Fine-tuning large language models (LLMs) on multi-task instruction-following data has been proven to be a powerful learning paradigm for improving their zero-shot capabilities on new tasks. Recent works about high-quality instruction-following data generation and selection require amounts of human labor to conceive model-understandable instructions for the given tasks and carefully filter the LLM-generated data. In this work, we introduce an automatic instruction augmentation method named INSTRAUG in multimodal tasks. It starts from a handful of basic and straightforward meta instructions but can expand an instruction-following dataset by 30 times. Results on two popular multimodal instructionfollowing benchmarks MULTIINSTRUCT and InstructBLIP show that INSTRAUG can significantly improve the alignment of multimodal large language models (MLLMs) across 12 multimodal tasks, which is even equivalent to the benefits of scaling up training data multiple times.

Mainstream approaches to aligning large language models (LLMs) heavily rely on human preference data, particularly when models require periodic updates. The standard process for iterative alignment of LLMs involves collecting new human feedback for each update. However, the data collection process is costly and challenging to scale. To address this issue, we introduce the "TS-Align" framework, which fine-tunes a policy model using pairwise feedback data automatically mined from its outputs. This automatic mining process is efficiently accomplished through the collaboration between a large-scale teacher model and a small-scale student model. The policy fine-tuning process can be iteratively repeated using on-policy generations within our proposed teacher-student collaborative framework. Through extensive experiments, we demonstrate that our final aligned policy outperforms the base policy model with an average win rate of 69.7% across seven conversational or instruction-following datasets. Furthermore, we show that the ranking capability of the teacher is effectively distilled into the student through our pipeline, resulting in a small-scale yet effective reward model for policy model alignment.

Semantic Textual Similarity (STS) constitutes a critical research direction in computational linguistics and serves as a key indicator of the encoding capabilities of embedding models. Driven by advances in pre-trained language models and contrastive learning techniques, leading sentence representation methods can already achieved average Spearman's correlation scores of approximately 86 across seven STS benchmarks in SentEval. However, further improvements have become increasingly marginal, with no existing method attaining an average score higher than 87 on these tasks. This paper conducts an in-depth analysis of this phenomenon and concludes that the upper limit for Spearman's correlation scores using contrastive learning is 87.5. To transcend this ceiling, we propose an innovative approach termed Pcc-tuning, which employs Pearson's correlation coefficient as a loss function to refine model performance beyond contrastive learning. Experimental results demonstrate that Pcc-tuning markedly surpasses previous state-of-the-art strategies, raising the Spearman's correlation score to above 90.

Large models have demonstrated exceptional generalization capabilities in computer vision and natural language processing. Recent efforts have focused on enhancing these models with multimodal processing abilities. However, addressing the challenges posed by scenarios where one modality is absent remains a significant hurdle. In response to this issue, we propose a robust latent representation tuning method for large models. Specifically, our approach introduces a modality latent translation module to maximize the correlation between modalities, resulting in a robust representation. Following this, a newly designed fusion module is employed to facilitate information interaction between the modalities. Within this framework, common semantics are refined during training, and robust performance is achieved even in the absence of one modality. Importantly, our method maintains the frozen state of the image and text foundation models to preserve their capabilities acquired through large-scale pretraining. We conduct experiments on several public datasets, and the results underscore the effectiveness of our proposed method.

Diffusion policies have demonstrated robust performance in generative modeling, prompting their application in robotic manipulation controlled via language descriptions. In this paper, we introduce a zero-shot, open-vocabulary diffusion policy method for robot manipulation. Using Vision-Language Models (VLMs), our method transforms linguistic task descriptions into actionable keyframes in 3D space. These keyframes serve to guide the diffusion process via inpainting. However, naively enforcing the diffusion process to adhere to the generated keyframes is problematic: the keyframes from the VLMs may be incorrect and lead to out-of-distribution (OOD) action sequences where the diffusion model performs poorly. To address these challenges, we develop an inpainting optimization strategy that balances adherence to the keyframes v.s. the training data distribution. Experimental evaluations demonstrate that our approach surpasses the performance of traditional fine-tuned language-conditioned methods in both simulated and real-world settings.

Most public instruction finetuning datasets are relatively small compared to the closed source datasets used to train industry models. To study questions about finetuning at scale, such as curricula and learning rate cooldown schedules, there is a need for industrial-scale datasets. However, this scale necessitates a data generation process that is almost entirely automated. In this work, we study methods for generating large instruction datasets from a single prompt. With little human oversight, we get LLMs to write diverse sets of instruction examples ranging from simple completion tasks to complex multi-turn dialogs across a variety of subject areas. When finetuning a Llama-3 8B base model, our dataset meets or exceeds both WizardLM and Ultrachat on both knowledge-intensive leaderboard tasks as well as conversational evaluations. We release our dataset, the "generator" prompts that created it, and our finetuned model checkpoints.

Large Vision-Language Models (LVLMs), despite their recent success, are hardly comprehensively tested for their cognitive abilities. Inspired by the prevalent use of the "Cookie Theft" task in human cognition test, we propose a novel evaluation benchmark to evaluate high-level cognitive ability of LVLMs using images with rich semantics. It defines eight reasoning capabilities and consists of an image description task and a visual question answering task. Our evaluation on well-known LVLMs shows that there is still a large gap in cognitive ability between LVLMs and humans.

In recent years, the input context sizes of large language models (LLMs) have increased dramatically. However, existing evaluation methods have not kept pace, failing to comprehensively assess the efficiency of models in handling long contexts. To bridge this gap, we introduce the BABILong benchmark, designed to test language models' ability to reason across facts distributed in extremely long documents. BABILong includes a diverse set of 20 reasoning tasks, including fact chaining, simple induction, deduction, counting, and handling lists/sets. These tasks are challenging on their own, and even more demanding when the required facts are scattered across long natural text. Our evaluations show that popular LLMs effectively utilize only 10-20\% of the context and their performance declines sharply with increased reasoning complexity. Among alternatives to in-context reasoning, Retrieval-Augmented Generation methods achieve a modest 60\% accuracy on single-fact question answering, independent of context length. Among context extension methods, the highest performance is demonstrated by recurrent memory transformers, enabling the processing of lengths up to 11 million tokens. The BABILong benchmark is extendable to any length to support the evaluation of new upcoming models with increased capabilities, and we provide splits up to 1 million token lengths.

Causal video question answering (QA) has garnered increasing interest, yet existing datasets often lack depth in causal reasoning. To address this gap, we capitalize on the unique properties of cartoons and construct CausalChaos!, a novel, challenging causal Why-QA dataset built upon the iconic "Tom and Jerry" cartoon series. Cartoons use the principles of animation that allow animators to create expressive, unambiguous causal relationships between events to form a coherent storyline. Utilizing these properties, along with thought-provoking questions and multi-level answers (answer and detailed causal explanation), our questions involve causal chains that interconnect multiple dynamic interactions between characters and visual scenes. These factors demand models to solve more challenging, yet well-defined causal relationships. We also introduce hard incorrect answer mining, including a causally confusing version that is even more challenging. While models perform well, there is much room for improvement, especially, on open-ended answers. We identify more advanced/explicit causal relationship modeling & joint modeling of vision and language as the immediate areas for future efforts to focus upon. Along with the other complementary datasets, our new challenging dataset will pave the way for these developments in the field.

Memes, which rapidly disseminate personal opinions and positions across the internet, also pose significant challenges in propagating social bias and prejudice. This study presents a novel approach to detecting harmful memes, particularly within the multicultural and multilingual context of Singapore. Our methodology integrates image captioning, Optical Character Recognition (OCR), and Large Language Model (LLM) analysis to comprehensively understand and classify harmful memes. Utilizing the BLIP model for image captioning, PP-OCR and TrOCR for text recognition across multiple languages, and the Qwen LLM for nuanced language understanding, our system is capable of identifying harmful content in memes created in English, Chinese, Malay, and Tamil. To enhance the system's performance, we fine-tuned our approach by leveraging additional data labeled using GPT-4V, aiming to distill the understanding capability of GPT-4V for harmful memes to our system. Our framework achieves top-1 at the public leaderboard of the Online Safety Prize Challenge hosted by AI Singapore, with the AUROC as 0.7749 and accuracy as 0.7087, significantly ahead of the other teams. Notably, our approach outperforms previous benchmarks, with FLAVA achieving an AUROC of 0.5695 and VisualBERT an AUROC of 0.5561.