Large language models (LLMs) have marked a significant milestone in the advancement of natural language processing (Radford et al., 2019; Brown et al., 2020; Ouyang et al., 2022; Bubeck et al., 2023), showcasing remarkable capabilities in understanding and generating human-like text. However, one pressing issue for the LLMs is their propensity to hallucinate (Rawte et al., 2023) and generate misleading or entirely fabricated information that can significantly undermine their trustworthiness and reliability. The task of uncertainty estimation has then emerged to be an important problem that aims to determine the confidence levels of LLMs' outputs. While the problem of uncertainty estimation and calibration has seen considerable development within the general machine learning and deep learning domains (Abdar et al., 2021; Gawlikowski et al., 2023), we see less development in the domain of LLMs. One of the major challenges is the difference in the format of the output: while machine learning and deep learning typically involve fixed-dimensional outputs, natural language generation (NLG) tasks central to LLM applications require handling variable outputs that carry semantic meanings. Existing uncertainty estimation approaches for LLMs usually involve designing uncertainty metrics for their outputs. For black-box LLMs, these metrics are computed by examining aspects like the generated outputs' consistency, similarity, entropy, and other relevant characteristics (Lin et al., 2023; Manakul et al., 2023; Kuhn et al., 2023; Hou et al., 2023; Farquhar et al., 2024). Given the complexity of LLMs' underlying architectures, semantic information may be diluted when processing through self-attention mechanisms and during token encoding/decoding. To address this issue, a growing stream of literature argues that hidden layers' activation values within the LLMs offer insights into the LLMs' knowledge and confidence (Slobodkin et al., 2023; Ahdritz et al., 2024; Duan et al., 2024).

Modern high-stakes systems, such as healthcare or robotics, often generate vast streaming event sequences. Our goal is to design an efficient, plug-and-play tool to elicit logic tree-based explanations from Large Language Models (LLMs) to provide customized insights into each observed event sequence. Built on the temporal point process model for events, our method employs the likelihood function as a score to evaluate generated logic trees. We propose an amortized Expectation-Maximization (EM) learning framework and treat the logic tree as latent variables. In the E-step, we evaluate the posterior distribution over the latent logic trees using an LLM prior and the likelihood of the observed event sequences. LLM provides a high-quality prior for the latent logic trees, however, since the posterior is built over a discrete combinatorial space, we cannot get the closed-form solution. We propose to generate logic tree samples from the posterior using a learnable GFlowNet, which is a diversity-seeking generator for structured discrete variables. The M-step employs the generated logic rules to approximate marginalization over the posterior, facilitating the learning of model parameters and refining the tunable LLM prior parameters. In the online setting, our locally built, lightweight model will iteratively extract the most relevant rules from LLMs for each sequence using only a few iterations. Empirical demonstrations showcase the promising performance and adaptability of our framework.

We present the first study of the robustness of existing watermarking techniques on Python code generated by large language models. Although existing works showed that watermarking can be robust for natural language, we show that it is easy to remove these watermarks on code by semantic-preserving transformations.

Large language models (LLMs) now support extremely long context windows, but the quadratic complexity of vanilla attention results in significantly long Time-to-First-Token (TTFT) latency. Existing approaches to address this complexity require additional pretraining or finetuning, and often sacrifice model accuracy. In this paper, we first provide both theoretical and empirical foundations for near-lossless sparse attention. We find dynamically capturing head-specific sparse patterns at runtime with low overhead is crucial. To address this, we propose SampleAttention, an adaptive structured and near-lossless sparse attention. Leveraging observed significant sparse patterns, SampleAttention attends to a fixed percentage of adjacent tokens to capture local window patterns, and employs a two-stage query-guided key-value filtering approach, which adaptively select a minimum set of key-values with low overhead, to capture column stripe patterns. Comprehensive evaluations show that SampleAttention can seamlessly replace vanilla attention in off-the-shelf LLMs with nearly no accuracy loss, and reduces TTFT by up to $2.42\times$ compared with FlashAttention.

During task-oriented dialogues (TODs), human users naturally introduce chitchat that is beyond the immediate scope of the task, interfering with the flow of the conversation. To address this issue without the need for expensive manual data creation, we use few-shot prompting with Llama-2-70B to enhance the MultiWOZ dataset with user backstories, a typical example of chitchat interference in TODs. We assess the impact of this addition by testing two models: one trained solely on TODs and another trained on TODs with a preliminary chitchat interaction. Our analysis demonstrates that our enhanced dataset poses a challenge for these systems. Moreover, we demonstrate that our dataset can be effectively used for training purposes, enabling a system to consistently acknowledge the user's backstory while also successfully moving the task forward in the same turn, as confirmed by human evaluation. These findings highlight the benefits of generating novel chitchat-TOD scenarios to test TOD systems more thoroughly and improve their resilience to natural user interferences

Recent studies indicated GPT-4 outperforms online crowd workers in data labeling accuracy, notably workers from Amazon Mechanical Turk (MTurk). However, these studies were criticized for deviating from standard crowdsourcing practices and emphasizing individual workers' performances over the whole data-annotation process. This paper compared GPT-4 and an ethical and well-executed MTurk pipeline, with 415 workers labeling 3,177 sentence segments from 200 scholarly articles using the CODA-19 scheme. Two worker interfaces yielded 127,080 labels, which were then used to infer the final labels through eight label-aggregation algorithms. Our evaluation showed that despite best practices, MTurk pipeline's highest accuracy was 81.5%, whereas GPT-4 achieved 83.6%. Interestingly, when combining GPT-4's labels with crowd labels collected via an advanced worker interface for aggregation, 2 out of the 8 algorithms achieved an even higher accuracy (87.5%, 87.0%). Further analysis suggested that, when the crowd's and GPT-4's labeling strengths are complementary, aggregating them could increase labeling accuracy.

Massive Over-activation Yielded Uplifts(MOYU) is an inherent property of large language models, and dynamic activation(DA) based on the MOYU property is a clever yet under-explored strategy designed to accelerate inference in these models. Existing methods that utilize MOYU often face a significant 'Impossible Trinity': struggling to simultaneously maintain model performance, enhance inference speed, and extend applicability across various architectures. Due to the theoretical ambiguities surrounding MOYU, this paper elucidates the root cause of the MOYU property and outlines the mechanisms behind two primary limitations encountered by current DA methods: 1) history-related activation uncertainty, and 2) semantic-irrelevant activation inertia. Our analysis not only underscores the limitations of current dynamic activation strategies within large-scale LLaMA models but also proposes opportunities for refining the design of future sparsity schemes.

Mixture-of-Experts (MoE) enjoys performance gain by increasing model capacity while keeping computation cost constant. When comparing MoE to dense models, prior work typically adopt the following setting: 1) use FLOPs or activated parameters as a measure of model complexity; 2) train all models to the same number of tokens. We argue that this setting favors MoE as FLOPs and activated parameters do not accurately measure the communication overhead in sparse layers, leading to a larger actual training budget for MoE. In this work, we revisit the settings by adopting step time as a more accurate measure of model complexity, and by determining the total compute budget under the Chinchilla compute-optimal settings. To efficiently run MoE on modern accelerators, we adopt a 3D sharding method that keeps the dense-to-MoE step time increase within a healthy range. We evaluate MoE and dense LLMs on a set of nine 0-shot and two 1-shot English tasks, as well as MMLU 5-shot and GSM8K 8-shot across three model scales at 6.4B, 12.6B, and 29.6B. Experimental results show that even under these settings, MoE consistently outperform dense LLMs on the speed-accuracy trade-off curve with meaningful gaps. Our full model implementation and sharding strategy has been released at https://github.com/apple/axlearn.

Some prior work has shown that LLMs perform well in NLG evaluation for different tasks. However, we discover that LLMs seem to confuse different evaluation criteria, which reduces their reliability. For further verification, we first consider avoiding issues of inconsistent conceptualization and vague expression in existing NLG quality criteria themselves. So we summarize a clear hierarchical classification system for 11 common aspects with corresponding different criteria from previous studies involved. Inspired by behavioral testing, we elaborately design 18 types of aspect-targeted perturbation attacks for fine-grained analysis of the evaluation behaviors of different LLMs. We also conduct human annotations beyond the guidance of the classification system to validate the impact of the perturbations. Our experimental results reveal confusion issues inherent in LLMs, as well as other noteworthy phenomena, and necessitate further research and improvements for LLM-based evaluation.

Accelerating the inference of large language models (LLMs) is an important challenge in artificial intelligence. This paper introduces distributed speculative inference (DSI), a novel distributed inference algorithm that is provably faster than speculative inference (SI) [Leviathan et al., 2023, Chen et al., 2023, Miao et al., 2023] and traditional autoregressive inference (non-SI). Like other SI algorithms, DSI works on frozen LLMs, requiring no training or architectural modifications, and it preserves the target distribution. Prior studies on SI have demonstrated empirical speedups (compared to non-SI) but require a fast and accurate drafter LLM. In practice, off-the-shelf LLMs often do not have matching drafters that are sufficiently fast and accurate. We show a gap: SI gets slower than non-SI when using slower or less accurate drafters. We close this gap by proving that DSI is faster than both SI and non-SI--given any drafters. By orchestrating multiple instances of the target and drafters, DSI is not only faster than SI but also supports LLMs that cannot be accelerated with SI. Our simulations show speedups of off-the-shelf LLMs in realistic settings: DSI is 1.29-1.92x