Large language models (LLMs) exhibit in-context learning abilities which enable the same model to perform several tasks without any task-specific training. In contrast, traditional adaptation approaches, such as fine-tuning, modify the underlying models for each specific task. In-context learning, however, consistently underperforms task-specific tuning approaches even when presented with the same examples. While most existing approaches (e.g., prompt engineering) focus on the LLM's learned representations to patch this performance gap, our experiments actually reveal that LLM representations contain sufficient information to make good predictions. As such, we focus on the LLM's reasoning abilities and demonstrate that this performance gap exists due to their inability to perform simple probabilistic reasoning tasks.

As the parameters of Large Language Models (LLMs) have scaled to hundreds of billions, the demand for efficient training methods -- balancing faster computation and reduced memory usage without sacrificing accuracy -- has become more critical than ever. In recent years, various mixed precision strategies, which involve different precision levels for optimization components, have been proposed to increase training speed with minimal accuracy degradation. However, these strategies often require manual adjustments and lack theoretical justification. In this work, we leverage stochastic rounding (SR) to address numerical errors of training with low-precision representation. We provide theoretical analyses of implicit regularization and convergence under the Adam optimizer when SR is utilized. With the insights from these analyses, we extend previous BF16 + SR strategy to be used in distributed settings, enhancing the stability and performance for large scale training. Empirical results from pre-training models with up to 6.7B parameters, for the first time, demonstrate that our BF16 with SR strategy outperforms (BF16, FP32) mixed precision strategies, achieving better validation perplexity, up to $1.54\times$ higher throughput, and $30\%$ less memory usage.

We present a method to compress the final linear layer of language models, reducing memory usage by up to 3.4x without significant performance loss. By grouping tokens based on Byte Pair Encoding (BPE) merges, we prevent materialisation of the memory-intensive logits tensor. Evaluations on the TinyStories dataset show that our method performs on par with GPT-Neo and GPT2 while significantly improving throughput by up to 3x, making it suitable for low-compute environments.

Large language models (LLMs) exhibit in-context learning abilities which enable the same model to perform several tasks without any task-specific training. In contrast, traditional adaptation approaches, such as fine-tuning, modify the underlying models for each specific task. In-context learning, however, consistently underperforms task-specific tuning approaches even when presented with the same examples. While most existing approaches (e.g., prompt engineering) focus on the LLM's learned representations to patch this performance gap, our experiments actually reveal that LLM representations contain sufficient information to make good predictions. As such, we focus on the LLM's reasoning abilities and demonstrate that this performance gap exists due to their inability to perform simple probabilistic reasoning tasks.

While Transformers underpin modern large language models (LMs), there is a growing list of alternative architectures with new capabilities, promises, and tradeoffs. This makes choosing the right LM architecture challenging. Recently-proposed $\textit{hybrid architectures}$ seek a best-of-all-worlds approach that reaps the benefits of all architectures. Hybrid design is difficult for two reasons: it requires manual expert-driven search, and new hybrids must be trained from scratch. We propose $\textbf{Manticore}$, a framework that addresses these challenges. Manticore $\textit{automates the design of hybrid architectures}$ while reusing pretrained models to create $\textit{pretrained}$ hybrids. Our approach augments ideas from differentiable Neural Architecture Search (NAS) by incorporating simple projectors that translate features between pretrained blocks from different architectures. We then fine-tune hybrids that combine pretrained models from different architecture families -- such as the GPT series and Mamba -- end-to-end. With Manticore, we enable LM selection without training multiple models, the construction of pretrained hybrids from existing pretrained models, and the ability to $\textit{program}$ pretrained hybrids to have certain capabilities. Manticore hybrids outperform existing manually-designed hybrids, achieve strong performance on Long Range Arena (LRA) tasks, and can improve on pretrained transformers and state space models.

Large language models (LLMs) can elicit social bias during generations, especially when inference with toxic prompts. Controlling the sensitive attributes in generation encounters challenges in data distribution, generalizability, and efficiency. Specifically, fine-tuning and retrieval demand extensive unbiased corpus, while direct prompting requires meticulously curated instructions for correcting the output in multiple rounds of thoughts but poses challenges on memory and inference latency. In this work, we propose the Expert-Guided Extinction of Toxic Tokens for Debiased Generation (EXPOSED) to eliminate the undesired harmful outputs for LLMs without the aforementioned requirements. EXPOSED constructs a debiasing expert based on the abundant toxic corpus to expose and elicit the potentially dangerous tokens. It then processes the output to the LLMs and constructs a fair distribution by suppressing and attenuating the toxic tokens. EXPOSED is evaluated on fairness benchmarks over three LLM families. Extensive experiments demonstrate that compared with other baselines, the proposed EXPOSED significantly reduces the potential social bias while balancing fairness and generation performance.

Recent Large Language Models (LLMs) have shown the ability to generate content that is difficult or impossible to distinguish from human writing. We investigate the ability of differently-sized LLMs to replicate human writing style in short, creative texts in the domain of Showerthoughts, thoughts that may occur during mundane activities. We compare GPT-2 and GPT-Neo fine-tuned on Reddit data as well as GPT-3.5 invoked in a zero-shot manner, against human-authored texts. We measure human preference on the texts across the specific dimensions that account for the quality of creative, witty texts. Additionally, we compare the ability of humans versus fine-tuned RoBERTa classifiers to detect AI-generated texts. We conclude that human evaluators rate the generated texts slightly worse on average regarding their creative quality, but they are unable to reliably distinguish between human-written and AI-generated texts. We further provide a dataset for creative, witty text generation based on Reddit Showerthoughts posts.

Recent advancements in the realm of deep learning, particularly in the development of large language models (LLMs), have demonstrated AI's ability to tackle complex mathematical problems or solving programming challenges. However, the capability to solve well-defined problems based on extensive training data differs significantly from the nuanced process of making scientific discoveries. Trained on almost all human knowledge available, today's sophisticated LLMs basically learn to predict sequences of tokens. They generate mathematical derivations and write code in a similar way as writing an essay, and do not have the ability to pioneer scientific discoveries in the manner a human scientist would do. In this study we delve into the potential of using deep learning to rediscover a fundamental mathematical concept: integrals. By defining integrals as area under the curve, we illustrate how AI can deduce the integral of a given function, exemplified by inferring $\int_{0}^{x} t^2 dt = \frac{x^3}{3}$ and $\int_{0}^{x} ae^{bt} dt = \frac{a}{b} e^{bx} - \frac{a}{b}$. Our experiments show that deep learning models can approach the task of inferring integrals either through a sequence-to-sequence model, akin to language translation, or by uncovering the rudimentary principles of integration, such as $\int_{0}^{x} t^n dt = \frac{x^{n+1}}{n+1}$.

Recent work found high mutual information between the learned representations of large language models (LLMs) and the geospatial property of its input, hinting an emergent internal model of space. However, whether this internal space model has any causal effects on the LLMs' behaviors was not answered by that work, led to criticism of these findings as mere statistical correlation. Our study focused on uncovering the causality of the spatial representations in LLMs. In particular, we discovered the potential spatial representations in DeBERTa, GPT-Neo using representational similarity analysis and linear and non-linear probing. Our casual intervention experiments showed that the spatial representations influenced the model's performance on next word prediction and a downstream task that relies on geospatial information. Our experiments suggested that the LLMs learn and use an internal model of space in solving geospatial related tasks.

As one of the most exciting features of large language models (LLMs), in-context learning is a mixed blessing. While it allows users to fast-prototype a task solver with only a few training examples, the performance is generally sensitive to various configurations of the prompt such as the choice or order of the training examples. In this paper, we for the first time theoretically and empirically identify that such a paradox is mainly due to the label shift of the in-context model to the data distribution, in which LLMs shift the label marginal $p(y)$ while having a good label conditional $p(x|y)$. With this understanding, we can simply calibrate the in-context predictive distribution by adjusting the label marginal, which is estimated via Monte-Carlo sampling over the in-context model, i.e., generation of LLMs. We call our approach as generative calibration. We conduct exhaustive experiments with 12 text classification tasks and 12 LLMs scaling from 774M to 33B, generally find that the proposed method greatly and consistently outperforms the ICL as well as state-of-the-art calibration methods, by up to 27% absolute in macro-F1. Meanwhile, the proposed method is also stable under different prompt configurations.