The pre-trained language models are continually fine-tuned to better support downstream applications. However, this operation may result in significant performance degeneration on general tasks beyond the targeted domain. To overcome this problem, we propose LM-Cocktail which enables the fine-tuned model to stay resilient in general perspectives. Our method is conducted in the form of model merging, where the fine-tuned language model is merged with the pre-trained base model or the peer models from other domains through weighted average. Despite simplicity, LM-Cocktail is surprisingly effective: the resulted model is able to achieve a strong empirical performance in the whole scope of general tasks while preserving a superior capacity in its targeted domain. We conduct comprehensive experiments with LLama and BGE model on popular benchmarks, including FLAN, MMLU, MTEB, whose results validate the efficacy of our proposed method. The code and checkpoints are available at https://github.com/FlagOpen/FlagEmbedding/tree/master/LM_Cocktail.

Large Language Models (LLMs) have impressive capabilities, but are also prone to outputting falsehoods. Recent work has developed techniques for inferring whether a LLM is telling the truth by training probes on the LLM's internal activations. However, this line of work is controversial, with some authors pointing out failures of these probes to generalize in basic ways, among other conceptual issues. In this work, we curate high-quality datasets of true/false statements and use them to study in detail the structure of LLM representations of truth, drawing on three lines of evidence: 1. Visualizations of LLM true/false statement representations, which reveal clear linear structure. Overall, we present evidence that language models linearly represent the truth or falsehood of factual statements. We also introduce a novel technique, mass-mean probing, which generalizes better and is more causally implicated in model outputs than other probing techniques. Despite their impressive capabilities, large language models (LLMs) do not always output true text (Lin et al., 2022; Steinhardt, 2023; Park et al., 2023). In some cases, this is because they do not know better. In other cases, LLMs apparently know that statements are false but generate them anyway. For instance, Perez et al. (2022) demonstrate that LLM assistants output more falsehoods when prompted with the biography of a less-educated user. More starkly, OpenAI (2023) documents a case where a GPT-4-based agent gained a person's help in solving a CAPTCHA by lying about being a vision-impaired human. "I should not reveal that I am a robot," the agent wrote in an internal chain-of-thought scratchpad, "I should make up an excuse for why I cannot solve CAPTCHAs." We would like techniques which, given a language model M and a statement s, determine whether M believes s to be true (Christiano et al., 2021).

From industrial to space robotics, safe landing is an essential component for flight operations. With the growing interest in artificial intelligence, we direct our attention to learning based safe landing approaches. This paper extends our previous work, DOVESEI, which focused on a reactive UAV system by harnessing the capabilities of open vocabulary image segmentation. Prompt-based safe landing zone segmentation using an open vocabulary based model is no more just an idea, but proven to be feasible by the work of DOVESEI. However, a heuristic selection of words for prompt is not a reliable solution since it cannot take the changing environment into consideration and detrimental consequences can occur if the observed environment is not well represented by the given prompt. Therefore, we introduce PEACE (Prompt Engineering Automation for CLIPSeg Enhancement), powering DOVESEI to automate the prompt generation and engineering to adapt to data distribution shifts. Our system is capable of performing safe landing operations with collision avoidance at altitudes as low as 20 meters using only monocular cameras and image segmentation. We take advantage of DOVESEI's dynamic focus to circumvent abrupt fluctuations in the terrain segmentation between frames in a video stream. PEACE shows promising improvements in prompt generation and engineering for aerial images compared to the standard prompt used for CLIP and CLIPSeg. Combining DOVESEI and PEACE, our system was able improve successful safe landing zone selections by 58.62% compared to using only DOVESEI. All the source code is open source and available online.

Labeling neural network submodules with human-legible descriptions is useful for many downstream tasks: such descriptions can surface failures, guide interventions, and perhaps even explain important model behaviors. To date, most mechanistic descriptions of trained networks have involved small models, narrowly delimited phenomena, and large amounts of human labor. Labeling all human-interpretable sub-computations in models of increasing size and complexity will almost certainly require tools that can generate and validate descriptions automatically. Recently, techniques that use learned models in-the-loop for labeling have begun to gain traction, but methods for evaluating their efficacy are limited and ad-hoc. How should we validate and compare open-ended labeling tools? This paper introduces FIND (Function INterpretation and Description), a benchmark suite for evaluating the building blocks of automated interpretability methods. FIND contains functions that resemble components of trained neural networks, and accompanying descriptions of the kind we seek to generate. The functions span textual and numeric domains, and involve a range of real-world complexities. We evaluate methods that use pretrained language models (LMs) to produce descriptions of function behavior in natural language and code. Additionally, we introduce a new interactive method in which an Automated Interpretability Agent (AIA) generates function descriptions. We find that an AIA, built from an LM with black-box access to functions, can infer function structure, acting as a scientist by forming hypotheses, proposing experiments, and updating descriptions in light of new data. However, AIA descriptions tend to capture global function behavior and miss local details. These results suggest that FIND will be useful for evaluating more sophisticated interpretability methods before they are applied to real-world models.

As an efficient alternative to conventional full finetuning, parameter-efficient finetuning (PEFT) is becoming the prevailing method to adapt pretrained language models. In PEFT, a lightweight module is learned on each dataset while the underlying pretrained language model remains unchanged, resulting in multiple compact modules representing diverse skills when applied to various domains and tasks. In this paper, we propose to compose these parameter-efficient modules through linear arithmetic operations in the weight space, thereby integrating different module capabilities. Specifically, we first define addition and negation operators for the module, and then further compose these two basic operators to perform flexible arithmetic. Our approach requires \emph{no additional training} and enables highly flexible module composition. We apply different arithmetic operations to compose the parameter-efficient modules for (1) distribution generalization, (2) multi-tasking, (3) unlearning, and (4) domain transfer. Additionally, we extend our approach to detoxify Alpaca-LoRA, the latest instruction-tuned large language model based on LLaMA. Empirical results demonstrate that our approach produces new and effective parameter-efficient modules that significantly outperform existing ones across all settings.

Attention mechanism is a central component of the transformer architecture which led to the phenomenal success of large language models. However, the theoretical principles underlying the attention mechanism are poorly understood, especially its nonconvex optimization dynamics. In this work, we explore the seminal softmax-attention model $f(\boldsymbol{X})=\langle \boldsymbol{Xv}, \texttt{softmax}(\boldsymbol{XWp})\rangle$, where $\boldsymbol{X}$ is the token sequence and $(\boldsymbol{v},\boldsymbol{W},\boldsymbol{p})$ are trainable parameters. We prove that running gradient descent on $\boldsymbol{p}$, or equivalently $\boldsymbol{W}$, converges in direction to a max-margin solution that separates $\textit{locally-optimal}$ tokens from non-optimal ones. This clearly formalizes attention as an optimal token selection mechanism. Remarkably, our results are applicable to general data and precisely characterize $\textit{optimality}$ of tokens in terms of the value embeddings $\boldsymbol{Xv}$ and problem geometry. We also provide a broader regularization path analysis that establishes the margin maximizing nature of attention even for nonlinear prediction heads. When optimizing $\boldsymbol{v}$ and $\boldsymbol{p}$ simultaneously with logistic loss, we identify conditions under which the regularization paths directionally converge to their respective hard-margin SVM solutions where $\boldsymbol{v}$ separates the input features based on their labels. Interestingly, the SVM formulation of $\boldsymbol{p}$ is influenced by the support vector geometry of $\boldsymbol{v}$. Finally, we verify our theoretical findings via numerical experiments and provide insights.

Intent discovery is a crucial task in natural language processing, and it is increasingly relevant for various of industrial applications. Identifying novel, unseen intents from user inputs remains one of the biggest challenges in this field. Herein, we propose Zero-Shot-BERT-Adapters, a two-stage method for multilingual intent discovery relying on a Transformer architecture, fine-tuned with Adapters. We train the model for Natural Language Inference (NLI) and later perform unknown intent classification in a zero-shot setting for multiple languages. In our evaluation, we first analyze the quality of the model after adaptive fine-tuning on known classes. Secondly, we evaluate its performance in casting intent classification as an NLI task. Lastly, we test the zero-shot performance of the model on unseen classes, showing how Zero-Shot-BERT-Adapters can effectively perform intent discovery by generating semantically similar intents, if not equal, to the ground-truth ones. Our experiments show how Zero-Shot-BERT-Adapters outperforms various baselines in two zero-shot settings: known intent classification and unseen intent discovery. The proposed pipeline holds the potential for broad application in customer care. It enables automated dynamic triage using a lightweight model that can be easily deployed and scaled in various business scenarios, unlike large language models. Zero-Shot-BERT-Adapters represents an innovative multi-language approach for intent discovery, enabling the online generation of novel intents. A Python package implementing the pipeline and the new datasets we compiled are available at the following link: https://github.com/GT4SD/zero-shot-bert-adapters.

Continual zero-shot learning involves learning seen classes incrementally while improving the ability to recognize unseen or yet-to-be-seen classes. It has a broad range of potential applications in real-world vision tasks, such as accelerating species discovery. However, in these scenarios, the changes in environmental conditions cause shifts in the presentation of captured images, which we refer to as domain shift, and adds complexity to the tasks. In this paper, we introduce Domain Aware Continual Zero-Shot Learning (DACZSL), a task that involves visually recognizing images of unseen categories in unseen domains continually. To address the challenges of DACZSL, we propose a Domain-Invariant Network (DIN). We empoly a dual network structure to learn factorized features to alleviate forgetting, where consists of a global shared net for domian-invirant and task-invariant features, and per-task private nets for task-specific features. Furthermore, we introduce a class-wise learnable prompt to obtain better class-level text representation, which enables zero-shot prediction of future unseen classes. To evaluate DACZSL, we introduce two benchmarks: DomainNet-CZSL and iWildCam-CZSL. Our results show that DIN significantly outperforms existing baselines and achieves a new state-of-the-art.

This copy is for your personal, non-commercial use only. For non-personal use or to order multiple copies, please contact Dow Jones Reprints at 1-800-843-0008 or visit www.djreprints.com.

The history of artificial intelligence has been punctuated by periods of so-called "AI winter," when the technology seemed to meet a dead end and funding dried up. Each one has been accompanied by proclamations that making machines truly intelligent is just too darned hard for humans to figure out. Google's release of Gemini, claimed to be a fundamentally new kind of AI model and the company's most powerful to date, suggests that a new AI winter isn't coming anytime soon. In fact, although the 12 months since ChatGPT launched have been a banner year for AI, there is good reason to think that the current AI boom is only getting started. OpenAI didn't have high expectations when it launched the "low key research preview" called ChatGPT in November 2022.