AI assistants such as ChatGPT are trained to respond to users by saying, "I am a large language model". This raises questions. Do such models know that they are LLMs and reliably act on this knowledge? Are they aware of their current circumstances, such as being deployed to the public? We refer to a model's knowledge of itself and its circumstances as situational awareness. To quantify situational awareness in LLMs, we introduce a range of behavioral tests, based on question answering and instruction following. These tests form the $\textbf{Situational Awareness Dataset (SAD)}$, a benchmark comprising 7 task categories and over 13,000 questions. The benchmark tests numerous abilities, including the capacity of LLMs to (i) recognize their own generated text, (ii) predict their own behavior, (iii) determine whether a prompt is from internal evaluation or real-world deployment, and (iv) follow instructions that depend on self-knowledge. We evaluate 16 LLMs on SAD, including both base (pretrained) and chat models. While all models perform better than chance, even the highest-scoring model (Claude 3 Opus) is far from a human baseline on certain tasks. We also observe that performance on SAD is only partially predicted by metrics of general knowledge (e.g. MMLU). Chat models, which are finetuned to serve as AI assistants, outperform their corresponding base models on SAD but not on general knowledge tasks. The purpose of SAD is to facilitate scientific understanding of situational awareness in LLMs by breaking it down into quantitative abilities. Situational awareness is important because it enhances a model's capacity for autonomous planning and action. While this has potential benefits for automation, it also introduces novel risks related to AI safety and control. Code and latest results available at https://situational-awareness-dataset.org .
Handling distribution shifts from training data, known as out-of-distribution (OOD) generalization, poses a significant challenge in the field of machine learning. While a pre-trained vision-language model like CLIP has demonstrated remarkable zero-shot performance, further adaptation of the model to downstream tasks leads to undesirable degradation for OOD data. In this work, we introduce Sparse Adaptation for Fine-Tuning (SAFT), a method that prevents fine-tuning from forgetting the general knowledge in the pre-trained model. SAFT only updates a small subset of important parameters whose gradient magnitude is large, while keeping the other parameters frozen. SAFT is straightforward to implement and conceptually simple. Extensive experiments show that with only 0.1% of the model parameters, SAFT can significantly improve the performance of CLIP. It consistently outperforms baseline methods across several benchmarks. On the few-shot learning benchmark of ImageNet and its variants, SAFT gives a gain of 5.15% on average over the conventional fine-tuning method in OOD settings.
While learning to align Large Language Models (LLMs) with human preferences has shown remarkable success, aligning these models to meet the diverse user preferences presents further challenges in preserving previous knowledge. This paper examines the impact of personalized preference optimization on LLMs, revealing that the extent of knowledge loss varies significantly with preference heterogeneity. Although previous approaches have utilized the KL constraint between the reference model and the policy model, we observe that they fail to maintain general knowledge and alignment when facing personalized preferences. To this end, we introduce Base-Anchored Preference Optimization (BAPO), a simple yet effective approach that utilizes the initial responses of reference model to mitigate forgetting while accommodating personalized alignment. BAPO effectively adapts to diverse user preferences while minimally affecting global knowledge or general alignment. Our experiments demonstrate the efficacy of BAPO in various setups.
Large Language Models (LLMs) trained on extensive corpora inevitably retain sensitive data, such as personal privacy information and copyrighted material. Recent advancements in knowledge unlearning involve updating LLM parameters to erase specific knowledge. However, current unlearning paradigms are mired in vague forgetting boundaries, often erasing knowledge indiscriminately. In this work, we introduce KnowUnDo, a benchmark containing copyrighted content and user privacy domains to evaluate if the unlearning process inadvertently erases essential knowledge. Our findings indicate that existing unlearning methods often suffer from excessive unlearning. To address this, we propose a simple yet effective method, MemFlex, which utilizes gradient information to precisely target and unlearn sensitive parameters. Experimental results show that MemFlex is superior to existing methods in both precise knowledge unlearning and general knowledge retaining of LLMs. Code and dataset are released at https://github.com/zjunlp/KnowUnDo.
To address data heterogeneity, the key strategy of Personalized Federated Learning (PFL) is to decouple general knowledge (shared among clients) and client-specific knowledge, as the latter can have a negative impact on collaboration if not removed. Existing PFL methods primarily adopt a parameter partitioning approach, where the parameters of a model are designated as one of two types: parameters shared with other clients to extract general knowledge and parameters retained locally to learn client-specific knowledge. However, as these two types of parameters are put together like a jigsaw puzzle into a single model during the training process, each parameter may simultaneously absorb both general and client-specific knowledge, thus struggling to separate the two types of knowledge effectively. In this paper, we introduce FedDecomp, a simple but effective PFL paradigm that employs parameter additive decomposition to address this issue. Instead of assigning each parameter of a model as either a shared or personalized one, FedDecomp decomposes each parameter into the sum of two parameters: a shared one and a personalized one, thus achieving a more thorough decoupling of shared and personalized knowledge compared to the parameter partitioning method. In addition, as we find that retaining local knowledge of specific clients requires much lower model capacity compared with general knowledge across all clients, we let the matrix containing personalized parameters be low rank during the training process. Moreover, a new alternating training strategy is proposed to further improve the performance. Experimental results across multiple datasets and varying degrees of data heterogeneity demonstrate that FedDecomp outperforms state-of-the-art methods up to 4.9\%. The code is available at https://github.com/XinghaoWu/FedDecomp.
We study the Knowledge-Based visual question-answering problem, for which given a question, the models need to ground it into the visual modality to find the answer. Although many recent works use question-dependent captioners to verbalize the given image and use Large Language Models to solve the VQA problem, the research results show they are not reasonably performing for multi-hop questions. Our study shows that replacing a complex question with several simpler questions helps to extract more relevant information from the image and provide a stronger comprehension of it. Moreover, we analyze the decomposed questions to find out the modality of the information that is required to answer them and use a captioner for the visual questions and LLMs as a general knowledge source for the non-visual KB-based questions. Our results demonstrate the positive impact of using simple questions before retrieving visual or non-visual information. We have provided results and analysis on three well-known VQA datasets including OKVQA, A-OKVQA, and KRVQA, and achieved up to 2% improvement in accuracy.
Recently, large language models (LLMs) with extensive general knowledge and powerful reasoning abilities have seen rapid development and widespread application. A systematic and reliable evaluation of LLMs or vision-language model (VLMs) is a crucial step in applying and developing them for various fields. There have been some early explorations about the usability of LLMs for limited urban tasks, but a systematic and scalable evaluation benchmark is still lacking. The challenge in constructing a systematic evaluation benchmark for urban research lies in the diversity of urban data, the complexity of application scenarios and the highly dynamic nature of the urban environment. In this paper, we design CityBench, an interactive simulator based evaluation platform, as the first systematic benchmark for evaluating the capabilities of LLMs for diverse tasks in urban research. First, we build CityData to integrate the diverse urban data and CitySimu to simulate fine-grained urban dynamics. Based on CityData and CitySimu, we design 8 representative urban tasks in 2 categories of perception-understanding and decision-making as the CityBench. With extensive results from 30 well-known LLMs and VLMs in 13 cities around the world, we find that advanced LLMs and VLMs can achieve competitive performance in diverse urban tasks requiring commonsense and semantic understanding abilities, e.g., understanding the human dynamics and semantic inference of urban images. Meanwhile, they fail to solve the challenging urban tasks requiring professional knowledge and high-level reasoning abilities, e.g., geospatial prediction and traffic control task. These observations provide valuable perspectives for utilizing and developing LLMs in the future. Codes are openly accessible via https://github.com/tsinghua-fib-lab/CityBench.
As large language models (LLMs) become increasingly integrated into daily applications, it is essential to ensure they operate fairly across diverse user demographics. In this work, we show that LLMs suffer from personalization bias, where their performance is impacted when they are personalized to a user's identity. We quantify personalization bias by evaluating the performance of LLMs along two axes - safety and utility. We measure safety by examining how benign LLM responses are to unsafe prompts with and without personalization. We measure utility by evaluating the LLM's performance on various tasks, including general knowledge, mathematical abilities, programming, and reasoning skills. We find that various LLMs, ranging from open-source models like Llama (Touvron et al., 2023) and Mistral (Jiang et al., 2023) to API-based ones like GPT-3.5 and GPT-4o (Ouyang et al., 2022), exhibit significant variance in performance in terms of safety-utility trade-offs depending on the user's identity. Finally, we discuss several strategies to mitigate personalization bias using preference tuning and prompt-based defenses.
Pre-trained Large Language Models (LLMs) have demonstrated remarkable capabilities but also pose risks by learning and generating copyrighted material, leading to significant legal and ethical concerns. In real-world scenarios, model owners need to continuously address copyright infringement as new requests for content removal emerge at different time points. This leads to the need for sequential unlearning, where copyrighted content is removed sequentially as new requests arise. Despite its practical relevance, sequential unlearning in the context of copyright infringement has not been rigorously explored in existing literature. To address this gap, we propose Stable Sequential Unlearning (SSU), a novel framework designed to unlearn copyrighted content from LLMs over multiple time steps. Our approach works by identifying and removing specific weight updates in the model's parameters that correspond to copyrighted content. We improve unlearning efficacy by introducing random labeling loss and ensuring the model retains its general-purpose knowledge by adjusting targeted parameters. Experimental results show that SSU achieves an effective trade-off between unlearning efficacy and general-purpose language abilities, outperforming existing baselines.
Large language models (LLMs) demonstrate an impressive ability to internalize knowledge and answer natural language questions. Although previous studies validate that LLMs perform well on general knowledge while presenting poor performance on long-tail nuanced knowledge, the community is still doubtful about whether the traditional knowledge graphs should be replaced by LLMs. In this paper, we ask if the schema of knowledge graph (i.e., taxonomy) is made obsolete by LLMs. Intuitively, LLMs should perform well on common taxonomies and at taxonomy levels that are common to people. Unfortunately, there lacks a comprehensive benchmark that evaluates the LLMs over a wide range of taxonomies from common to specialized domains and at levels from root to leaf so that we can draw a confident conclusion. To narrow the research gap, we constructed a novel taxonomy hierarchical structure discovery benchmark named TaxoGlimpse to evaluate the performance of LLMs over taxonomies. TaxoGlimpse covers ten representative taxonomies from common to specialized domains with in-depth experiments of different levels of entities in this taxonomy from root to leaf. Our comprehensive experiments of eighteen state-of-the-art LLMs under three prompting settings validate that LLMs can still not well capture the knowledge of specialized taxonomies and leaf-level entities.
Large Language Models (LLMs) are trained on vast amounts of data, most of which is automatically scraped from the internet. This data includes encyclopedic documents that harbor a vast amount of general knowledge (e.g., Wikipedia) but also potentially overlap with benchmark datasets used for evaluating LLMs. Consequently, evaluating models on test splits that might have leaked into the training set is prone to misleading conclusions. To foster sound evaluation of language models, we introduce a new test dataset named RepLiQA, suited for question-answering and topic retrieval tasks. RepLiQA is a collection of five splits of test sets, four of which have not been released to the internet or exposed to LLM APIs prior to this publication. Each sample in RepLiQA comprises (1) a reference document crafted by a human annotator and depicting an imaginary scenario (e.g., a news article) absent from the internet; (2) a question about the document's topic; (3) a ground-truth answer derived directly from the information in the document; and (4) the paragraph extracted from the reference document containing the answer. As such, accurate answers can only be generated if a model can find relevant content within the provided document. We run a large-scale benchmark comprising several state-of-the-art LLMs to uncover differences in performance across models of various types and sizes in a context-conditional language modeling setting. Released splits of RepLiQA can be found here: https://huggingface.co/datasets/ServiceNow/repliqa.
Generative models have demonstrated human-level proficiency in various benchmarks across domains like programming, natural sciences, and general knowledge. Despite these promising results on competitive benchmarks, they still struggle with seemingly simple problem-solving tasks typically carried out by elementary-level students. How do state-of-the-art models perform on standardized tests designed to assess computational thinking and problem-solving skills at schools? In this paper, we curate a novel benchmark involving computational thinking tests grounded in elementary visual programming domains. Our initial results show that state-of-the-art models like GPT-4o and Llama3 barely match the performance of an average school student. To further boost the performance of these models, we fine-tune them using a novel synthetic data generation methodology. The key idea is to develop a comprehensive dataset using symbolic methods that capture different skill levels, ranging from recognition of visual elements to multi-choice quizzes to synthesis-style tasks. We showcase how various aspects of symbolic information in synthetic data help improve fine-tuned models' performance. We will release the full implementation and datasets to facilitate further research on enhancing computational thinking in generative models.
Entity matching is the task of linking records from different sources that refer to the same real-world entity. Past work has primarily treated entity linking as a standard supervised learning problem. However, supervised entity matching models often do not generalize well to new data, and collecting exhaustive labeled training data is often cost prohibitive. Further, recent efforts have adopted LLMs for this task in few/zero-shot settings, exploiting their general knowledge. But LLMs are prohibitively expensive for performing inference at scale for real-world entity matching tasks. As an efficient alternative, we re-cast entity matching as a conditional generation task as opposed to binary classification. This enables us to "distill" LLM reasoning into smaller entity matching models via natural language explanations. This approach achieves strong performance, especially on out-of-domain generalization tests (10.85% F-1) where standalone generative methods struggle. We perform ablations that highlight the importance of explanations, both for performance and model robustness.
Robustness against noisy imaging is crucial for practical image anomaly detection systems. This study introduces a Robust Anomaly Detection (RAD) dataset with free views, uneven illuminations, and blurry collections to systematically evaluate the robustness of current anomaly detection methods. Specifically, RAD aims to identify foreign objects on working platforms as anomalies. The collection process incorporates various sources of imaging noise, such as viewpoint changes, uneven illuminations, and blurry collections, to replicate real-world inspection scenarios. Subsequently, we assess and analyze 11 state-of-the-art unsupervised and zero-shot methods on RAD. Our findings indicate that: 1) Variations in viewpoint, illumination, and blurring affect anomaly detection methods to varying degrees; 2) Methods relying on memory banks and assisted by synthetic anomalies demonstrate stronger robustness; 3) Effectively leveraging the general knowledge of foundational models is a promising avenue for enhancing the robustness of anomaly detection methods. The dataset is available at https://github.com/hustCYQ/RAD-dataset.
Retrieval-Augmented Generation (RAG) offers a promising solution to address various limitations of Large Language Models (LLMs), such as hallucination and difficulties in keeping up with real-time updates. This approach is particularly critical in expert and domain-specific applications where LLMs struggle to cover expert knowledge. Therefore, evaluating RAG models in such scenarios is crucial, yet current studies often rely on general knowledge sources like Wikipedia to assess the models' abilities in solving common-sense problems. In this paper, we evaluated LLMs by RAG settings in a domain-specific context, college enrollment. We identified six required abilities for RAG models, including the ability in conversational RAG, analyzing structural information, faithfulness to external knowledge, denoising, solving time-sensitive problems, and understanding multi-document interactions. Each ability has an associated dataset with shared corpora to evaluate the RAG models' performance. We evaluated popular LLMs such as Llama, Baichuan, ChatGLM, and GPT models. Experimental results indicate that existing closed-book LLMs struggle with domain-specific questions, highlighting the need for RAG models to solve expert problems. Moreover, there is room for RAG models to improve their abilities in comprehending conversational history, analyzing structural information, denoising, processing multi-document interactions, and faithfulness in expert knowledge. We expect future studies could solve these problems better.
Endowing Large Multimodal Models (LMMs) with visual grounding capability can significantly enhance AIs' understanding of the visual world and their interaction with humans. However, existing methods typically fine-tune the parameters of LMMs to learn additional segmentation tokens and overfit grounding and segmentation datasets. Such a design would inevitably cause a catastrophic diminution in the indispensable conversational capability of general AI assistants. In this paper, we comprehensively evaluate state-of-the-art grounding LMMs across a suite of multimodal question-answering benchmarks, observing pronounced performance drops that indicate vanishing general knowledge comprehension and weakened instruction following ability. To address this issue, we present F-LMM -- grounding frozen off-the-shelf LMMs in human-AI conversations -- a straightforward yet effective design based on the fact that word-pixel correspondences conducive to visual grounding inherently exist in the attention weights of well-trained LMMs. Using only a few trainable CNN layers, we can translate word-pixel attention weights to mask logits, which a SAM-based mask refiner can further optimise. Our F-LMM neither learns special segmentation tokens nor utilises high-quality grounded instruction-tuning data, but achieves competitive performance on referring expression segmentation and panoptic narrative grounding benchmarks while completely preserving LMMs' original conversational ability. Additionally, with instruction-following ability preserved and grounding ability obtained, our F-LMM can perform visual chain-of-thought reasoning and better resist object hallucinations.
Recommender systems are widely used to suggest engaging content, and Large Language Models (LLMs) have given rise to generative recommenders. Such systems can directly generate items, including for open-set tasks like question suggestion. While the world knowledge of LLMs enable good recommendations, improving the generated content through user feedback is challenging as continuously fine-tuning LLMs is prohibitively expensive. We present a training-free approach for optimizing generative recommenders by connecting user feedback loops to LLM-based optimizers. We propose a generative explore-exploit method that can not only exploit generated items with known high engagement, but also actively explore and discover hidden population preferences to improve recommendation quality. We evaluate our approach on question generation in two domains (e-commerce and general knowledge), and model user feedback with Click Through Rate (CTR). Experiments show our LLM-based explore-exploit approach can iteratively improve recommendations, and consistently increase CTR. Ablation analysis shows that generative exploration is key to learning user preferences, avoiding the pitfalls of greedy exploit-only approaches. A human evaluation strongly supports our quantitative findings.
This study investigates the loss of generalization ability in neural networks, revisiting warm-starting experiments from Ash & Adams. Our empirical analysis reveals that common methods designed to enhance plasticity by maintaining trainability provide limited benefits to generalization. While reinitializing the network can be effective, it also risks losing valuable prior knowledge. To this end, we introduce the Hare & Tortoise, inspired by the brain's complementary learning system. Hare & Tortoise consists of two components: the Hare network, which rapidly adapts to new information analogously to the hippocampus, and the Tortoise network, which gradually integrates knowledge akin to the neocortex. By periodically reinitializing the Hare network to the Tortoise's weights, our method preserves plasticity while retaining general knowledge. Hare & Tortoise can effectively maintain the network's ability to generalize, which improves advanced reinforcement learning algorithms on the Atari-100k benchmark. The code is available at https://github.com/dojeon-ai/hare-tortoise.
Personalization has emerged as a critical research area in modern intelligent systems, focusing on mining users' behavioral history and adapting to their preferences for delivering tailored experiences. Despite the remarkable few-shot capabilities exhibited by black-box large language models (LLMs), the inherent opacity of their model parameters presents significant challenges in aligning the generated output with individual expectations. Existing solutions have primarily focused on prompt design to incorporate user-specific profiles and behaviors; however, such approaches often struggle to generalize effectively due to their inability to capture shared knowledge among all users. To address these challenges, we propose HYDRA, a model factorization framework that captures both user-specific behavior patterns from historical data and shared general knowledge among all users to deliver personalized generation. In order to capture user-specific behavior patterns, we first train a reranker to prioritize the most useful information from top-retrieved relevant historical records. By combining the prioritized history with the corresponding query, we train an adapter to align the output with individual user-specific preferences, eliminating the reliance on access to inherent model parameters of black-box LLMs. Both the reranker and the adapter can be decomposed into a base model with multiple user-specific heads, resembling a hydra. The base model maintains shared knowledge across users, while the multiple personal heads capture user-specific preferences. Experimental results demonstrate that HYDRA outperforms existing state-of-the-art prompt-based methods by an average relative improvement of 9.01% across five diverse personalization tasks in the LaMP benchmark. Our implementation is available at https://github.com/night-chen/HYDRA.
Normalizing flow-based generative models have been widely used in applications where the exact density estimation is of major importance. Recent research proposes numerous methods to improve their expressivity. However, conditioning on a context is largely overlooked area in the bijective flow research. Conventional conditioning with the vector concatenation is limited to only a few flow types. More importantly, this approach cannot support a practical setup where a set of context-conditioned (specialist) models are trained with the fixed pretrained general-knowledge (generalist) model. We propose ContextFlow++ approach to overcome these limitations using an additive conditioning with explicit generalist-specialist knowledge decoupling. Furthermore, we support discrete contexts by the proposed mixed-variable architecture with context encoders. Particularly, our context encoder for discrete variables is a surjective flow from which the context-conditioned continuous variables are sampled. Our experiments on rotated MNIST-R, corrupted CIFAR-10C, real-world ATM predictive maintenance and SMAP unsupervised anomaly detection benchmarks show that the proposed ContextFlow++ offers faster stable training and achieves higher performance metrics. Our code is publicly available at https://github.com/gudovskiy/contextflow.