In this work, we introduce a semiparametric token-sequence co-supervision training method. It trains a language model by simultaneously leveraging supervision from the traditional next token prediction loss which is calculated over the parametric token embedding space and the next sequence prediction loss which is calculated over the nonparametric sequence embedding space. The nonparametric sequence embedding space is constructed by a separate language model tasked to condense an input text into a single representative embedding. Our experiments demonstrate that a model trained via both supervisions consistently surpasses models trained via each supervision independently. Analysis suggests that this co-supervision encourages a broader generalization capability across the model. Especially, the robustness of parametric token space which is established during the pretraining step tends to effectively enhance the stability of nonparametric sequence embedding space, a new space established by another language model.

Generative models, widely utilized in various applications, can often struggle with prompts corresponding to partial tokens. This struggle stems from tokenization, where partial tokens fall out of distribution during inference, leading to incorrect or nonsensical outputs. This paper examines a technique to alleviate the tokenization artifact on text completion in generative models, maintaining performance even in regular non-subword cases. The method, termed token alignment, involves backtracking to the last complete tokens and ensuring the model's generation aligns with the prompt. This approach showcases marked improvement across many partial token scenarios, including nuanced cases like space-prefix and partial indentation, with only a minor time increase. The technique and analysis detailed in this paper contribute to the continuous advancement of generative models in handling partial inputs, bearing relevance for applications like code completion and text autocompletion. Generative models have shown remarkable efficacy in a range of applications. However, they have been observed to falter when dealing with partially provided inputs or subwords during text completion. For instance, a generative model might struggle to predict the remaining part of the word where a prompt ending in a subword often leads to incorrect or nonsensical outputs. This issue arises due to the artifact of tokenization where a partial token can be out-of-distribution during inference.

Large language model evaluation plays a pivotal role in the enhancement of its capacity. Previously, numerous methods for evaluating large language models have been proposed in this area. Despite their effectiveness, these existing works mainly focus on assessing objective questions, overlooking the capability to evaluate subjective questions which is extremely common for large language models. Additionally, these methods predominantly utilize centralized datasets for evaluation, with question banks concentrated within the evaluation platforms themselves. Moreover, the evaluation processes employed by these platforms often overlook personalized factors, neglecting to consider the individual characteristics of both the evaluators and the models being evaluated. To address these limitations, we propose a novel anonymous crowd-sourcing evaluation platform, BingJian, for large language models that employs a competitive scoring mechanism where users participate in ranking models based on their performance. This platform stands out not only for its support of centralized evaluations to assess the general capabilities of models but also for offering an open evaluation gateway. Through this gateway, users have the opportunity to submit their questions, testing the models on a personalized and potentially broader range of capabilities. Furthermore, our platform introduces personalized evaluation scenarios, leveraging various forms of human-computer interaction to assess large language models in a manner that accounts for individual user preferences and contexts. The demonstration of BingJian can be accessed at https://github.com/Mingyue-Cheng/Bingjian.

Large language models (LLMs) have achieved significant advances in recent days. Extensive efforts have been made before the public release of LLMs to align their behaviors with human values. The primary goal of alignment is to ensure their helpfulness, honesty and harmlessness. However, even meticulously aligned LLMs remain vulnerable to malicious manipulations such as jailbreaking, leading to unintended behaviors. The jailbreak is to intentionally develop a malicious prompt that escapes from the LLM security restrictions to produce uncensored detrimental contents. Previous works explore different jailbreak methods for red teaming LLMs, yet they encounter challenges regarding to effectiveness and scalability. In this work, we propose Tastle, a novel black-box jailbreak framework for automated red teaming of LLMs. We designed malicious content concealing and memory reframing with an iterative optimization algorithm to jailbreak LLMs, motivated by the research about the distractibility and over-confidence phenomenon of LLMs. Extensive experiments of jailbreaking both open-source and proprietary LLMs demonstrate the superiority of our framework in terms of effectiveness, scalability and transferability. We also evaluate the effectiveness of existing jailbreak defense methods against our attack and highlight the crucial need to develop more effective and practical defense strategies.

Large Language Models (LLMs) are poised to play an increasingly important role in our lives, providing assistance across a wide array of tasks. In the geospatial domain, LLMs have demonstrated the ability to answer generic questions, such as identifying a country's capital; nonetheless, their utility is hindered when it comes to answering fine-grained questions about specific places, such as grocery stores or restaurants, which constitute essential aspects of people's everyday lives. This is mainly because the places in our cities haven't been systematically fed into LLMs, so as to understand and memorize them. This study introduces a novel framework for fine-tuning a pre-trained model on city-specific data, to enable it to provide accurate recommendations, while minimizing hallucinations. We share our model, LAMP, and the data used to train it. We conduct experiments to analyze its ability to correctly retrieving spatial objects, and compare it to well-known open- and closed- source language models, such as GPT-4. Finally, we explore its emerging capabilities through a case study on day planning.

Language, a prominent human ability to express through sequential symbols, has been computationally mastered by recent advances of large language models (LLMs). By predicting the next word recurrently with huge neural models, LLMs have shown unprecedented capabilities in understanding and reasoning. Circuit, as the "language" of electronic design, specifies the functionality of an electronic device by cascade connections of logic gates. Then, can circuits also be mastered by a a sufficiently large "circuit model", which can conquer electronic design tasks by simply predicting the next logic gate? In this work, we take the first step to explore such possibilities. Two primary barriers impede the straightforward application of LLMs to circuits: their complex, non-sequential structure, and the intolerance of hallucination due to strict constraints (e.g., equivalence). For the first barrier, we encode a circuit as a memory-less, depth-first traversal trajectory, which allows Transformer-based neural models to better leverage its structural information, and predict the next gate on the trajectory as a circuit model. For the second barrier, we introduce an equivalence-preserving decoding process, which ensures that every token in the generated trajectory adheres to the specified equivalence constraints. Moreover, the circuit model can also be regarded as a stochastic policy to tackle optimization-oriented circuit design tasks. Experimentally, we trained a Transformer-based model of 88M parameters, named "Circuit Transformer", which demonstrates impressive performance in end-to-end logic synthesis. With Monte-Carlo tree search, Circuit Transformer significantly improves over resyn2 while retaining strict equivalence, showcasing the potential of generative AI in conquering electronic design challenges.

A substantial gap persists in understanding the reasons behind the exceptional performance of the Transformer architecture in NLP. A particularly unexplored area involves the mechanistic description of how the distribution of parameters evolves over time during training. In this work we suggest that looking at the time evolution of the statistic distribution of model parameters, and specifically at bifurcation effects, can help understanding the model quality, potentially reducing training costs and evaluation efforts and empirically showing the reasons behind the effectiveness of weights sparsification.

AcademiaOS is a first attempt to automate grounded theory development in qualitative research with large language models. Using recent large language models' language understanding, generation, and reasoning capabilities, AcademiaOS codes curated qualitative raw data such as interview transcripts and develops themes and dimensions to further develop a grounded theoretical model, affording novel insights. A user study (n=19) suggests that the system finds acceptance in the academic community and exhibits the potential to augment humans in qualitative research. AcademiaOS has been made open-source for others to build upon and adapt to their use cases.

Since the breakthrough of ChatGPT, large language models (LLMs) have garnered significant attention in the research community. With the development of LLMs, the question of text style transfer for conversational models has emerged as a natural extension, where chatbots may possess their own styles or even characters. However, standard evaluation metrics have not yet been established for this new settings. This paper aims to address this issue by proposing the LMStyle Benchmark, a novel evaluation framework applicable to chat-style text style transfer (C-TST), that can measure the quality of style transfer for LLMs in an automated and scalable manner. In addition to conventional style strength metrics, LMStyle Benchmark further considers a novel aspect of metrics called appropriateness, a high-level metrics take account of coherence, fluency and other implicit factors without the aid of reference samples. Our experiments demonstrate that the new evaluation methods introduced by LMStyle Benchmark have a higher correlation with human judgments in terms of appropriateness. Based on LMStyle Benchmark, we present a comprehensive list of evaluation results for popular LLMs, including LLaMA, Alpaca, and Vicuna, reflecting their stylistic properties, such as formality and sentiment strength, along with their appropriateness.

With the rapid development of Large Language Models (LLMs), we have witnessed intense competition among the major LLM products like ChatGPT, LLaMa, and Gemini. However, various issues (e.g. privacy leakage and copyright violation) of the training corpus still remain underexplored. For example, the Times sued OpenAI and Microsoft for infringing on its copyrights by using millions of its articles for training. From the perspective of LLM practitioners, handling such unintended privacy violations can be challenging. Previous work addressed the ``unlearning" problem of LLMs using gradient information, while they mostly introduced significant overheads like data preprocessing or lacked robustness. In this paper, contrasting with the methods based on first-order information, we revisit the unlearning problem via the perspective of second-order information (Hessian). Our unlearning algorithms, which are inspired by classic Newton update, are not only data-agnostic/model-agnostic but also proven to be robust in terms of utility preservation or privacy guarantee. Through a comprehensive evaluation with four NLP datasets as well as a case study on real-world datasets, our methods consistently show superiority over the first-order methods.