Employing Large Language Models (LLMs) to assess the quality of generated responses, such as prompting instruct-tuned models or fine-tuning judge models, has become a widely adopted evaluation method. It is also known that such evaluators are vulnerable to biases, such as favoring longer responses. While it is important to overcome this problem, the specifics of these biases remain under-explored. In this work, we qualitatively identify six types of biases inherent in various judge models. We propose EvalBiasBench as a meta-evaluation collection of hand-crafted test cases for each bias type. Additionally, we present de-biasing dataset construction methods and the associated preference dataset OffsetBias. Experimental results demonstrate that fine-tuning on our dataset significantly enhances the robustness of judge models against biases and improves performance across most evaluation scenarios. We release our datasets and the fine-tuned judge model to public.

Diplomacy is one of the most sophisticated activities in human society. The complex interactions among multiple parties/ agents involve various abilities like social reasoning, negotiation arts, and long-term strategy planning. Previous AI agents surely have proved their capability of handling multi-step games and larger action spaces on tasks involving multiple agents. However, diplomacy involves a staggering magnitude of decision spaces, especially considering the negotiation stage required. Recently, LLM agents have shown their potential for extending the boundary of previous agents on a couple of applications, however, it is still not enough to handle a very long planning period in a complex multi-agent environment. Empowered with cutting-edge LLM technology, we make the first stab to explore AI's upper bound towards a human-like agent for such a highly comprehensive multi-agent mission by combining three core and essential capabilities for stronger LLM-based societal agents: 1) strategic planner with memory and reflection; 2) goal-oriented negotiate with social reasoning; 3) augmenting memory by self-play games to self-evolving without any human in the loop.

Serving large language models (LLMs) in production can incur substantial costs, which has prompted recent advances in inference system optimizations. Today, these systems are evaluated against conventional latency and throughput metrics (eg. TTFT, TBT, Normalised Latency and TPOT). However, these metrics fail to fully capture the nuances of LLM inference, leading to an incomplete assessment of user-facing performance crucial for real-time applications such as chat and translation. In this paper, we first identify the pitfalls of current performance metrics in evaluating LLM inference systems. We then propose Metron, a comprehensive performance evaluation framework that includes fluidity-index -- a novel metric designed to reflect the intricacies of the LLM inference process and its impact on real-time user experience. Finally, we evaluate various existing open-source platforms and model-as-a-service offerings using Metron, discussing their strengths and weaknesses. Metron is available at https://github.com/project-metron/metron.

As transformer-based object detection models progress, their impact in critical sectors like autonomous vehicles and aviation is expected to grow. Soft errors causing bit flips during inference have significantly impacted DNN performance, altering predictions. Traditional range restriction solutions for CNNs fall short for transformers. This study introduces the Global Clipper and Global Hybrid Clipper, effective mitigation strategies specifically designed for transformer-based models. It significantly enhances their resilience to soft errors and reduces faulty inferences to ~ 0\%. We also detail extensive testing across over 64 scenarios involving two transformer models (DINO-DETR and Lite-DETR) and two CNN models (YOLOv3 and SSD) using three datasets, totalling approximately 3.3 million inferences, to assess model robustness comprehensively. Moreover, the paper explores unique aspects of attention blocks in transformers and their operational differences from CNNs.

Despite the impressive capabilities of Large Language Models (LLMs) in various tasks, their vulnerability to unsafe prompts remains a critical issue. These prompts can lead LLMs to generate responses on illegal or sensitive topics, posing a significant threat to their safe and ethical use. Existing approaches attempt to address this issue using classification models, but they have several drawbacks. With the increasing complexity of unsafe prompts, similarity search-based techniques that identify specific features of unsafe prompts provide a more robust and effective solution to this evolving problem. This paper investigates the potential of sentence encoders to distinguish safe from unsafe prompts, and the ability to classify various unsafe prompts according to a safety taxonomy. We introduce new pairwise datasets and the Categorical Purity (CP) metric to measure this capability. Our findings reveal both the effectiveness and limitations of existing sentence encoders, proposing directions to improve sentence encoders to operate as more robust safety detectors. Our code is available at https://github.com/JwdanielJung/Safe-Embed.

Radiation therapy (RT) is one of the most effective treatments for cancer, and its success relies on the accurate delineation of targets. However, target delineation is a comprehensive medical decision that currently relies purely on manual processes by human experts. Manual delineation is time-consuming, laborious, and subject to interobserver variations. Although the advancements in artificial intelligence (AI) techniques have significantly enhanced the auto-contouring of normal tissues, accurate delineation of RT target volumes remains a challenge. In this study, we propose a visual language model-based RT target volume auto-delineation network termed Radformer. The Radformer utilizes a hierarichal vision transformer as the backbone and incorporates large language models to extract text-rich features from clinical data. We introduce a visual language attention module (VLAM) for integrating visual and linguistic features for language-aware visual encoding (LAVE). The Radformer has been evaluated on a dataset comprising 2985 patients with head-and-neck cancer who underwent RT. Metrics, including the Dice similarity coefficient (DSC), intersection over union (IOU), and 95th percentile Hausdorff distance (HD95), were used to evaluate the performance of the model quantitatively. Our results demonstrate that the Radformer has superior segmentation performance compared to other state-of-the-art models, validating its potential for adoption in RT practice.

The early detection of suicide risk is important since it enables the intervention to prevent potential suicide attempts. This paper studies the automatic detection of suicide risk based on spontaneous speech from adolescents, and collects a Mandarin dataset with 15 hours of suicide speech from more than a thousand adolescents aged from ten to eighteen for our experiments. To leverage the diverse acoustic and linguistic features embedded in spontaneous speech, both the Whisper speech model and textual large language models (LLMs) are used for suicide risk detection. Both all-parameter finetuning and parameter-efficient finetuning approaches are used to adapt the pre-trained models for suicide risk detection, and multiple audio-text fusion approaches are evaluated to combine the representations of Whisper and the LLM. The proposed system achieves a detection accuracy of 0.807 and an F1-score of 0.846 on the test set with 119 subjects, indicating promising potential for real suicide risk detection applications.

To support software developers in understanding and maintaining programs, various automatic (source) code summarization techniques have been proposed to generate a concise natural language summary (i.e., comment) for a given code snippet. Recently, the emergence of large language models (LLMs) has led to a great boost in the performance of code-related tasks. In this paper, we undertake a systematic and comprehensive study on code summarization in the era of LLMs, which covers multiple aspects involved in the workflow of LLM-based code summarization. Specifically, we begin by examining prevalent automated evaluation methods for assessing the quality of summaries generated by LLMs and find that the results of the GPT-4 evaluation method are most closely aligned with human evaluation. Then, we explore the effectiveness of five prompting techniques (zero-shot, few-shot, chain-of-thought, critique, and expert) in adapting LLMs to code summarization tasks. Contrary to expectations, advanced prompting techniques may not outperform simple zero-shot prompting. Next, we investigate the impact of LLMs' model settings (including top\_p and temperature parameters) on the quality of generated summaries. We find the impact of the two parameters on summary quality varies by the base LLM and programming language, but their impacts are similar. Moreover, we canvass LLMs' abilities to summarize code snippets in distinct types of programming languages. The results reveal that LLMs perform suboptimally when summarizing code written in logic programming languages compared to other language types. Finally, we unexpectedly find that CodeLlama-Instruct with 7B parameters can outperform advanced GPT-4 in generating summaries describing code implementation details and asserting code properties. We hope that our findings can provide a comprehensive understanding of code summarization in the era of LLMs.

We present and evaluate a method called grammar masking, which is used to guide large language models (LLMs) toward producing syntactically correct models for a given context-free grammar. Prompt engineering methods such as few-shot learning or priming can be used to improve the chances of an LLM producing correct syntax, but the more complex the grammar, the more time-consuming and less promising these methods become. Previous work is focused primarily on the usage of either language model training or prompt engineering. In this work, a method is presented that restricts the output to a given grammar using constrained decoding to ensure the output adheres to a valid syntax. We use several DSLs built with MontiCore and task multiple LLMs to produce models with and without constrained decoding. A corresponding parser is used to confirm the syntactic correctness of each model. We show that grammar masking can dramatically improve the modeling capabilities of several LLMs, reducing the need for well-refined prompting while increasing the chance of producing correct models.

The effectiveness of large language models (LLMs) is often hindered by duplicated data in their extensive pre-training datasets. Current approaches primarily focus on detecting and removing duplicates, which risks the loss of valuable information and neglects the varying degrees of duplication. To address this, we propose a soft deduplication method that maintains dataset integrity while selectively reducing the sampling weight of data with high commonness. Central to our approach is the concept of "data commonness", a metric we introduce to quantify the degree of duplication by measuring the occurrence probabilities of samples using an n-gram model. Empirical analysis shows that this method significantly improves training efficiency, achieving comparable perplexity scores with at least a 26% reduction in required training steps. Additionally, it enhances average few-shot downstream accuracy by 1.77% when trained for an equivalent duration. Importantly, this approach consistently improves performance, even on rigorously deduplicated datasets, indicating its potential to complement existing methods and become a standard pre-training process for LLMs.