Understanding AI systems' inner workings is critical for ensuring value alignment and safety. This review explores mechanistic interpretability: reverse-engineering the computational mechanisms and representations learned by neural networks into human-understandable algorithms and concepts to provide a granular, causal understanding. We establish foundational concepts such as features encoding knowledge within neural activations and hypotheses about their representation and computation. We survey methodologies for causally dissecting model behaviors and assess the relevance of mechanistic interpretability to AI safety. We investigate challenges surrounding scalability, automation, and comprehensive interpretation. We advocate for clarifying concepts, setting standards, and scaling techniques to handle complex models and behaviors and expand to domains such as vision and reinforcement learning. Mechanistic interpretability could help prevent catastrophic outcomes as AI systems become more powerful and inscrutable.

Traditionally, success in multilingual machine translation can be attributed to three key factors in training data: large volume, diverse translation directions, and high quality. In the current practice of fine-tuning large language models (LLMs) for translation, we revisit the importance of all these factors. We find that LLMs display strong translation capability after being fine-tuned on as few as 32 training instances, and that fine-tuning on a single translation direction effectively enables LLMs to translate in multiple directions. However, the choice of direction is critical: fine-tuning LLMs with English on the target side can lead to task misinterpretation, which hinders translations into non-English languages. A similar problem arises when noise is introduced into the target side of parallel data, especially when the target language is well-represented in the LLM's pre-training. In contrast, noise in an under-represented language has a less pronounced effect. Our findings suggest that attaining successful alignment hinges on teaching the model to maintain a "superficial" focus, thereby avoiding the learning of erroneous biases beyond translation.

The rapidly developing Large Vision Language Models (LVLMs) have shown notable capabilities on a range of multi-modal tasks, but still face the hallucination phenomena where the generated texts do not align with the given contexts, significantly restricting the usages of LVLMs. Most previous work detects and mitigates hallucination at the coarse-grained level or requires expensive annotation (e.g., labeling by proprietary models or human experts). To address these issues, we propose detecting and mitigating hallucinations in LVLMs via fine-grained AI feedback. The basic idea is that we generate a small-size sentence-level hallucination annotation dataset by proprietary models, whereby we train a hallucination detection model which can perform sentence-level hallucination detection, covering primary hallucination types (i.e., object, attribute, and relationship). Then, we propose a detect-then-rewrite pipeline to automatically construct preference dataset for training hallucination mitigating model. Furthermore, we propose differentiating the severity of hallucinations, and introducing a Hallucination Severity-Aware Direct Preference Optimization (HSA-DPO) for mitigating hallucination in LVLMs by incorporating the severity of hallucinations into preference learning. Extensive experiments demonstrate the effectiveness of our method.

A fundamental skill among human developers is the ability to understand and reason about program execution. As an example, a programmer can mentally simulate code execution in natural language to debug and repair code (aka. rubber duck debugging). However, large language models (LLMs) of code are typically trained on the surface textual form of programs, thus may lack a semantic understanding of how programs execute at run-time. To address this issue, we propose NExT, a method to teach LLMs to inspect the execution traces of programs (variable states of executed lines) and reason about their run-time behavior through chain-of-thought (CoT) rationales. Specifically, NExT uses self-training to bootstrap a synthetic training set of execution-aware rationales that lead to correct task solutions (e.g., fixed programs) without laborious manual annotation. Experiments on program repair tasks based on MBPP and HumanEval demonstrate that NExT improves the fix rate of a PaLM 2 model, by 26.1% and 14.3% absolute, respectively, with significantly improved rationale quality as verified by automated metrics and human raters. Our model can also generalize to scenarios where program traces are absent at test-time.

We describe an approach for aligning an LLM-based dialogue agent based on global (i.e., dialogue-level) rewards, while also taking into account naturally-occurring multimodal signals. At a high level, our approach (dubbed GELI) learns a local, turn-level reward model by decomposing the human-provided Global Explicit (GE) session-level reward, using Local Implicit (LI) multimodal reward signals to crossmodally shape the reward decomposition step. This decomposed reward model is then used as part of the standard RHLF pipeline improve an LLM-based dialog agent. We run quantitative and qualitative human studies to evaluate the performance of our GELI approach, and find that it shows consistent improvements across various conversational metrics compared to baseline methods.

We present the results and the main findings of SemEval-2024 Task 8: Multigenerator, Multidomain, and Multilingual Machine-Generated Text Detection. The task featured three subtasks. Subtask A is a binary classification task determining whether a text is written by a human or generated by a machine. This subtask has two tracks: a monolingual track focused solely on English texts and a multilingual track. Subtask B is to detect the exact source of a text, discerning whether it is written by a human or generated by a specific LLM. Subtask C aims to identify the changing point within a text, at which the authorship transitions from human to machine. The task attracted a large number of participants: subtask A monolingual (126), subtask A multilingual (59), subtask B (70), and subtask C (30). In this paper, we present the task, analyze the results, and discuss the system submissions and the methods they used. For all subtasks, the best systems used LLMs.

Medical errors in clinical text pose significant risks to patient safety. The MEDIQA-CORR 2024 shared task focuses on detecting and correcting these errors across three subtasks: identifying the presence of an error, extracting the erroneous sentence, and generating a corrected sentence. In this paper, we present our approach that achieved top performance in all three subtasks. For the MS dataset, which contains subtle errors, we developed a retrieval-based system leveraging external medical question-answering datasets. For the UW dataset, reflecting more realistic clinical notes, we created a pipeline of modules to detect, localize, and correct errors. Both approaches utilized the DSPy framework for optimizing prompts and few-shot examples in large language model (LLM) based programs. Our results demonstrate the effectiveness of LLM based programs for medical error correction. However, our approach has limitations in addressing the full diversity of potential errors in medical documentation. We discuss the implications of our work and highlight future research directions to advance the robustness and applicability of medical error detection and correction systems.

This paper presents a framework that can interpret humans' navigation commands containing temporal elements and directly translate their natural language instructions into robot motion planning. Central to our framework is utilizing Large Language Models (LLMs). To enhance the reliability of LLMs in the framework and improve user experience, we propose methods to resolve the ambiguity in natural language instructions and capture user preferences. The process begins with an ambiguity classifier, identifying potential uncertainties in the instructions. Ambiguous statements trigger a GPT-4-based mechanism that generates clarifying questions, incorporating user responses for disambiguation. Also, the framework assesses and records user preferences for non-ambiguous instructions, enhancing future interactions. The last part of this process is the translation of disambiguated instructions into a robot motion plan using Linear Temporal Logic. This paper details the development of this framework and the evaluation of its performance in various test scenarios.

The task of accurate and efficient language translation is an extremely important information processing task. Machine learning enabled and automated translation that is accurate and fast is often a large topic of interest in the machine learning and data science communities. In this study, we examine using local Generative Pretrained Transformer (GPT) models to perform automated zero shot black-box, sentence wise, multi-natural-language translation into English text. We benchmark 16 different open-source GPT models, with no custom fine-tuning, from the Huggingface LLM repository for translating 50 different non-English languages into English using translated TED Talk transcripts as the reference dataset. These GPT model inference calls are performed strictly locally, on single A100 Nvidia GPUs. Benchmark metrics that are reported are language translation accuracy, using BLEU, GLEU, METEOR, and chrF text overlap measures, and wall-clock time for each sentence translation. The best overall performing GPT model for translating into English text for the BLEU metric is ReMM-v2-L2-13B with a mean score across all tested languages of $0.152$, for the GLEU metric is ReMM-v2-L2-13B with a mean score across all tested languages of $0.256$, for the chrF metric is Llama2-chat-AYT-13B with a mean score across all tested languages of $0.448$, and for the METEOR metric is ReMM-v2-L2-13B with a mean score across all tested languages of $0.438$.

Citation text plays a pivotal role in elucidating the connection between scientific documents, demanding an in-depth comprehension of the cited paper. Constructing citations is often time-consuming, requiring researchers to delve into extensive literature and grapple with articulating relevant content. To address this challenge, the field of citation text generation (CTG) has emerged. However, while earlier methods have primarily centered on creating single-sentence citations, practical scenarios frequently necessitate citing multiple papers within a single paragraph. To bridge this gap, we propose a method that leverages Large Language Models (LLMs) to generate multi-citation sentences. Our approach involves a single source paper and a collection of target papers, culminating in a coherent paragraph containing multi-sentence citation text. Furthermore, we introduce a curated dataset named MCG-S2ORC, composed of English-language academic research papers in Computer Science, showcasing multiple citation instances. In our experiments, we evaluate three LLMs LLaMA, Alpaca, and Vicuna to ascertain the most effective model for this endeavor. Additionally, we exhibit enhanced performance by integrating knowledge graphs from target papers into the prompts for generating citation text. This research underscores the potential of harnessing LLMs for citation generation, opening a compelling avenue for exploring the intricate connections between scientific documents.