the isolated causal effect of some focal language-encoded intervention on an external outcome. A core challenge of estimating isolated effects is the need to approximate all non-focal language outside of the intervention. In this paper, we introduce a formal estimation framework for isolated causal effects and explore how different approximations of non-focal language impact effect estimates. Drawing on the principle of omitted variable bias, we present metrics for evaluating the quality of isolated effect estimation and non-focal language approximation along the axes of fidelity and overlap. In experiments on semi-synthetic and real-world data, we validate the ability of our framework to recover ground truth isolated effects, and we demonstrate the utility of our proposed metrics as measures of quality for both isolated effect estimates and non-focal language approximations.

Large Language Models (LLMs) have revolutionized natural language processing and hold immense potential for advancing Artificial Intelligence. However, the core architecture of most mainstream LLMs--the Transformer--has inherent limitations in computational depth, rendering them theoretically incapable of solving many reasoning tasks that demand increasingly deep computations. Chain of Thought (CoT) prompting has emerged as a technique to address these architectural limitations, as evidenced by several theoretical studies. It offers a promising approach to solving complex reasoning tasks that were previously beyond the capabilities of these models. Despite its successes, CoT and its variants (such as Tree of Thought, Graph of Thought, etc.) rely on a "one-prompt-for-all" approach, using a single prompt structure (e.g., "think step by step") for a wide range of tasks--from counting and sorting to solving mathematical and algorithmic problems. This approach poses significant challenges for models to generate the correct reasoning steps, as the model must navigate through a vast prompt template space to find the appropriate template for each task. In this work, we build upon previous theoretical analyses of CoT to demonstrate how the one-promptfor-all approach can negatively affect the computability of LLMs. Our findings show that task-specific supervision is essential for navigating the prompt space accurately and achieving optimal performance. Through experiments with state-of-the-art LLMs, we reveal a gap in reasoning performance when supervision is applied versus when it is not. Our goal is to provide deeper insights into the mechanisms underlying CoT, offering guidance for the effective design of CoT variants. Additionally, we underscore the limitations of traditional "unsupervised" prompting methods, arguing that users of CoT cannot simply "sit back" and rely entirely on the model.

Natural Language Processing (NLP) research on AI Safety and social bias in AI has focused on safety for humans and social bias against human minorities. However, some AI ethicists have argued that the moral significance of nonhuman animals has been ignored in AI research. Therefore, the purpose of this study is to investigate whether there is speciesism, i.e., discrimination against nonhuman animals, in NLP research. First, we explain why nonhuman animals are relevant in NLP research. Next, we survey the findings of existing research on speciesism in NLP researchers, data, and models and further investigate this problem in this study. The findings of this study suggest that speciesism exists within researchers, data, and models, respectively. Specifically, our survey and experiments show that (a) among NLP researchers, even those who study social bias in AI, do not recognize speciesism or speciesist bias; (b) among NLP data, speciesist bias is inherent in the data annotated in the datasets used to evaluate NLP models; (c) OpenAI GPTs, recent NLP models, exhibit speciesist bias by default. Finally, we discuss how we can reduce speciesism in NLP research.

LLMs' performance on complex tasks is still unsatisfactory. A key issue is that presently LLMs learn in a data-driven schema, while the instructions about these complex tasks are both scarce and hard to collect or construct. On the contrary, a prominent phenomenon is that LLMs can learn rather fast on simpler tasks with adequate prior knowledge captured during pretraining stage. Thus, if the prerequisite and mechanism of such rapid generalization could be elucidated, it could enhance the efficiency and effectiveness of the LLM's ability to learn complex tasks. Thus, in this paper, we employ a gradient-based method, to dissect the process that the SFT process adapts LLMs to downstream tasks via the perspective of attention patterns. We find that: (1) LLMs selectively activate task-specific attention heads during SFT; (2) activation patterns for complex tasks are combinations of basic task patterns; and (3) changes in a few parameters can significantly impact activation patterns after SFT on a small number of samples.Based on these insights, experiments are conducted to actually enhance the efficiency and effectiveness of SFT.