Finite state machines (FSMs) are widely used to manage robot behavior logic, particularly in real-world applications that require a high degree of reliability and structure. However, traditional manual FSM design and modification processes can be time-consuming and error-prone. We propose that large language models (LLMs) can assist developers in editing FSM code for real-world robotic use cases. LLMs, with their ability to use context and process natural language, offer a solution for FSM modification with high correctness, allowing developers to update complex control logic through natural language instructions. Our approach leverages few-shot prompting and language-guided code generation to reduce the amount of time it takes to edit an FSM. To validate this approach, we evaluate it on a real-world robotics dataset, demonstrating its effectiveness in practical scenarios.

Recent studies have discovered that LLMs have serious privacy leakage concerns, where an LLM may be fooled into outputting private information under carefully crafted adversarial prompts. These risks include leaking system prompts, personally identifiable information, training data, and model parameters. Most existing red-teaming approaches for privacy leakage rely on humans to craft the adversarial prompts. A few automated methods are proposed for system prompt extraction, but they cannot be applied to more severe risks (e.g., training data extraction) and have limited effectiveness even for system prompt extraction. In this paper, we propose PrivAgent, a novel black-box red-teaming framework for LLM privacy leakage. We formulate different risks as a search problem with a unified attack goal. Our framework trains an open-source LLM through reinforcement learning as the attack agent to generate adversarial prompts for different target models under different risks. We propose a novel reward function to provide effective and fine-grained rewards for the attack agent. Finally, we introduce customizations to better fit our general framework to system prompt extraction and training data extraction. Through extensive evaluations, we first show that PrivAgent outperforms existing automated methods in system prompt leakage against six popular LLMs. Notably, our approach achieves a 100% success rate in extracting system prompts from real-world applications in OpenAI's GPT Store. We also show PrivAgent's effectiveness in extracting training data from an open-source LLM with a success rate of 5.9%. We further demonstrate PrivAgent's effectiveness in evading the existing guardrail defense and its helpfulness in enabling better safety alignment. Finally, we validate our customized designs through a detailed ablation study. We release our code here https://github.com/rucnyz/RedAgent.

Large multimodal models (LMMs) have shown remarkable performance in the visual commonsense reasoning (VCR) task, which aims to answer a multiple-choice question based on visual commonsense within an image. However, the ability of LMMs to correct potential visual commonsense errors in the distractor upon their occurrence is yet under-explored. Drawing inspiration from how a human teacher crafts challenging distractors to test students' comprehension of the concepts or skills and assists them in identifying and correcting errors toward the answer, we are the pioneering research for LMMs to simulate this error correction process. To this end, we employ GPT-4 as a ``teacher'' to collect the explainable feedback dataset VCR-DF for error correction, which serves as a benchmark to evaluate the ability of LMMs to identify misconceptions and clarify reasons behind the error in VCR distractors toward final answers. In addition, we propose an LMM-based Pedagogical Expert Instructed Feedback Generation (PEIFG) model to incorporate the learnable expert prompts and multimodal instruction as guidance for feedback generation. Experimental results show that our PEIFG significantly outperforms existing LMMs. We believe that our benchmark provides a new direction for evaluating the capabilities of LMMs.

Equivalence checking of two SQL queries is an intractable problem encountered in diverse contexts ranging from grading student submissions in a DBMS course to debugging query rewriting rules in an optimizer, and many more. While a lot of progress has been made in recent years in developing practical solutions for this problem, the existing methods can handle only a small subset of SQL, even for bounded equivalence checking. They cannot support sophisticated SQL expressions one encounters in practice. At the same time, large language models (LLMs) -- such as GPT-4 -- have emerged as power generators of SQL from natural language specifications. This paper explores whether LLMs can also demonstrate the ability to reason with SQL queries and help advance SQL equivalence checking. Towards this, we conducted a detailed evaluation of several LLMs over collections with SQL pairs of varying levels of complexity. We explored the efficacy of different prompting techniques, the utility of synthetic examples & explanations, as well as logical plans generated by query parsers. Our main finding is that with well-designed prompting using an unoptimized SQL Logical Plan, LLMs can perform equivalence checking beyond the capabilities of current techniques, achieving nearly 100% accuracy for equivalent pairs and up to 70% for non-equivalent pairs of SQL queries. While LLMs lack the ability to generate formal proofs, their synthetic examples and human-readable explanations offer valuable insights to students (& instructors) in a classroom setting and to database administrators (DBAs) managing large database installations. Additionally, we also show that with careful fine-tuning, we can close the performance gap between smaller (and efficient) models and larger models such as GPT, thus paving the way for potential LLM-integration in standalone data processing systems.

Realizing scaling laws in embodied AI has become a focus. However, previous work has been scattered across diverse simulation platforms, with assets and models lacking unified interfaces, which has led to inefficiencies in research. To address this, we introduce InfiniteWorld, a unified and scalable simulator for general vision-language robot interaction built on Nvidia Isaac Sim. InfiniteWorld encompasses a comprehensive set of physics asset construction methods and generalized free robot interaction benchmarks. Specifically, we first built a unified and scalable simulation framework for embodied learning that integrates a series of improvements in generation-driven 3D asset construction, Real2Sim, automated annotation framework, and unified 3D asset processing. This framework provides a unified and scalable platform for robot interaction and learning. In addition, to simulate realistic robot interaction, we build four new general benchmarks, including scene graph collaborative exploration and open-world social mobile manipulation. The former is often overlooked as an important task for robots to explore the environment and build scene knowledge, while the latter simulates robot interaction tasks with different levels of knowledge agents based on the former. They can more comprehensively evaluate the embodied agent's capabilities in environmental understanding, task planning and execution, and intelligent interaction. We hope that this work can provide the community with a systematic asset interface, alleviate the dilemma of the lack of high-quality assets, and provide a more comprehensive evaluation of robot interactions.

This paper presents a case study on the development of Auto-AdvER, a specialised named entity recognition schema and dataset for text in the car advertisement genre. Developed with industry needs in mind, Auto-AdvER is designed to enhance text mining analytics in this domain and contributes a linguistically unique NER dataset. We present a schema consisting of three labels: "Condition", "Historic" and "Sales Options". We outline the guiding principles for annotation, describe the methodology for schema development, and show the results of an annotation study demonstrating inter-annotator agreement of 92% F1-Score. Furthermore, we compare the performance by using encoder-only models: BERT, DeBERTaV3 and decoder-only open and closed source Large Language Models (LLMs): Llama, Qwen, GPT-4 and Gemini. Our results show that the class of LLMs outperforms the smaller encoder-only models. However, the LLMs are costly and far from perfect for this task. We present this work as a stepping stone toward more fine-grained analysis and discuss Auto-AdvER's potential impact on advertisement analytics and customer insights, including applications such as the analysis of market dynamics and data-driven predictive maintenance. Our schema, as well as our associated findings, are suitable for both private and public entities considering named entity recognition in the automotive domain, or other specialist domains.

Large Language Models (LLMs) are distinguished by their massive parameter counts, which typically result in significant redundancy. This work introduces MaskLLM, a learnable pruning method that establishes Semi-structured (or ``N:M'') Sparsity in LLMs, aimed at reducing computational overhead during inference. Instead of developing a new importance criterion, MaskLLM explicitly models N:M patterns as a learnable distribution through Gumbel Softmax sampling. This approach facilitates end-to-end training on large-scale datasets and offers two notable advantages: 1) High-quality Masks - our method effectively scales to large datasets and learns accurate masks; 2) Transferability - the probabilistic modeling of mask distribution enables the transfer learning of sparsity across domains or tasks. We assessed MaskLLM using 2:4 sparsity on various LLMs, including LLaMA-2, Nemotron-4, and GPT-3, with sizes ranging from 843M to 15B parameters, and our empirical results show substantial improvements over state-of-the-art methods. For instance, leading approaches achieve a perplexity (PPL) of 10 or greater on Wikitext compared to the dense model's 5.12 PPL, but MaskLLM achieves a significantly lower 6.72 PPL solely by learning the masks with frozen weights. Furthermore, MaskLLM's learnable nature allows customized masks for lossless application of 2:4 sparsity to downstream tasks or domains. Code is available at https://github.com/NVlabs/MaskLLM.

Large Language Models (LLMs) exhibit impressive performance across various tasks, but deploying them for inference poses challenges. Their high resource demands often necessitate complex, costly multi-GPU pipelines, or the use of smaller, less capable models. While quantization offers a promising solution utilizing lower precision for model storage, existing methods frequently experience significant performance drops at lower precision levels. Additionally, they typically provide only a limited set of solutions at specific bit levels, many of which are extensively manually tuned. To address these challenges, we propose a new method called SKIM: Scaled K-means clustering wIth Mixed precision. Our approach introduces two novel techniques: 1. A greedy algorithm to solve approximately optimal bit allocation across weight channels, and 2. A trainable scaling vector for non-differentiable K-means clustering. These techniques substantially improve performance and can be adapted to any given bit. Notably, in terms of model perplexity, our method narrows the gap between 3-bit quantized LLaMA models and their full precision counterparts by 16.3% on average.

Originally formalized with symbolic representations, syntactic trees may also be effectively represented in the activations of large language models (LLMs). Indeed, a 'Structural Probe' can find a subspace of neural activations, where syntactically related words are relatively close to one-another. However, this syntactic code remains incomplete: the distance between the Structural Probe word embeddings can represent the existence but not the type and direction of syntactic relations. Here, we hypothesize that syntactic relations are, in fact, coded by the relative direction between nearby embeddings. To test this hypothesis, we introduce a 'Polar Probe' trained to read syntactic relations from both the distance and the direction between word embeddings. Our approach reveals three main findings. First, our Polar Probe successfully recovers the type and direction of syntactic relations, and substantially outperforms the Structural Probe by nearly two folds. Second, we confirm that this polar coordinate system exists in a low-dimensional subspace of the intermediate layers of many LLMs and becomes increasingly precise in the latest frontier models. Third, we demonstrate with a new benchmark that similar syntactic relations are coded similarly across the nested levels of syntactic trees. Overall, this work shows that LLMs spontaneously learn a geometry of neural activations that explicitly represents the main symbolic structures of linguistic theory.

Molecule optimization is a critical task in drug discovery to optimize desired properties of a given molecule through chemical modification. Despite Large Language Models (LLMs) holding the potential to efficiently simulate this task by using natural language to direct the optimization, straightforwardly utilizing them shows limited performance. In this work, we facilitate utilizing LLMs in an iterative paradigm by proposing a simple yet highly effective domain feedback provider, namely $\text{Re}^3$DF. In detail, $\text{Re}^3$DF harnesses an external toolkit, RDKit, to handle the molecule hallucination, if the modified molecule is chemically invalid. Otherwise, its desired properties are computed and compared to the original one, establishing reliable domain feedback with correct direction and distance towards the objective, followed by a retrieved example, to guide the LLM to refine the modified molecule. We conduct experiments across both single- and multi-property objectives with 2 thresholds, where $\text{Re}^3$DF shows significant improvements. Particularly, for 20 single-property objectives, $\text{Re}^3$DF enhances Hit ratio by 16.96% and 20.76% under loose (\texttt{l}) and strict (\texttt{s}) thresholds, respectively. For 32 multi-property objectives, $\text{Re}^3$DF enhances Hit ratio by 6.04% and 5.25%.