Prompt learning approaches have made waves in natural language processing by inducing better few-shot performance while they still follow a parametric-based learning paradigm; the oblivion and rote memorization problems in learning may encounter unstable generalization issues. Specifically, vanilla prompt learning may struggle to utilize atypical instances by rote during fully-supervised training or overfit shallow patterns with low-shot data. To alleviate such limitations, we develop RetroPrompt with the motivation of decoupling knowledge from memorization to help the model strike a balance between generalization and memorization. In contrast with vanilla prompt learning, RetroPrompt constructs an open-book knowledge-store from training instances and implements a retrieval mechanism during the process of input, training and inference, thus equipping the model with the ability to retrieve related contexts from the training corpus as cues for enhancement. Extensive experiments demonstrate that RetroPrompt can obtain better performance in both few-shot and zero-shot settings. Besides, we further illustrate that our proposed RetroPrompt can yield better generalization abilities with new datasets. Detailed analysis of memorization indeed reveals RetroPrompt can reduce the reliance of language models on memorization; thus, improving generalization for downstream tasks.

Language models have been shown to propagate social bias through their output, particularly in the representation of gender and ethnicity. This paper investigates gender and ethnicity biases in AI-generated occupational stories. Representation biases are measured before and after applying our proposed mitigation strategy, Bias Analysis and Mitigation through Explanation (BAME), revealing improvements in demographic representation ranging from 2% to 20%. BAME leverages model-generated explanations to inform targeted prompt engineering, effectively reducing biases without modifying model parameters. By analyzing stories generated across 25 occupational groups, three large language models (Claude 3.5 Sonnet, Llama 3.1 70B Instruct, and GPT-4 Turbo), and multiple demographic dimensions, we identify persistent patterns of overrepresentation and underrepresentation linked to training data stereotypes. Our findings demonstrate that guiding models with their own internal reasoning mechanisms can significantly enhance demographic parity, thereby contributing to the development of more transparent generative AI systems.

Computation-intensive tensor operators constitute over 90\% of the computations in Large Language Models (LLMs) and Deep Neural Networks.Automatically and efficiently generating high-performance tensor operators with hardware primitives is crucial for diverse and ever-evolving hardware architectures like RISC-V, ARM, and GPUs, as manually optimized implementation takes at least months and lacks portability.LLMs excel at generating high-level language codes, but they struggle to fully comprehend hardware characteristics and produce high-performance tensor operators. We introduce a tensor-operator auto-generation framework with a one-line user prompt (QiMeng-TensorOp), which enables LLMs to automatically exploit hardware characteristics to generate tensor operators with hardware primitives, and tune parameters for optimal performance across diverse hardware. Experimental results on various hardware platforms, SOTA LLMs, and typical tensor operators demonstrate that QiMeng-TensorOp effectively unleashes the computing capability of various hardware platforms, and automatically generates tensor operators of superior performance. Compared with vanilla LLMs, QiMeng-TensorOp achieves up to $1291 \times$ performance improvement. Even compared with human experts, QiMeng-TensorOp could reach $251 \%$ of OpenBLAS on RISC-V CPUs, and $124 \%$ of cuBLAS on NVIDIA GPUs. Additionally, QiMeng-TensorOp also significantly reduces development costs by $200 \times$ compared with human experts.

Prompt learning approaches have made waves in natural language processing by inducing better few-shot performance while they still follow a parametric-based learning paradigm; the oblivion and rote memorization problems in learning may encounter unstable generalization issues. Specifically, vanilla prompt learning may struggle to utilize atypical instances by rote during fully-supervised training or overfit shallow patterns with low-shot data. To alleviate such limitations, we develop RetroPrompt with the motivation of decoupling knowledge from memorization to help the model strike a balance between generalization and memorization. In contrast with vanilla prompt learning, RetroPrompt constructs an open-book knowledge-store from training instances and implements a retrieval mechanism during the process of input, training and inference, thus equipping the model with the ability to retrieve related contexts from the training corpus as cues for enhancement. Extensive experiments demonstrate that RetroPrompt can obtain better performance in both few-shot and zero-shot settings.

Traditionally, offline datasets have been used to evaluate task-oriented dialogue (TOD) models. These datasets lack context awareness, making them suboptimal benchmarks for conversational systems. In contrast, user-agents, which are context-aware, can simulate the variability and unpredictability of human conversations, making them better alternatives as evaluators. Prior research has utilized large language models (LLMs) to develop user-agents. Our work builds upon this by using LLMs to create user-agents for the evaluation of TOD systems. This involves prompting an LLM, using in-context examples as guidance, and tracking the user-goal state. Our evaluation of diversity and task completion metrics for the user-agents shows improved performance with the use of better prompts. Additionally, we propose methodologies for the automatic evaluation of TOD models within this dynamic framework.

Leading open-source large language models (LLMs) such as Llama-3.1-Instruct-405B are extremely capable at generating text, answering questions, and solving a variety of natural language understanding tasks. However, they incur higher inference cost and latency compared to smaller LLMs. Knowledge distillation provides a way to use outputs from these large, capable teacher models to train smaller student models which can be used for inference at lower cost and latency, while retaining comparable accuracy. We investigate the efficacy of distillation using the Llama-3.1-405B-Instruct teacher and the smaller Llama-3.1-8B-Instruct and Llama-3.1-70B-Instruct student models. Contributions of this work include (a) We evaluate the generalizability of distillation with the above Llama-3.1 teacher-student pairs across different tasks and datasets (b) We show that using synthetic data during distillation significantly improves the accuracy of 8B and 70B models, and when used with reasoning chains, even matches or surpasses the zero-shot accuracy of 405B model on some datasets (c) We empirically show that distillation enables 8B and 70B models to internalize 405B's reasoning ability by using only standard fine-tuning (without customizing any loss function). This allows cost and latency-efficient student model inference. (d) We show pitfalls in evaluation of distillation, and present task-specific evaluation, including both human and LLM-grading, and ground-truth based traditional accuracy benchmarks. This methodical study brings out the fundamental importance of synthetic data quality in knowledge distillation, and of combining multiple, task-specific ways of accuracy and quality evaluation in assessing the effectiveness of distillation.

In this paper, we present a very first study to investigate trust and ethical implications of on-device artificial intelligence (AI), focusing on ''small'' language models (SLMs) amenable for personal devices like smartphones. While on-device SLMs promise enhanced privacy, reduced latency, and improved user experience compared to cloud-based services, we posit that they might also introduce significant challenges and vulnerabilities compared to on-server counterparts. As part of our trust assessment study, we conduct a systematic evaluation of the state-of-the-art on-devices SLMs, contrasted to their on-server counterparts, based on a well-established trustworthiness measurement framework. Our results show on-device SLMs to be (statistically) significantly less trustworthy, specifically demonstrating more stereotypical, unfair and privacy-breaching behavior. Informed by these findings, we then perform our ethics assessment study by inferring whether SLMs would provide responses to potentially unethical vanilla prompts, collated from prior jailbreaking and prompt engineering studies and other sources. Strikingly, the on-device SLMs did answer valid responses to these prompts, which ideally should be rejected. Even more seriously, the on-device SLMs responded with valid answers without any filters and without the need for any jailbreaking or prompt engineering. These responses can be abused for various harmful and unethical scenarios including: societal harm, illegal activities, hate, self-harm, exploitable phishing content and exploitable code, all of which indicates the high vulnerability and exploitability of these on-device SLMs. Overall, our findings highlight gaping vulnerabilities in state-of-the-art on-device AI which seem to stem from resource constraints faced by these models and which may make typical defenses fundamentally challenging to be deployed in these environments.

Code translation between programming languages is a long-existing and critical task in software engineering, facilitating the modernization of legacy systems, ensuring cross-platform compatibility, and enhancing software performance. With the recent advances in large language models (LLMs) and their applications to code translation, there is an increasing need for comprehensive evaluation of these models. In this study, we empirically analyze the generated outputs of eleven popular instruct-tuned LLMs with parameters ranging from 1B up to 46.7B on 3,820 translation pairs across five languages, including C, C++, Go, Java, and Python. Our analysis found that between 26.4% and 73.7% of code translations produced by our evaluated LLMs necessitate post-processing, as these translations often include a mix of code, quotes, and text rather than being purely source code. Overlooking the output format of these models can inadvertently lead to underestimation of their actual performance. This is particularly evident when evaluating them with execution-based metrics such as Computational Accuracy (CA). Our results demonstrate that a strategic combination of prompt engineering and regular expression can effectively extract the source code from the model generation output. In particular, our method can help eleven selected models achieve an average Code Extraction Success Rate (CSR) of 92.73%. Our findings shed light on and motivate future research to conduct more reliable benchmarks of LLMs for code translation.

Prompt-tuning has demonstrated parameter-efficiency in fusing unimodal foundation models for multimodal tasks. However, its limited adaptivity and expressiveness lead to suboptimal performance when compared with other tuning methods. In this paper, we address this issue by disentangling the vanilla prompts to adaptively capture dataset-level and instance-level features. Building upon this disentanglement, we introduce the mixture of prompt experts (MoPE) technique to enhance expressiveness. MoPE leverages multimodal pairing priors to route the most effective prompt on a per-instance basis. Compared to vanilla prompting, our MoPE-based conditional prompting exhibits greater expressiveness for multimodal fusion, scaling better with the training data and the overall number of trainable parameters. We also study a regularization term for expert routing, leading to emergent expert specialization, where different experts focus on different concepts, enabling interpretable soft prompting. Extensive experiments across three multimodal datasets demonstrate that our method achieves state-of-the-art results, matching or even surpassing the performance of fine-tuning, while requiring only 0.8% of the trainable parameters. Code will be released: https://github.com/songrise/MoPE.

Large Language Models (LLMs) have garnered significant attention for their powerful ability in natural language understanding and reasoning. In this paper, we present a comprehensive empirical study to explore the performance of LLMs on misinformation detection tasks. This study stands as the pioneering investigation into the understanding capabilities of multiple LLMs regarding both content and propagation across social media platforms. Our empirical studies on five misinformation detection datasets show that LLMs with diverse prompts achieve comparable performance in text-based misinformation detection but exhibit notably constrained capabilities in comprehending propagation structure compared to existing models in propagation-based misinformation detection. Besides, we further design four instruction-tuned strategies to enhance LLMs for both content and propagation-based misinformation detection. These strategies boost LLMs to actively learn effective features from multiple instances or hard instances, and eliminate irrelevant propagation structures, thereby achieving better detection performance. Extensive experiments further demonstrate LLMs would play a better capacity in content and propagation structure under these proposed strategies and achieve promising detection performance. These findings highlight the potential ability of LLMs to detect misinformation.