Image classification is one of the most fundamental capabilities of machine vision intelligence. In this work, we revisit the image classification task using visually-grounded language models (VLMs) such as GPT-4V and LLaVA. We find that existing proprietary and public VLMs, despite often using CLIP as a vision encoder and having many more parameters, significantly underperform CLIP on standard image classification benchmarks like ImageNet. To understand the reason, we explore several hypotheses concerning the inference algorithms, training objectives, and data processing in VLMs. Our analysis reveals that the primary cause is data-related: critical information for image classification is encoded in the VLM's latent space but can only be effectively decoded with enough training data. Specifically, there is a strong correlation between the frequency of class exposure during VLM training and instruction-tuning and the VLM's performance in those classes; when trained with sufficient data, VLMs can match the accuracy of state-of-the-art classification models. Based on these findings, we enhance a VLM by integrating classification-focused datasets into its training, and demonstrate that the enhanced classification performance of the VLM transfers to its general capabilities, resulting in an improvement of 11.8% on the newly collected ImageWikiQA dataset.

Leveraging the vast genetic diversity within microbiomes offers unparalleled insights into complex phenotypes, yet the task of accurately predicting and understanding such traits from genomic data remains challenging. We propose a framework taking advantage of existing large models for gene vectorization to predict habitat specificity from entire microbial genome sequences. Based on our model, we develop attribution techniques to elucidate gene interaction effects that drive microbial adaptation to diverse environments. We train and validate our approach on a large dataset of high quality microbiome genomes from different habitats. We not only demonstrate solid predictive performance, but also how sequence-level information of entire genomes allows us to identify gene associations underlying complex phenotypes. Our attribution recovers known important interaction networks and proposes new candidates for experimental follow up.

To parameter-efficiently fine-tune (PEFT) large language models (LLMs), the low-rank adaptation (LoRA) method approximates the model changes $\Delta W \in \mathbb{R}^{m \times n}$ through the product of two matrices $A \in \mathbb{R}^{m \times r}$ and $B \in \mathbb{R}^{r \times n}$, where $r \ll \min(m, n)$, $A$ is initialized with Gaussian noise, and $B$ with zeros. LoRA freezes the original model $W$ and updates the "Noise & Zero" adapter, which may lead to slow convergence. To overcome this limitation, we introduce Principal Singular values and Singular vectors Adaptation (PiSSA). PiSSA shares the same architecture as LoRA, but initializes the adaptor matrices $A$ and $B$ with the principal components of the original matrix $W$, and put the remaining components into a residual matrix $W^{res} \in \mathbb{R}^{m \times n}$ which is frozen during fine-tuning. Compared to LoRA, PiSSA updates the principal components while freezing the "residual" parts, allowing faster convergence and enhanced performance. Comparative experiments of PiSSA and LoRA across 12 different models, ranging from 184M to 70B, encompassing 5 NLG and 8 NLU tasks, reveal that PiSSA consistently outperforms LoRA under identical experimental setups. On the GSM8K benchmark, Mistral-7B fine-tuned with PiSSA achieves an accuracy of 72.86%, surpassing LoRA's 67.7% by 5.16%. Due to the same architecture, PiSSA is also compatible with quantization to further reduce the memory requirement of fine-tuning. Compared to QLoRA, QPiSSA (PiSSA with 4-bit quantization) exhibits smaller quantization errors in the initial stages. Fine-tuning LLaMA-3-70B on GSM8K, QPiSSA attains an accuracy of 86.05%, exceeding the performances of QLoRA at 81.73%. Leveraging a fast SVD technique, PiSSA can be initialized in only a few seconds, presenting a negligible cost for transitioning from LoRA to PiSSA.

Source-Free Domain Adaptation (SFDA) aims to adapt a pre-trained source model to a target domain using only unlabeled target data. Current SFDA methods face challenges in effectively leveraging pre-trained knowledge and exploiting target domain data. Multimodal Large Language Models (MLLMs) offer remarkable capabilities in understanding visual and textual information, but their applicability to SFDA poses challenges such as instruction-following failures, intensive computational demands, and difficulties in performance measurement prior to adaptation. To alleviate these issues, we propose Reliability-based Curriculum Learning (RCL), a novel framework that integrates multiple MLLMs for knowledge exploitation via pseudo-labeling in SFDA. Our framework incorporates proposed Reliable Knowledge Transfer, Self-correcting and MLLM-guided Knowledge Expansion, and Multi-hot Masking Refinement to progressively exploit unlabeled data in the target domain. RCL achieves state-of-the-art (SOTA) performance on multiple SFDA benchmarks, e.g., +9.4% on DomainNet, demonstrating its effectiveness in enhancing adaptability and robustness without requiring access to source data.

Large Language Models have gained considerable attention for their revolutionary capabilities. However, there is also growing concern on their safety implications, making a comprehensive safety evaluation for LLMs urgently needed before model deployment. In this work, we propose S-Eval, a new comprehensive, multi-dimensional and open-ended safety evaluation benchmark. At the core of S-Eval is a novel LLM-based automatic test prompt generation and selection framework, which trains an expert testing LLM Mt combined with a range of test selection strategies to automatically construct a high-quality test suite for the safety evaluation. The key to the automation of this process is a novel expert safety-critique LLM Mc able to quantify the riskiness score of an LLM's response, and additionally produce risk tags and explanations. Besides, the generation process is also guided by a carefully designed risk taxonomy with four different levels, covering comprehensive and multi-dimensional safety risks of concern. Based on these, we systematically construct a new and large-scale safety evaluation benchmark for LLMs consisting of 220,000 evaluation prompts, including 20,000 base risk prompts (10,000 in Chinese and 10,000 in English) and 200,000 corresponding attack prompts derived from 10 popular adversarial instruction attacks against LLMs. Moreover, considering the rapid evolution of LLMs and accompanied safety threats, S-Eval can be flexibly configured and adapted to include new risks, attacks and models. S-Eval is extensively evaluated on 20 popular and representative LLMs. The results confirm that S-Eval can better reflect and inform the safety risks of LLMs compared to existing benchmarks. We also explore the impacts of parameter scales, language environments, and decoding parameters on the evaluation, providing a systematic methodology for evaluating the safety of LLMs.

The need for fair AI is increasingly clear in the era of general-purpose systems such as ChatGPT, Gemini, and other large language models (LLMs). However, the increasing complexity of human-AI interaction and its social impacts have raised questions of how fairness standards could be applied. Here, we review the technical frameworks that machine learning researchers have used to evaluate fairness, such as group fairness and fair representations, and find that their application to LLMs faces inherent limitations. We show that each framework either does not logically extend to LLMs or presents a notion of fairness that is intractable for LLMs, primarily due to the multitudes of populations affected, sensitive attributes, and use cases. To address these challenges, we develop guidelines for the more realistic goal of achieving fairness in particular use cases: the criticality of context, the responsibility of LLM developers, and the need for stakeholder participation in an iterative process of design and evaluation. Moreover, it may eventually be possible and even necessary to use the general-purpose capabilities of AI systems to address fairness challenges as a form of scalable AI-assisted alignment.

Going beyond mimicking limited human experiences, recent studies show initial evidence that, like humans, large language models (LLMs) are capable of improving their abilities purely by self-correction, i.e., correcting previous responses through self-examination, in certain circumstances. Nevertheless, little is known about how such capabilities arise. In this work, based on a simplified setup akin to an alignment task, we theoretically analyze self-correction from an in-context learning perspective, showing that when LLMs give relatively accurate self-examinations as rewards, they are capable of refining responses in an in-context way. Notably, going beyond previous theories on over-simplified linear transformers, our theoretical construction underpins the roles of several key designs of realistic transformers for self-correction: softmax attention, multi-head attention, and the MLP block. We validate these findings extensively on synthetic datasets. Inspired by these findings, we also illustrate novel applications of self-correction, such as defending against LLM jailbreaks, where a simple self-correction step does make a large difference. We believe that these findings will inspire further research on understanding, exploiting, and enhancing self-correction for building better foundation models.

Elon Musk's artificial intelligence company xAI has closed a 6bn ( 4.7bn) investment round that will make it among the best-funded challengers to OpenAI. The startup is only a year old, but it has rapidly built its own large language model (LLM), the technology underpinning many of the recent advances in generative artificial intelligence capable of creating human-like text, pictures, video, and voices. The funding round, one of the biggest yet in the burgeoning AI field, values the company at 18bn before taking into account the 6bn investment, Musk said on X, the social network he owns. Generative AI has so far proven very expensive to develop, in part because of the need for huge amounts of computing power and energy to train LLMs. In a blogpost, xAI said: "The funds from the round will be used to take xAI's first products to market, build advanced infrastructure, and accelerate the research and development of future technologies."

"Oh, stop it--you're making me blush," the throaty voice said, laughing off a compliment. Barret Zoph, who'd given the compliment, looked pleased. As he should--Zoph represents OpenAI, the company behind the voice. "We are looking at the future of interaction between ourselves and the machines," promised Mira Murati, OpenAI's chief technology officer. ChatGPT-4o is just one of a wave of new conversational A.I., including the rollout of Meta AI last month.

When OpenAI's new voice assistant said it was "doing fantastic" in a launch demo this month, Scarlett Johansson was not. The Hollywood star said she was "shocked, angered and in disbelief" that the updated version of ChatGPT, which can listen to spoken prompts and respond verbally, had a voice "eerily similar" to hers. One of Johansson's signature roles was as the voice of a futuristic version of Siri in the 2013 film Her and, for the actor, the similarity was stark. The OpenAI chief executive, Sam Altman, appeared to acknowledge the film's influence with a one-word post on X on the day of the launch: "her". In a statement, Johansson said Altman had approached her last year to be a voice of ChatGPT and that she had declined for "personal reasons".