Building an intelligent and efficient medical assistant is still a challenging AI problem. The major limitation comes from the data modality scarceness, which reduces comprehensive patient perception. This demo paper presents the GigaPevt, the first multimodal medical assistant that combines the dialog capabilities of large language models with specialized medical models. Such an approach shows immediate advantages in dialog quality and metric performance, with a 1.18\% accuracy improvement in the question-answering task.

For a LLM to be trustworthy, its confidence level should be well-calibrated with its actual performance. While it is now common sense that LLM performances are greatly impacted by prompts, the confidence calibration in prompting LLMs has yet to be thoroughly explored. In this paper, we explore how different prompting strategies influence LLM confidence calibration and how it could be improved. We conduct extensive experiments on six prompting methods in the question-answering context and we observe that, while these methods help improve the expected LLM calibration, they also trigger LLMs to be over-confident when responding to some instances. Inspired by human cognition, we propose Fact-and-Reflection (FaR) prompting, which improves the LLM calibration in two steps. First, FaR elicits the known "facts" that are relevant to the input prompt from the LLM. And then it asks the model to "reflect" over them to generate the final answer. Experiments show that FaR prompting achieves significantly better calibration; it lowers the Expected Calibration Error by 23.5% on our multi-purpose QA tasks. Notably, FaR prompting even elicits the capability of verbally expressing concerns in less confident scenarios, which helps trigger retrieval augmentation for solving these harder instances.

Knowledge distillation (KD) has been widely adopted to compress large language models (LLMs). Existing KD methods investigate various divergence measures including the Kullback-Leibler (KL), reverse Kullback-Leibler (RKL), and Jensen-Shannon (JS) divergences. However, due to limitations inherent in their assumptions and definitions, these measures fail to deliver effective supervision when few distribution overlap exists between the teacher and the student. In this paper, we show that the aforementioned KL, RKL, and JS divergences respectively suffer from issues of mode-averaging, mode-collapsing, and mode-underestimation, which deteriorates logits-based KD for diverse NLP tasks. We propose the Sinkhorn Knowledge Distillation (SinKD) that exploits the Sinkhorn distance to ensure a nuanced and precise assessment of the disparity between teacher and student distributions. Besides, profit by properties of the Sinkhorn metric, we can get rid of sample-wise KD that restricts the perception of divergence in each teacher-student sample pair. Instead, we propose a batch-wise reformulation to capture geometric intricacies of distributions across samples in the high-dimensional space. Comprehensive evaluation on GLUE and SuperGLUE, in terms of comparability, validity, and generalizability, highlights our superiority over state-of-the-art methods on all kinds of LLMs with encoder-only, encoder-decoder, and decoder-only architectures.

The safety of Large Language Models (LLMs) has gained increasing attention in recent years, but there still lacks a comprehensive approach for detecting safety issues within LLMs' responses in an aligned, customizable and explainable manner. In this paper, we propose ShieldLM, an LLM-based safety detector, which aligns with general human safety standards, supports customizable detection rules, and provides explanations for its decisions. To train ShieldLM, we compile a large bilingual dataset comprising 14,387 query-response pairs, annotating the safety of responses based on various safety standards. Through extensive experiments, we demonstrate that ShieldLM surpasses strong baselines across four test sets, showcasing remarkable customizability and explainability. Besides performing well on standard detection datasets, ShieldLM has also been shown to be effective in real-world situations as a safety evaluator for advanced LLMs. We release ShieldLM at \url{https://github.com/thu-coai/ShieldLM} to support accurate and explainable safety detection under various safety standards, contributing to the ongoing efforts to enhance the safety of LLMs.

Pre-trained language models (PLMs) are shown to be vulnerable to minor word changes, which poses a big threat to real-world systems. While previous studies directly focus on manipulating word inputs, they are limited by their means of generating adversarial samples, lacking generalization to versatile real-world attack. This paper studies the basic structure of transformer-based PLMs, the self-attention (SA) mechanism. (1) We propose a powerful perturbation technique \textit{HackAttend}, which perturbs the attention scores within the SA matrices via meticulously crafted attention masks. We show that state-of-the-art PLMs fall into heavy vulnerability that minor attention perturbations $(1\%)$ can produce a very high attack success rate $(98\%)$. Our paper expands the conventional text attack of word perturbations to more general structural perturbations. (2) We introduce \textit{S-Attend}, a novel smoothing technique that effectively makes SA robust via structural perturbations. We empirically demonstrate that this simple yet effective technique achieves robust performance on par with adversarial training when facing various text attackers. Code is publicly available at \url{github.com/liongkj/HackAttend}.

Information on the depth of floodwater is crucial for rapid mapping of areas affected by floods. However, previous approaches for estimating floodwater depth, including field surveys, remote sensing, and machine learning techniques, can be time-consuming and resource-intensive. This paper presents an automated and fast approach for estimating floodwater depth from on-site flood photos. A pre-trained large multimodal model, GPT-4 Vision, was used specifically for estimating floodwater. The input data were flooding photos that contained referenced objects, such as street signs, cars, people, and buildings. Using the heights of the common objects as references, the model returned the floodwater depth as the output. Results show that the proposed approach can rapidly provide a consistent and reliable estimation of floodwater depth from flood photos. Such rapid estimation is transformative in flood inundation mapping and assessing the severity of the flood in near-real time, which is essential for effective flood response strategies.

Retrieval Augmented Generation (RAG) systems have seen huge popularity in augmenting Large-Language Model (LLM) outputs with domain specific and time sensitive data. Very recently a shift is happening from simple RAG setups that query a vector database for additional information with every user input to more sophisticated forms of RAG. However, different concrete approaches compete on mostly anecdotal evidence at the moment. In this paper we present a rigorous dataset creation and evaluation workflow to quantitatively compare different RAG strategies. We use a dataset created this way for the development and evaluation of a boolean agent RAG setup: A system in which a LLM can decide whether to query a vector database or not, thus saving tokens on questions that can be answered with internal knowledge. We publish our code and generated dataset online.

Automated story generation has been one of the long-standing challenges in NLP. Among all dimensions of stories, suspense is very common in human-written stories but relatively under-explored in AI-generated stories. While recent advances in large language models (LLMs) have greatly promoted language generation in general, state-of-the-art LLMs are still unreliable when it comes to suspenseful story generation. We propose a novel iterative-prompting-based planning method that is grounded in two theoretical foundations of story suspense from cognitive psychology and narratology. This theory-grounded method works in a fully zero-shot manner and does not rely on any supervised story corpora. To the best of our knowledge, this paper is the first attempt at suspenseful story generation with LLMs. Extensive human evaluations of the generated suspenseful stories demonstrate the effectiveness of our method.

Language model intelligence is revolutionizing the way we program materials simulations. However, the diversity of simulation scenarios renders it challenging to precisely transform human language into a tailored simulator. Here, using three functionalized types of language model, we propose a language-to-simulation (Lang2Sim) framework that enables interactive navigation on languaging a simulation engine, by taking a scenario instance of water sorption in porous matrices. Unlike line-by-line coding of a target simulator, the language models interpret each simulator as an assembly of invariant tool function and its variant input-output pair. Lang2Sim enables the precise transform of textual description by functionalizing and sequentializing the language models of, respectively, rationalizing the tool categorization, customizing its input-output combinations, and distilling the simulator input into executable format. Importantly, depending on its functionalized type, each language model features a distinct processing of chat history to best balance its memory limit and information completeness, thus leveraging the model intelligence to unstructured nature of human request. Overall, this work establishes language model as an intelligent platform to unlock the era of languaging a simulation engine.

AI alignment in the shape of Reinforcement Learning from Human Feedback (RLHF) is increasingly treated as a crucial ingredient for high performance large language models. Proximal Policy Optimization (PPO) has been positioned by recent literature as the canonical method for the RL part of RLHF. However, it involves both high computational cost and sensitive hyperparameter tuning. We posit that most of the motivational principles that led to the development of PPO are less of a practical concern in RLHF and advocate for a less computationally expensive method that preserves and even increases performance. We revisit the formulation of alignment from human preferences in the context of RL. Keeping simplicity as a guiding principle, we show that many components of PPO are unnecessary in an RLHF context and that far simpler REINFORCE-style optimization variants outperform both PPO and newly proposed "RL-free" methods such as DPO and RAFT. Our work suggests that careful adaptation to LLMs alignment characteristics enables benefiting from online RL optimization at low cost.