Correct reasoning steps are important for language models to achieve high performance on many tasks, such as commonsense reasoning, question answering, and mathematical problem-solving [Wei et al., 2022, Kojima et al., 2022, Suzgun et al., 2022]. One way to elicit reasoning is through the chain-of-thought (CoT) method Wei et al. [2022], Kojima et al. [2022], which asks the model to provide step-by-step reasoning. Another approach encourages the model to provide similar problems Yasunaga et al. [2024] as the query, indirectly compelling the model to first understand the original query. Similarly, repeating and rephrasing the query Deng et al. [2023], Mekala et al. [2023] requires the model to first understand the problem and then modify the query into its own words. This rephrasing might help simplify the problem for the model. Additionally, reasoning can be generated by indirectly providing reasoning examples in demonstrations, referred to as in-context learning (ICL) Brown et al. [2020], Min et al. [2022], Xie et al. [2021]. While these methods have demonstrated significant performance improvements, language models are still prone to errors due to incorrect context understanding or analytical steps. Furthermore, they are subject to instability when requests are paraphrased. This instability is particularly concerning in the context of adversarial prompts, where recent research [Zou et al., 2023, Zeng et al., 2024] has shown that adversaries can intentionally rewrite prompts to coax safety-aligned language models into generating objectionable content that they would not generate otherwise.

The generation of corner cases has become increasingly crucial for efficiently testing autonomous vehicles prior to road deployment. However, existing methods struggle to accommodate diverse testing requirements and often lack the ability to generalize to unseen situations, thereby reducing the convenience and usability of the generated scenarios. A method that facilitates easily controllable scenario generation for efficient autonomous vehicles (AV) testing with realistic and challenging situations is greatly needed. To address this, we proposed OmniTester: a multimodal Large Language Model (LLM) based framework that fully leverages the extensive world knowledge and reasoning capabilities of LLMs. OmniTester is designed to generate realistic and diverse scenarios within a simulation environment, offering a robust solution for testing and evaluating AVs. In addition to prompt engineering, we employ tools from Simulation of Urban Mobility to simplify the complexity of codes generated by LLMs. Furthermore, we incorporate Retrieval-Augmented Generation and a self-improvement mechanism to enhance the LLM's understanding of scenarios, thereby increasing its ability to produce more realistic scenes. In the experiments, we demonstrated the controllability and realism of our approaches in generating three types of challenging and complex scenarios. Additionally, we showcased its effectiveness in reconstructing new scenarios described in crash report, driven by the generalization capability of LLMs.

With the remarkable success achieved by Multimodal Large Language Models (MLLMs), numerous benchmarks have been designed to assess MLLMs' ability to guide their development in image perception tasks (e.g., image captioning and visual question answering). Moreover, these benchmarks contain many overly simple problems or challenging samples, which do not effectively differentiate the capabilities among various MLLMs. To address these challenges, we propose a pipeline to process the existing benchmarks, which consists of two modules (i.e., Semi-Automated Screening Process and Eliminating Answer Leakage). The Semi-Automated Screening Process filters out samples that cannot distinguish the model's capabilities by synthesizing various MLLMs and manually evaluating them. The Eliminate Answer Leakage filters the samples whose answers can be inferred without images. Finally, we curate the LIME-M: Less Is More for Evaluation of Multimodal LLMs, which is a lightweight Multimodal benchmark and can more effectively evaluate the performance of different models. Our experiments demonstrate that 1. LIME-M can better distinguish the performance of different MLLMs within less sample numbers (24% of original) and time spent (23% of original); 2. LIME-M eliminates answer leakage and mainly focuses on the information within images; 3. The current automatic metric (i.e., CIDEr) is insufficient for evaluating MLLMs' capabilities in captioning. Moreover, we find that removing the caption task score when calculating the overall score demonstrates a more precise reflection of model performance differences. In order to better understand the model's capabilities and guide to address the shortcomings of MLLMs, researchers have developed numerous benchmarks for various tasks (Antol et al., 2015; Wei et al., 2023; Fu et al., 2023; Yue et al., 2024; Wu et al., 2024a). These benchmarks thoroughly explore the capabilities of MLLMs in various tasks such as image captioning, image question answering, and Multimodal retrieving. However, existing MLLM benchmarks and unified evaluation frameworks cannot effectively and efficiently reflect the ability of MLLMs.

Human motion generation is an important area of research in many fields. In this work, we tackle the problem of motion stitching and in-betweening. Current methods either require manual efforts, or are incapable of handling longer sequences. To address these challenges, we propose a diffusion model with a transformer-based denoiser to generate realistic human motion. Our method demonstrated strong performance in generating in-betweening sequences, transforming a variable number of input poses into smooth and realistic motion sequences consisting of 75 frames at 15 fps, resulting in a total duration of 5 seconds. We present the performance evaluation of our method using quantitative metrics such as Frechet Inception Distance (FID), Diversity, and Multimodality, along with visual assessments of the generated outputs.

The progress in text summarization techniques has been remarkable. However the task of accurately extracting and summarizing necessary information from highly specialized documents such as research papers has not been sufficiently investigated. We are focusing on the task of extracting research questions (RQ) from research papers and construct a new dataset consisting of machine learning papers, RQ extracted from these papers by GPT-4, and human evaluations of the extracted RQ from multiple perspectives. Using this dataset, we systematically compared recently proposed LLM-based evaluation functions for summarizations, and found that none of the functions showed sufficiently high correlations with human evaluations. We expect our dataset provides a foundation for further research on developing better evaluation functions tailored to the RQ extraction task, and contribute to enhance the performance of the task. The dataset is available at https://github.com/auto-res/PaperRQ-HumanAnno-Dataset.

The Extended Church-Turing Thesis (ECTT) posits that all effective information processing, including unbounded and non-uniform interactive computations, can be described in terms of interactive Turing machines with advice. Does this assertion also apply to the abilities of contemporary large language models (LLMs)? From a broader perspective, this question calls for an investigation of the computational power of LLMs by the classical means of computability and computational complexity theory, especially the theory of automata. Along these lines, we establish a number of fundamental results. Firstly, we argue that any fixed (non-adaptive) LLM is computationally equivalent to a, possibly very large, deterministic finite-state transducer. This characterizes the base level of LLMs. We extend this to a key result concerning the simulation of space-bounded Turing machines by LLMs. Secondly, we show that lineages of evolving LLMs are computationally equivalent to interactive Turing machines with advice. The latter finding confirms the validity of the ECTT for lineages of LLMs. From a computability viewpoint, it also suggests that lineages of LLMs possess super-Turing computational power. Consequently, in our computational model knowledge generation is in general a non-algorithmic process realized by lineages of LLMs. Finally, we discuss the merits of our findings in the broader context of several related disciplines and philosophies.

Developing a consistent and reliable AI game master for text-based games is a challenging task due to the limitations of large language models (LLMs) and the complexity of the game master's role. This paper presents a novel approach to enhance AI game masters by leveraging function calling in the context of the table-top role-playing game "Jim Henson's Labyrinth: The Adventure Game." Our methodology involves integrating game-specific controls through functions, which we show improves the narrative quality and state update consistency of the AI game master. The experimental results, based on human evaluations and unit tests, demonstrate the effectiveness of our approach in enhancing gameplay experience and maintaining coherence with the game state. This work contributes to the advancement of game AI and interactive storytelling, offering insights into the design of more engaging and consistent AI-driven game masters.

We introduce a novel benchmark for evaluating the role-playing capabilities of language models. Our approach leverages language models themselves to emulate users in dynamic, multi-turn conversations and to assess the resulting dialogues. The framework consists of three main components: a player model assuming a specific character role, an interrogator model simulating user behavior, and a judge model evaluating conversation quality. We conducted experiments comparing automated evaluations with human annotations to validate our approach, demonstrating strong correlations across multiple criteria. This work provides a foundation for a robust and dynamic evaluation of model capabilities in interactive scenarios.

Reward models (RM) capture the values and preferences of humans and play a central role in Reinforcement Learning with Human Feedback (RLHF) to align pretrained large language models (LLMs). Traditionally, training these models relies on extensive human-annotated preference data, which poses significant challenges in terms of scalability and cost. To overcome these limitations, we propose Semi-Supervised Reward Modeling (SSRM), an approach that enhances RM training using unlabeled data. Given an unlabeled dataset, SSRM involves three key iterative steps: pseudo-labeling unlabeled examples, selecting high-confidence examples through a confidence threshold, and supervised finetuning on the refined dataset. Across extensive experiments on various model configurations, we demonstrate that SSRM significantly improves reward models without incurring additional labeling costs. Notably, SSRM can achieve performance comparable to models trained entirely on labeled data of equivalent volumes. Overall, SSRM substantially reduces the dependency on large volumes of human-annotated data, thereby decreasing the overall cost and time involved in training effective reward models.

The occurrence of bubbles in pipeline parallelism is an inherent limitation that can account for more than 40% of the large language model (LLM) training time and is one of the main reasons for the underutilization of GPU resources in LLM training. Harvesting these bubbles for GPU side tasks can increase resource utilization and reduce training costs but comes with challenges. First, because bubbles are discontinuous with various shapes, programming side tasks becomes difficult while requiring excessive engineering effort. Second, a side task can compete with pipeline training for GPU resources and incur significant overhead. To address these challenges, we propose FreeRide, a system designed to harvest bubbles in pipeline parallelism for side tasks. FreeRide provides programmers with interfaces to implement side tasks easily, manages bubbles and side tasks during pipeline training, and controls access to GPU resources by side tasks to reduce overhead. We demonstrate that FreeRide achieves 7.8% average cost savings with a negligible overhead of about 1% in training LLMs while serving model training, graph analytics, and image processing side tasks.