Generative large language models (LLMs) exhibit impressive capabilities, which can be further augmented by integrating a pre-trained vision model into the original LLM to create a multimodal LLM (MLLM). However, this integration often significantly decreases performance on natural language understanding and generation tasks, compared to the original LLM. This study investigates this issue using the LLaVA MLLM, treating the integration as a continual learning problem. We evaluate five continual learning methods to mitigate forgetting and identify a technique that enhances visual understanding while minimizing linguistic performance loss. Our approach reduces linguistic performance degradation by up to 15% over the LLaVA recipe, while maintaining high multimodal accuracy. We also demonstrate the robustness of our method through continual learning on a sequence of vision-language tasks, effectively preserving linguistic skills while acquiring new multimodal capabilities. Figure 1: Summary results of the best CL methods we evaluated for training LLaVA 1.5 compared to the unimodal base LLM and the original version of LLaVA 1.5. All results are with Pythia 2.8B as the base LLM. The best method has almost the same vision-language (VL) accuracy while providing a large increase in linguistic performance on 1 NLG and 4 NLU tasks by 8% and 2% (absolute), resp.

The rapid development of large language and multimodal models has sparked significant interest in using proprietary models, such as GPT-4o, to develop autonomous agents capable of handling real-world scenarios like web navigation. Although recent open-source efforts have tried to equip agents with the ability to explore environments and continuously improve over time, they are building text-only agents in synthetic environments where the reward signals are clearly defined. Such agents struggle to generalize to realistic settings that require multimodal perception abilities and lack ground-truth signals. In this paper, we introduce an open-source framework designed to facilitate the development of multimodal web agent that can autonomously conduct real-world exploration and improve itself. We first train the base model with imitation learning to gain the basic abilities. We then let the agent explore the open web and collect feedback on its trajectories. After that, it further improves its policy by learning from well-performing trajectories judged by another general-purpose model. This exploration-feedback-optimization cycle can continue for several iterations. Experimental results show that our web agent successfully improves itself after each iteration, demonstrating strong performance across multiple test sets.

Large language models (LLMs) have demonstrated their remarkable performance across various language understanding tasks. While emerging benchmarks have been proposed to evaluate LLMs in various domains such as mathematics and computer science, they merely measure the accuracy in terms of the final prediction on multi-choice questions. However, it remains insufficient to verify the essential understanding of LLMs given a chosen choice. To fill this gap, we present CLR-Bench to comprehensively evaluate the LLMs in complex college-level reasoning. Specifically, (i) we prioritize 16 challenging college disciplines in computer science and artificial intelligence. The dataset contains 5 types of questions, while each question is associated with detailed explanations from experts. (ii) To quantify a fair evaluation of LLMs' reasoning ability, we formalize the criteria with two novel metrics. Q$\rightarrow$A is utilized to measure the performance of direct answer prediction, and Q$\rightarrow$AR effectively considers the joint ability to answer the question and provide rationale simultaneously. Extensive experiments are conducted with 40 LLMs over 1,018 discipline-specific questions. The results demonstrate the key insights that LLMs, even the best closed-source LLM, i.e., GPT-4 turbo, tend to `guess' the college-level answers. It shows a dramatic decrease in accuracy from 63.31% Q$\rightarrow$A to 39.00% Q$\rightarrow$AR, indicating an unsatisfactory reasoning ability.

Despite its widespread adoption, Adam's advantage over Stochastic Gradient Descent (SGD) lacks a comprehensive theoretical explanation. This paper investigates Adam's sensitivity to rotations of the parameter space. We demonstrate that Adam's performance in training transformers degrades under random rotations of the parameter space, indicating a crucial sensitivity to the choice of basis. This reveals that conventional rotation-invariant assumptions are insufficient to capture Adam's advantages theoretically. To better understand the rotation-dependent properties that benefit Adam, we also identify structured rotations that preserve or even enhance its empirical performance. We then examine the rotation-dependent assumptions in the literature, evaluating their adequacy in explaining Adam's behavior across various rotation types. This work highlights the need for new, rotation-dependent theoretical frameworks to fully understand Adam's empirical success in modern machine learning tasks.

Despite the success of Large Language Models (LLMs), they still face challenges related to high inference costs and memory requirements. To address these issues, Knowledge Distillation (KD) has emerged as a popular method for model compression, with student-generated outputs (SGOs) being particularly notable for reducing the mismatch between training and inference. However, SGOs often produce noisy and biased sequences, which can lead to misguidance from the teacher model, especially in long sequences. To mitigate these challenges, we propose SWITCH (Studying WIth TeaCHer for Knowledge Distillation), a novel approach that strategically incorporates the teacher model during the student's sequence generation. SWITCH identifies discrepancies between the token probabilities of the teacher and student models, allowing the teacher to intervene selectively, particularly in long sequences that are more prone to teacher misguidance. Extensive experimental results across three model families and five instruction-following datasets show that SWITCH surpasses traditional KD methods, particularly excelling in the generation of long sequential data.

Hacking poses a significant threat to cybersecurity, inflicting billions of dollars in damages annually. To mitigate these risks, ethical hacking, or penetration testing, is employed to identify vulnerabilities in systems and networks. Recent advancements in large language models (LLMs) have shown potential across various domains, including cybersecurity. However, there is currently no comprehensive, open, end-to-end automated penetration testing benchmark to drive progress and evaluate the capabilities of these models in security contexts. This paper introduces a novel open benchmark for LLM-based automated penetration testing, addressing this critical gap. We first evaluate the performance of LLMs, including GPT-4o and Llama 3.1-405B, using the state-of-the-art PentestGPT tool. Our findings reveal that while Llama 3.1 demonstrates an edge over GPT-4o, both models currently fall short of performing fully automated, end-to-end penetration testing. Next, we advance the state-of-the-art and present ablation studies that provide insights into improving the PentestGPT tool. Our research illuminates the challenges LLMs face in each aspect of Pentesting, e.g. enumeration, exploitation, and privilege escalation. This work contributes to the growing body of knowledge on AI-assisted cybersecurity and lays the foundation for future research in automated penetration testing using large language models.

Navigating efficiently to an object in an unexplored environment is a critical skill for general-purpose intelligent robots. Recent approaches to this object goal navigation problem have embraced a modular strategy, integrating classical exploration algorithms-notably frontier exploration-with a learned semantic mapping/exploration module. This paper introduces a novel informative path planning and 3D object probability mapping approach. The mapping module computes the probability of the object of interest through semantic segmentation and a Bayes filter. Additionally, it stores probabilities for common objects, which semantically guides the exploration based on common sense priors from a large language model. The planner terminates when the current viewpoint captures enough voxels identified with high confidence as the object of interest. Although our planner follows a zero-shot approach, it achieves state-of-the-art performance as measured by the Success weighted by Path Length (SPL) and Soft SPL in the Habitat ObjectNav Challenge 2023, outperforming other works by more than 20%. Furthermore, we validate its effectiveness on real robots. Project webpage: https://ippon-paper.github.io/

Retrieval-augmented generation (RAG) is a key means to effectively enhance large language models (LLMs) in many knowledge-based tasks. However, existing RAG methods struggle with knowledge-intensive reasoning tasks, because useful information required to these tasks are badly scattered. This characteristic makes it difficult for existing RAG methods to accurately identify key information and perform global reasoning with such noisy augmentation. In this paper, motivated by the cognitive theories that humans convert raw information into various structured knowledge when tackling knowledge-intensive reasoning, we proposes a new framework, StructRAG, which can identify the optimal structure type for the task at hand, reconstruct original documents into this structured format, and infer answers based on the resulting structure. Extensive experiments across various knowledge-intensive tasks show that StructRAG achieves state-of-the-art performance, particularly excelling in challenging scenarios, demonstrating its potential as an effective solution for enhancing LLMs in complex real-world applications.

Large language models (LLMs) are susceptible to memorizing training data, raising concerns due to the potential extraction of sensitive information. Current methods to measure memorization rates of LLMs, primarily discoverable extraction (Carlini et al., 2022), rely on single-sequence greedy sampling, potentially underestimating the true extent of memorization. This paper introduces a probabilistic relaxation of discoverable extraction that quantifies the probability of extracting a target sequence within a set of generated samples, considering various sampling schemes and multiple attempts. This approach addresses the limitations of reporting memorization rates through discoverable extraction by accounting for the probabilistic nature of LLMs and user interaction patterns. Our experiments demonstrate that this probabilistic measure can reveal cases of higher memorization rates compared to rates found through discoverable extraction. We further investigate the impact of different sampling schemes on extractability, providing a more comprehensive and realistic assessment of LLM memorization and its associated risks. Our contributions include a new probabilistic memorization definition, empirical evidence of its effectiveness, and a thorough evaluation across different models, sizes, sampling schemes, and training data repetitions.

The leakage of benchmark data into the training data has emerged as a significant challenge for evaluating the capabilities of large language models (LLMs). In this work, we use experimental evidence and theoretical estimates to challenge the common assumption that small-scale contamination renders benchmark evaluations invalid. First, we experimentally quantify the magnitude of benchmark overfitting based on scaling along three dimensions: The number of model parameters (up to 1.6B), the number of times an example is seen (up to 144), and the number of training tokens (up to 40B). We find that if model and data follow the Chinchilla scaling laws, minor contamination indeed leads to overfitting. At the same time, even 144 times of contamination can be forgotten if the training data is scaled beyond five times Chinchilla, a regime characteristic of many modern LLMs. We then derive a simple theory of example forgetting via cumulative weight decay. It allows us to bound the number of gradient steps required to forget past data for any training run where we know the hyperparameters of AdamW. This indicates that many LLMs, including Llama 3, have forgotten the data seen at the beginning of training. Experimentally, we demonstrate that forgetting occurs faster than what is predicted by our bounds. Taken together, our results suggest that moderate amounts of contamination can be forgotten at the end of realistically scaled training runs.