Large language models (LLMs) memorize text from several sources of documents. In pretraining, LLM trains to maximize the likelihood of text but neither receives the source of the text nor memorizes the source. Accordingly, LLM can not provide document information on the generated content, and users do not obtain any hint of reliability, which is crucial for factuality or privacy infringement. This work introduces token-level source identification in the decoding step, which maps the token representation to the reference document. We propose a bi-gram source identifier, a multi-layer perceptron with two successive token representations as input for better generalization. We conduct extensive experiments on Wikipedia and PG19 datasets with several LLMs, layer locations, and identifier sizes. The overall results show a possibility of token-level source identifiers for tracing the document, a crucial problem for the safe use of LLMs.

For both human speakers and which we test in a zero-shot manner. We machine learning models, the task requires connecting show that LLaVA (Liu et al., 2024), a powerful the visual data causally, to generate a narrative foundation model, performs best on the task, but consistent with the contents of the images. As for only slightly so than TAPM (Yu et al., 2021), a model-generated stories, evaluation is one of the model designed for visual storytelling which is 50 key challenges due to the inherently creative nature times smaller than LLaVA. Second, given insights of the task. Since human-written stories are derived from our proposed distance-based evaluation typically used to train visual storytelling models-- method, we upgrade the visual and language under the assumption that these stories provide a components on TAPM, resulting in a model that good learning signal--most previous work evaluated achieves comparable performance to LLaVA with model-generated stories by directly comparing a significantly lower number of parameters.

Sparsely-activated Mixture-of-Experts (MoE) architecture has increasingly been adopted to further scale large language models (LLMs) due to its sub-linear scaling for computation costs. However, frequent failures still pose significant challenges as training scales. The cost of even a single failure is significant, as all GPUs need to wait idle until the failure is resolved, potentially losing considerable training progress as training has to restart from checkpoints. Existing solutions for efficient fault-tolerant training either lack elasticity or rely on building resiliency into pipeline parallelism, which cannot be applied to MoE models due to the expert parallelism strategy adopted by the MoE architecture. We present Lazarus, a system for resilient and elastic training of MoE models. Lazarus adaptively allocates expert replicas to address the inherent imbalance in expert workload and speeds-up training, while a provably optimal expert placement algorithm is developed to maximize the probability of recovery upon failures. Through adaptive expert placement and a flexible token dispatcher, Lazarus can also fully utilize all available nodes after failures, leaving no GPU idle. Our evaluation shows that Lazarus outperforms existing MoE training systems by up to 5.7x under frequent node failures and 3.4x on a real spot instance trace.

State-of-the-art language models are autoregressive and operate on subword units known as tokens. Specifically, one must encode the conditioning string into a list of tokens before passing to the language models for next-token prediction. We show that popular encoding schemes, such as maximum prefix encoding (MPE) and byte-pair-encoding (BPE), induce a sampling bias that cannot be mitigated with more training or data. To counter this universal problem, for each encoding scheme above, we propose a novel algorithm to obtain unbiased estimates from any language model trained on tokenized data. Our methods do not require finetuning the model, and the complexity, defined as the number of model runs, scales linearly with the sequence length in the case of MPE. As a result, we show that one can simulate token-free behavior from a tokenized language model. We empirically verify the correctness of our method through a Markov-chain setup, where it accurately recovers the transition probabilities, as opposed to the conventional method of directly prompting tokens into the language model.

In the Emotion Recognition in Conversation task, recent investigations have utilized attention mechanisms exploring relationships among utterances from intra- and inter-speakers for modeling emotional interaction between them. However, attributes such as speaker personality traits remain unexplored and present challenges in terms of their applicability to other tasks or compatibility with diverse model architectures. Therefore, this work introduces a novel framework named BiosERC, which investigates speaker characteristics in a conversation. By employing Large Language Models (LLMs), we extract the "biographical information" of the speaker within a conversation as supplementary knowledge injected into the model to classify emotional labels for each utterance. Our proposed method achieved state-of-the-art (SOTA) results on three famous benchmark datasets: IEMOCAP, MELD, and EmoryNLP, demonstrating the effectiveness and generalization of our model and showcasing its potential for adaptation to various conversation analysis tasks. Our source code is available at https://github.com/yingjie7/BiosERC.

Can a mere next-token predictor faithfully model human intelligence? We crystallize this emerging concern and correct popular misconceptions surrounding it, and advocate a simple multi-token objective. As a starting point, we argue that the two often-conflated phases of next-token prediction -- autoregressive inference and teacher-forced training -- must be treated distinctly. The popular criticism that errors can compound during autoregressive inference, crucially assumes that teacher-forcing has learned an accurate next-token predictor. This assumption sidesteps a more deep-rooted problem we expose: in certain classes of tasks, teacher-forcing can simply fail to learn an accurate next-token predictor in the first place. We describe a general mechanism of how teacher-forcing can fail, and design a minimal planning task where both the Transformer and the Mamba architecture empirically fail in that manner -- remarkably, despite the task being straightforward to learn. Finally, we provide preliminary evidence that this failure can be resolved using a simple modification that predicts multiple tokens in advance. We hope this finding can ground future debates and inspire explorations beyond the next-token prediction paradigm. We make our code available under https://github.com/gregorbachmann/Next-Token-Failures

AI assistants such as ChatGPT are trained to respond to users by saying, "I am a large language model". This raises questions. Do such models know that they are LLMs and reliably act on this knowledge? Are they aware of their current circumstances, such as being deployed to the public? We refer to a model's knowledge of itself and its circumstances as situational awareness. To quantify situational awareness in LLMs, we introduce a range of behavioral tests, based on question answering and instruction following. These tests form the $\textbf{Situational Awareness Dataset (SAD)}$, a benchmark comprising 7 task categories and over 13,000 questions. The benchmark tests numerous abilities, including the capacity of LLMs to (i) recognize their own generated text, (ii) predict their own behavior, (iii) determine whether a prompt is from internal evaluation or real-world deployment, and (iv) follow instructions that depend on self-knowledge. We evaluate 16 LLMs on SAD, including both base (pretrained) and chat models. While all models perform better than chance, even the highest-scoring model (Claude 3 Opus) is far from a human baseline on certain tasks. We also observe that performance on SAD is only partially predicted by metrics of general knowledge (e.g. MMLU). Chat models, which are finetuned to serve as AI assistants, outperform their corresponding base models on SAD but not on general knowledge tasks. The purpose of SAD is to facilitate scientific understanding of situational awareness in LLMs by breaking it down into quantitative abilities. Situational awareness is important because it enhances a model's capacity for autonomous planning and action. While this has potential benefits for automation, it also introduces novel risks related to AI safety and control. Code and latest results available at https://situational-awareness-dataset.org .

This study evaluates the capabilities of ChatGPT versions 3.5 and 4 in generating code across a diverse range of programming languages. Our objective is to assess the effectiveness of these AI models for generating scientific programs. To this end, we asked ChatGPT to generate three distinct codes: a simple numerical integration, a conjugate gradient solver, and a parallel 1D stencil-based heat equation solver. The focus of our analysis was on the compilation, runtime performance, and accuracy of the codes. While both versions of ChatGPT successfully created codes that compiled and ran (with some help), some languages were easier for the AI to use than others (possibly because of the size of the training sets used). Parallel codes -- even the simple example we chose to study here -- also difficult for the AI to generate correctly.

AI-supported programming has taken giant leaps with tools such as Meta's Llama and openAI's chatGPT. These are examples of stochastic sources of programs and have already greatly influenced how we produce code and teach programming. If we consider input to such models as a stochastic source, a natural question is, what is the relation between the input and the output distributions, between the chatGPT prompt and the resulting program? In this paper, we will show how the relation between random Python programs generated from chatGPT can be described geometrically and topologically using Tree-edit distances between the program's syntax trees and without explicit modeling of the underlying space. A popular approach to studying high-dimensional samples in a metric space is to use low-dimensional embedding using, e.g., multidimensional scaling. Such methods imply errors depending on the data and dimension of the embedding space. In this article, we propose to restrict such projection methods to purely visualization purposes and instead use geometric summary statistics, methods from spatial point statistics, and topological data analysis to characterize the configurations of random programs that do not rely on embedding approximations. To demonstrate their usefulness, we compare two publicly available models: ChatGPT-4 and TinyLlama, on a simple problem related to image processing. Application areas include understanding how questions should be asked to obtain useful programs; measuring how consistently a given large language model answers; and comparing the different large language models as a programming assistant. Finally, we speculate that our approach may in the future give new insights into the structure of programming languages.

Recent progress in large language models (LLMs) has led to their widespread adoption in various domains. However, these advancements have also introduced additional safety risks and raised concerns regarding their detrimental impact on already marginalized populations. Despite growing mitigation efforts to develop safety safeguards, such as supervised safety-oriented fine-tuning and leveraging safe reinforcement learning from human feedback, multiple concerns regarding the safety and ingrained biases in these models remain. Furthermore, previous work has demonstrated that models optimized for safety often display exaggerated safety behaviors, such as a tendency to refrain from responding to certain requests as a precautionary measure. As such, a clear trade-off between the helpfulness and safety of these models has been documented in the literature. In this paper, we further investigate the effectiveness of safety measures by evaluating models on already mitigated biases. Using the case of Llama 2 as an example, we illustrate how LLMs' safety responses can still encode harmful assumptions. To do so, we create a set of non-toxic prompts, which we then use to evaluate Llama models. Through our new taxonomy of LLMs responses to users, we observe that the safety/helpfulness trade-offs are more pronounced for certain demographic groups which can lead to quality-of-service harms for marginalized populations.