Massive activations, which manifest in specific feature dimensions of hidden states, introduce a significant bias in large language models (LLMs), leading to an overemphasis on the corresponding token. In this paper, we identify that massive activations originate not from the hidden state but from the intermediate state of a feed-forward network module in an early layer. Expanding on the previous observation that massive activations occur only in specific feature dimensions, we dive deep into the weights that cause massive activations. Specifically, we define top-$k$ massive weights as the weights that contribute to the dimensions with the top-$k$ magnitudes in the intermediate state. When these massive weights are set to zero, the functionality of LLMs is entirely disrupted. However, when all weights except for massive weights are set to zero, it results in a relatively minor performance drop, even though a much larger number of weights are set to zero. This implies that during the pre-training process, learning is dominantly focused on massive weights. Building on this observation, we propose a simple plug-and-play method called MacDrop (massive weights curriculum dropout), to rely less on massive weights during parameter-efficient fine-tuning. This method applies dropout to the pre-trained massive weights, starting with a high dropout probability and gradually decreasing it as fine-tuning progresses. Through experiments, we demonstrate that MacDrop generally improves performance across zero-shot downstream tasks and generation tasks.

Large language models (LLMs) have shown remarkable advances in supporting long-context comprehension and processing tasks. However, scaling the generation inference of LLMs to such long contexts incurs significant additional computation load, and demands a substantial GPU memory footprint to maintain the key-value (KV) cache of transformer-based LLMs. Existing KV cache compression methods, such as quantization, face memory bottlenecks as context length increases, while static-sized caches, such as eviction, suffer from inefficient policies. These limitations restrict deployment on consumer-grade devices like a single Nvidia 4090 GPU. To overcome this, we propose Locret, a framework for long-context LLM inference that introduces retaining heads to evaluate the causal importance of KV cache units, allowing for more accurate eviction within a fixed cache size. Locret is fine-tuned on top of the frozen backbone LLM using a minimal amount of data from standard long-context SFT datasets. During inference, we evict low-importance cache units along with a chunked prefill pattern, significantly reducing peak GPU memory usage. We conduct an extensive empirical study to evaluate Locret, where the experimental results show that Locret outperforms the recent competitive approaches, including InfLLM, Quantization, SirLLM, and MInference, in terms of memory efficiency and the quality of generated contents -- Locret achieves over a 20x and 8x KV cache compression ratio compared to the full KV cache for Phi-3-mini-128K and Llama-3.1-8B-instruct. Additionally, Locret can be combined with other methods, such as quantization and token merging. To our knowledge, Locret is the first framework capable of deploying Llama-3.1-8B or similar models on a single Nvidia 4090 GPU, enabling 128K long-context inference without compromising generation quality, and requiring little additional system optimizations.

We present a spatially efficient decomposition of matrices and arbitrary-order tensors as linear combinations of tensor products of $\{-1, 1\}$-valued vectors. For any matrix $A \in \mathbb{R}^{m \times n}$, $$A - R_w = S_w C_w T_w^\top = \sum_{j=1}^w c_j \cdot \mathbf{s}_j \mathbf{t}_j^\top$$ is a {\it $w$-width signed cut decomposition of $A$}. Here $C_w = "diag"(\mathbf{c}_w)$ for some $\mathbf{c}_w \in \mathbb{R}^w,$ and $S_w, T_w$, and the vectors $\mathbf{s}_j, \mathbf{t}_j$ are $\{-1, 1\}$-valued. To store $(S_w, T_w, C_w)$, we may pack $w \cdot (m + n)$ bits, and require only $w$ floating point numbers. As a function of $w$, $\|R_w\|_F$ exhibits exponential decay when applied to #f32 matrices with i.i.d. $\mathcal N (0, 1)$ entries. Choosing $w$ so that $(S_w, T_w, C_w)$ has the same memory footprint as a \textit{f16} or \textit{bf16} matrix, the relative error is comparable. Our algorithm yields efficient signed cut decompositions in $20$ lines of pseudocode. It reflects a simple modification from a celebrated 1999 paper [1] of Frieze and Kannan. As a first application, we approximate the weight matrices in the open \textit{Mistral-7B-v0.1} Large Language Model to a $50\%$ spatial compression. Remarkably, all $226$ remainder matrices have a relative error $<6\%$ and the expanded model closely matches \textit{Mistral-7B-v0.1} on the {\it huggingface} leaderboard [2]. Benchmark performance degrades slowly as we reduce the spatial compression from $50\%$ to $25\%$. We optimize our open source \textit{rust} implementation [3] with \textit{simd} instructions on \textit{avx2} and \textit{avx512} architectures. We also extend our algorithm from matrices to tensors of arbitrary order and use it to compress a picture of the first author's cat Angus.

Predicting phenotypes with complex genetic bases based on a small, interpretable set of variant features remains a challenging task. Conventionally, data-driven approaches are utilized for this task, yet the high dimensional nature of genotype data makes the analysis and prediction difficult. Motivated by the extensive knowledge encoded in pre-trained LLMs and their success in processing complex biomedical concepts, we set to examine the ability of LLMs in feature selection and engineering for tabular genotype data, with a novel knowledge-driven framework. We develop FREEFORM, Free-flow Reasoning and Ensembling for Enhanced Feature Output and Robust Modeling, designed with chain-of-thought and ensembling principles, to select and engineer features with the intrinsic knowledge of LLMs. Evaluated on two distinct genotype-phenotype datasets, genetic ancestry and hereditary hearing loss, we find this framework outperforms several data-driven methods, particularly on low-shot regimes. FREEFORM is available as open-source framework at GitHub: https://github.com/PennShenLab/FREEFORM.

We introduce Loki, an open-source tool designed to address the growing problem of misinformation. Loki adopts a human-centered approach, striking a balance between the quality of fact-checking and the cost of human involvement. It decomposes the fact-checking task into a five-step pipeline: breaking down long texts into individual claims, assessing their check-worthiness, generating queries, retrieving evidence, and verifying the claims. Instead of fully automating the claim verification process, Loki provides essential information at each step to assist human judgment, especially for general users such as journalists and content moderators. Moreover, it has been optimized for latency, robustness, and cost efficiency at a commercially usable level. Loki is released under an MIT license and is available on GitHub. We also provide a video presenting the system and its capabilities.

In this article, we investigate the alignment of Large Language Models according to human preferences. We discuss the features of training a Preference Model, which simulates human preferences, and the methods and details we found essential for achieving the best results. We also discuss using Reinforcement Learning to fine-tune Large Language Models and describe the challenges we faced and the ways to overcome them. Additionally, we present our experience with the Direct Preference Optimization method, which enables us to align a Large Language Model with human preferences without creating a separate Preference Model. As our contribution, we introduce the approach for collecting a preference dataset through perplexity filtering, which makes the process of creating such a dataset for a specific Language Model much easier and more cost-effective.

The advancements in artificial intelligence in recent years, such as Large Language Models (LLMs), have fueled expectations for breakthroughs in genomic foundation models (GFMs). The code of nature, hidden in diverse genomes since the very beginning of life's evolution, holds immense potential for impacting humans and ecosystems through genome modeling. Recent breakthroughs in GFMs, such as Evo, have attracted significant investment and attention to genomic modeling, as they address long-standing challenges and transform in-silico genomic studies into automated, reliable, and efficient paradigms. In the context of this flourishing era of consecutive technological revolutions in genomics, GFM studies face two major challenges: the lack of GFM benchmarking tools and the absence of open-source software for diverse genomics. These challenges hinder the rapid evolution of GFMs and their wide application in tasks such as understanding and synthesizing genomes, problems that have persisted for decades. To address these challenges, we introduce GFMBench, a framework dedicated to GFM-oriented benchmarking. GFMBench standardizes benchmark suites and automates benchmarking for a wide range of open-source GFMs. It integrates millions of genomic sequences across hundreds of genomic tasks from four large-scale benchmarks, democratizing GFMs for a wide range of in-silico genomic applications. Additionally, GFMBench is released as open-source software, offering user-friendly interfaces and diverse tutorials, applicable for AutoBench and complex tasks like RNA design and structure prediction. To facilitate further advancements in genome modeling, we have launched a public leaderboard showcasing the benchmark performance derived from AutoBench. GFMBench represents a step toward standardizing GFM benchmarking and democratizing GFM applications.

Retrieval-Augmented Generation (RAG) has been shown to enhance the factual accuracy of Large Language Models (LLMs), but existing methods often suffer from limited reasoning capabilities in effectively using the retrieved evidence, particularly when using open-source LLMs. To mitigate this gap, we introduce a novel framework, Open-RAG, designed to enhance reasoning capabilities in RAG with open-source LLMs. Our framework transforms an arbitrary dense LLM into a parameter-efficient sparse mixture of experts (MoE) model capable of handling complex reasoning tasks, including both single- and multi-hop queries. Open-RAG uniquely trains the model to navigate challenging distractors that appear relevant but are misleading. As a result, Open-RAG leverages latent learning, dynamically selecting relevant experts and integrating external knowledge effectively for more accurate and contextually relevant responses. In addition, we propose a hybrid adaptive retrieval method to determine retrieval necessity and balance the trade-off between performance gain and inference speed. Experimental results show that the Llama2-7B-based Open-RAG outperforms state-of-the-art LLMs and RAG models such as ChatGPT, Self-RAG, and Command R+ in various knowledge-intensive tasks. We open-source our code and models at https://openragmoe.github.io/

LLMs are ideal for decision-making due to their ability to reason over long contexts and identify critical factors. However, challenges arise when processing transcripts of spoken speech describing complex scenarios. These transcripts often contain ungrammatical or incomplete sentences, repetitions, hedging, and vagueness. For example, during a company's earnings call, an executive might project a positive revenue outlook to reassure investors, despite significant uncertainty regarding future earnings. It is crucial for LLMs to incorporate this uncertainty systematically when making decisions. In this paper, we introduce DeFine, a new framework that constructs probabilistic factor profiles from complex scenarios. DeFine then integrates these profiles with analogical reasoning, leveraging insights from similar past experiences to guide LLMs in making critical decisions in novel situations. Our framework separates the tasks of quantifying uncertainty in complex scenarios and incorporating it into LLM decision-making. This approach is particularly useful in fields such as medical consultations, negotiations, and political debates, where making decisions under uncertainty is vital.

We study the depth of grade-school math (GSM) problem-solving capabilities of LLMs. To this end, we evaluate their performance on pairs of existing math word problems together so that the answer to the second problem depends on correctly answering the first problem. Our findings reveal a significant reasoning gap in most LLMs, that is performance difference between solving the compositional pairs and solving each question independently. This gap is more pronounced in smaller, more cost-efficient, and math-specialized models. Moreover, instruction-tuning recipes and code generation have varying effects across LLM sizes, while finetuning on GSM can lead to task overfitting. Our analysis indicates that large reasoning gaps are not because of test-set leakage, but due to distraction from additional context and poor second-hop reasoning. Overall, LLMs exhibit systematic differences in their reasoning abilities, despite what their performance on standard benchmarks indicates.