We investigate challenges and possibilities of formal reasoning for encoder-only transformers (EOT), meaning sound and complete methods for verifying or interpreting behaviour. In detail, we condense related formal reasoning tasks in the form of a naturally occurring satisfiability problem (SAT). We find that SAT is undecidable if we consider EOT, commonly considered in the expressiveness community. Furthermore, we identify practical scenarios where SAT is decidable and establish corresponding complexity bounds. Besides trivial cases, we find that quantized EOT, namely those restricted by some fixed-width arithmetic, lead to the decidability of SAT due to their limited attention capabilities. However, the problem remains difficult, as we establish those scenarios where SAT is NEXPTIME-hard and those where we can show that it is solvable in NEXPTIME for quantized EOT. To complement our theoretical results, we put our findings and their implications in the overall perspective of formal reasoning.

While large multi-modal models (LMMs) have exhibited impressive capabilities across diverse tasks, their effectiveness in handling complex tasks has been limited by the prevailing single-step reasoning paradigm. To this end, this paper proposes VoCoT, a multi-step Visually grounded object-centric Chain-of-Thought reasoning framework tailored for inference with LMMs. VoCoT is characterized by two key features: (1) object-centric reasoning paths that revolve around cross-modal shared object-level information, and (2) visually grounded representation of object concepts in a multi-modal interleaved and aligned manner, which effectively bridges the modality gap within LMMs during long-term generation. Additionally, we construct an instruction dataset to facilitate LMMs in adapting to reasoning with VoCoT. By introducing VoCoT into the prevalent open-source LMM architecture, we introduce VolCano. With only 7B parameters and limited input resolution, VolCano demonstrates excellent performance across various scenarios, surpassing SOTA models, including GPT-4V, in tasks requiring complex reasoning. Our code, data and model will be available at https://github.com/RupertLuo/VoCoT.

Transformer architectures have shown promising results in time series processing. However, despite recent advances in subquadratic attention mechanisms or state-space models, processing very long sequences still imposes significant computational requirements. Token merging, which involves replacing multiple tokens with a single one calculated as their linear combination, has shown to considerably improve the throughput of vision transformer architectures while maintaining accuracy. In this work, we go beyond computer vision and perform the first investigations of token merging in time series analysis on both time series transformers and state-space models. To effectively scale token merging to long sequences, we introduce local merging, a domain-specific token merging algorithm that selectively combines tokens within a local neighborhood, adjusting the computational complexity from linear to quadratic based on the neighborhood size. Our comprehensive empirical evaluation demonstrates that token merging offers substantial computational benefits with minimal impact on accuracy across various models and datasets. On the recently proposed Chronos foundation model, we achieve accelerations up to 5400 % with only minor accuracy degradations.

Existing first-order optimizers mainly include two branches: classical optimizers represented by Stochastic Gradient Descent (SGD) and adaptive optimizers represented by Adam, along with their many derivatives. The debate over the merits and demerits of these two types of optimizers has persisted for a decade. In practical experience, it is generally considered that SGD is more suitable for tasks like Computer Vision(CV), while adaptive optimizers are widely used in tasks with sparse gradients, such as Large Language Models(LLM). Although adaptive optimizers always offer better convergence speeds in almost all tasks, they can lead to over-fitting in some cases, resulting in poorer generalization performance compared to SGD in certain tasks. Even in Large Language Models, Adam continues to face challenges, and its original strategy may not always have an advantage due to the introduction of improvements such as gradient clipping. With a wide variety of optimization methods available, it is essential to introduce a unified, interpretable theory. This paper will discuss under the framework of first-order optimizers and, through the intervention of the balance theory, will for the first time propose a unified strategy to integrate all first-order optimization methods.

Understanding how sentences are internally represented in the human brain, as well as in large language models (LLMs) such as ChatGPT, is a major challenge for cognitive science. Classic linguistic theories propose that the brain represents a sentence by parsing it into hierarchically organized constituents. In contrast, LLMs do not explicitly parse linguistic constituents and their latent representations remains poorly explained. Here, we demonstrate that humans and LLMs construct similar latent representations of hierarchical linguistic constituents by analyzing their behaviors during a novel one-shot learning task, in which they infer which words should be deleted from a sentence. Both humans and LLMs tend to delete a constituent, instead of a nonconstituent word string. In contrast, a naive sequence processing model that has access to word properties and ordinal positions does not show this property. Based on the word deletion behaviors, we can reconstruct the latent constituency tree representation of a sentence for both humans and LLMs. These results demonstrate that a latent tree-structured constituency representation can emerge in both the human brain and LLMs.

Large language models (LLMs) reflect societal norms and biases, especially about gender. While societal biases and stereotypes have been extensively researched in various NLP applications, there is a surprising gap for emotion analysis. However, emotion and gender are closely linked in societal discourse. E.g., women are often thought of as more empathetic, while men's anger is more socially accepted. To fill this gap, we present the first comprehensive study of gendered emotion attribution in five state-of-the-art LLMs (open- and closed-source). We investigate whether emotions are gendered, and whether these variations are based on societal stereotypes. We prompt the models to adopt a gendered persona and attribute emotions to an event like 'When I had a serious argument with a dear person'. We then analyze the emotions generated by the models in relation to the gender-event pairs. We find that all models consistently exhibit gendered emotions, influenced by gender stereotypes. These findings are in line with established research in psychology and gender studies. Our study sheds light on the complex societal interplay between language, gender, and emotion. The reproduction of emotion stereotypes in LLMs allows us to use those models to study the topic in detail, but raises questions about the predictive use of those same LLMs for emotion applications.

Post-training quantization (PTQ) techniques applied to weights, activations, and the KV cache greatly reduce memory usage, latency, and power consumption of Large Language Models (LLMs), but may lead to large quantization errors when outliers are present. Recent findings suggest that rotating activation or weight matrices helps remove outliers and benefits quantization. In this work, we identify a collection of applicable rotation parameterizations that lead to identical outputs in full-precision Transformer architectures, and find that some random rotations lead to much better quantization than others, with an up to 13 points difference in downstream zero-shot reasoning performance. As a result, we propose SpinQuant that optimizes (or learns) the rotation matrices with Cayley optimization on a small validation set. With 4-bit quantization of weight, activation, and KV-cache, SpinQuant narrows the accuracy gap on zero-shot reasoning tasks with full precision to merely 2.9 points on the LLaMA-2 7B model, surpassing LLM-QAT by 19.1 points and SmoothQuant by 25.0 points. SpinQuant also outperforms concurrent work QuaRot, which applies random rotations to remove outliers. In particular, for LLaMA-2 7B/LLaMA-3 8B models that are hard to quantize, SpinQuant reduces the gap to full precision by 30.2%/34.1% relative to QuaRot.

The emergence of online recruitment services has revolutionized the traditional landscape of job seeking and recruitment, necessitating the development of high-quality industrial applications to improve person-job fitting. Existing methods generally rely on modeling the latent semantics of resumes and job descriptions and learning a matching function between them. Inspired by the powerful role-playing capabilities of Large Language Models (LLMs), we propose to introduce a mock interview process between LLM-played interviewers and candidates. The mock interview conversations can provide additional evidence for candidate evaluation, thereby augmenting traditional person-job fitting based solely on resumes and job descriptions. However, characterizing these two roles in online recruitment still presents several challenges, such as developing the skills to raise interview questions, formulating appropriate answers, and evaluating two-sided fitness. To this end, we propose MockLLM, a novel applicable framework that divides the person-job matching process into two modules: mock interview generation and two-sided evaluation in handshake protocol, jointly enhancing their performance through collaborative behaviors between interviewers and candidates. We design a role-playing framework as a multi-role and multi-behavior paradigm to enable a single LLM agent to effectively behave with multiple functions for both parties. Moreover, we propose reflection memory generation and dynamic prompt modification techniques to refine the behaviors of both sides, enabling continuous optimization of the augmented additional evidence. Extensive experimental results show that MockLLM can achieve the best performance on person-job matching accompanied by high mock interview quality, envisioning its emerging application in real online recruitment in the future.

Recent advancements in Large Vision-Language Models (VLMs) have underscored their superiority in various multimodal tasks. However, the adversarial robustness of VLMs has not been fully explored. Existing methods mainly assess robustness through unimodal adversarial attacks that perturb images, while assuming inherent resilience against text-based attacks. Different from existing attacks, in this work we propose a more comprehensive strategy that jointly attacks both text and image modalities to exploit a broader spectrum of vulnerability within VLMs. Specifically, we propose a dual optimization objective aimed at guiding the model to generate affirmative responses with high toxicity. Our attack method begins by optimizing an adversarial image prefix from random noise to generate diverse harmful responses in the absence of text input, thus imbuing the image with toxic semantics. Subsequently, an adversarial text suffix is integrated and co-optimized with the adversarial image prefix to maximize the probability of eliciting affirmative responses to various harmful instructions. The discovered adversarial image prefix and text suffix are collectively denoted as a Universal Master Key (UMK). When integrated into various malicious queries, UMK can circumvent the alignment defenses of VLMs and lead to the generation of objectionable content, known as jailbreaks. The experimental results demonstrate that our universal attack strategy can effectively jailbreak MiniGPT-4 with a 96% success rate, highlighting the vulnerability of VLMs and the urgent need for new alignment strategies.

Recent advancements in AI have sped up the evolution of versatile robot designs. Chess provides a standardized environment that allows for the evaluation of the influence of robot behaviors on human behavior. This article presents an open-source chess robot for humanrobot interaction (HRI) research, specifically focusing on verbal and non-verbal interactions. OpenChessRobot recognizes chess pieces using computer vision, executes moves, and interacts with the human player using voice and robotic gestures. We detail the software design, provide quantitative evaluations of the robot's efficacy and offer a guide for its reproducibility. Keywords: Artificial Intelligence, Chess, Human-robot Interaction, Open-source, Transfer Learning 1. Introduction Robots are becoming increasingly common across a variety of traditionally human-controlled domains. Examples range from automated mowers that maintain community lawns to robots in assembly lines and agricultural settings. Recent scientific advancements in AI have enabled new opportunities for intelligent sensing, reasoning, and acting by robots. In particular, the rapid development of large language models, such as ChatGPT, and vision-language models, have lowered the barrier of human-to-robot communication by being able to transform text and images into interpretable actions or vice versa. As technology advances, it is likely that robots will attain greater capabilities and will be able to tackle tasks previously within the exclusive realm of human expertise. This ongoing evolution may also lead to closer and more productive interactions between humans and robots. At the same time, integrating different AI-based robotic components remains a challenge, and the human-robot interaction (HRI) field lags in terms of endorsing reproducibility principles (Gunes et al., 2022). Encouraging transparent and reproducible research, therefore, remains an ongoing task. Furthermore, chess has played an important role in advancing the field of AI, starting with Claude Shannon's chess-playing algorithm (Shannon, 1950) to the success of IBM's Deep Blue (Campbell et al., 2002) and DeepMind's self-play learning algorithm (Silver et al., 2018). In this paper, we incorporate modern AI algorithms into the design of a chess-playing robot to be used for studying HRI. HRI research may benefit from a chess-based setup because the game of chess provides a controlled rule-based environment in which the impact of robots on human players can be precisely measured.