As world knowledge advances and new task schemas emerge, Continual Learning (CL) becomes essential for keeping Large Language Models (LLMs) current and addressing their shortcomings. This process typically involves continual instruction tuning (CIT) and continual pre-training (CPT) to enable these models to adapt to novel tasks and acquire critical knowledge. However, collecting sufficient CPT data and efficiently bridging knowledge gaps remain significant challenges. Inspired by the 'summarizing mistakes' strategy, we propose the Continue Evolving from Mistakes (CEM) method, a data-efficient approach aiming to collect CPT data and continually improve LLMs' performance through iterative evaluation and supplementation with mistake-relevant knowledge. To further optimize data usage and mitigate forgetting, we introduce a novel training paradigm that combines CIT and CPT. Experiments show that CEM substantially enhances multiple models' performance on both in-domain and out-of-domain QA tasks, achieving gains of up to 29.63%. Code and datasets are available on https://anonymous.4open.science/r/cem-BB25.
Multi-modal knowledge graph completion (MMKGC) aims to discover unobserved knowledge from given knowledge graphs, collaboratively leveraging structural information from the triples and multi-modal information of the entities to overcome the inherent incompleteness. Existing MMKGC methods usually extract multi-modal features with pre-trained models, resulting in coarse handling of multi-modal entity information, overlooking the nuanced, fine-grained semantic details and their complex interactions. To tackle this shortfall, we introduce a novel framework MyGO to tokenize, fuse, and augment the fine-grained multi-modal representations of entities and enhance the MMKGC performance. Motivated by the tokenization technology, MyGO tokenizes multi-modal entity information as fine-grained discrete tokens and learns entity representations with a cross-modal entity encoder. To further augment the multi-modal representations, MyGO incorporates fine-grained contrastive learning to highlight the specificity of the entity representations. Experiments on standard MMKGC benchmarks reveal that our method surpasses 19 of the latest models, underlining its superior performance. Code and data can be found in https://github.com/zjukg/MyGO
Human intention-based systems enable robots to perceive and interpret user actions to interact with humans and adapt to their behavior proactively. Therefore, intention prediction is pivotal in creating a natural interaction with social robots in human-designed environments. In this paper, we examine using Large Language Models (LLMs) to infer human intention in a collaborative object categorization task with a physical robot. We propose a novel multimodal approach that integrates user non-verbal cues, like hand gestures, body poses, and facial expressions, with environment states and user verbal cues to predict user intentions in a hierarchical architecture. Our evaluation of five LLMs shows the potential for reasoning about verbal and non-verbal user cues, leveraging their context-understanding and real-world knowledge to support intention prediction while collaborating on a task with a social robot. Video: https://youtu.be/tBJHfAuzohI
The task of reading comprehension (RC), often implemented as context-based question answering (QA), provides a primary means to assess language models' natural language understanding (NLU) capabilities. Yet, when applied to large language models (LLMs) with extensive built-in world knowledge, this method can be deceptive. If the context aligns with the LLMs' internal knowledge, it is hard to discern whether the models' answers stem from context comprehension or from LLMs' internal information. Conversely, using data that conflicts with the models' knowledge creates erroneous trends which distort the results. To address this issue, we suggest to use RC on imaginary data, based on fictitious facts and entities. This task is entirely independent of the models' world knowledge, enabling us to evaluate LLMs' linguistic abilities without the interference of parametric knowledge. Testing ChatGPT, GPT-4, LLaMA 2 and Mixtral on such imaginary data, we uncover a class of linguistic phenomena posing a challenge to current LLMs, involving thinking in terms of alternative, hypothetical scenarios. While all the models handle simple affirmative and negative contexts with high accuracy, they are much more prone to error when dealing with modal and conditional contexts. Crucially, these phenomena also trigger the LLMs' vulnerability to knowledge-conflicts again. In particular, while some models prove virtually unaffected by knowledge conflicts in affirmative and negative contexts, when faced with more semantically involved modal and conditional environments, they often fail to separate the text from their internal knowledge.
Open-vocabulary human-object interaction (HOI) detection, which is concerned with the problem of detecting novel HOIs guided by natural language, is crucial for understanding human-centric scenes. However, prior zero-shot HOI detectors often employ the same levels of feature maps to model HOIs with varying distances, leading to suboptimal performance in scenes containing human-object pairs with a wide range of distances. In addition, these detectors primarily rely on category names and overlook the rich contextual information that language can provide, which is essential for capturing open vocabulary concepts that are typically rare and not well-represented by category names alone. In this paper, we introduce a novel end-to-end open vocabulary HOI detection framework with conditional multi-level decoding and fine-grained semantic enhancement (CMD-SE), harnessing the potential of Visual-Language Models (VLMs). Specifically, we propose to model human-object pairs with different distances with different levels of feature maps by incorporating a soft constraint during the bipartite matching process. Furthermore, by leveraging large language models (LLMs) such as GPT models, we exploit their extensive world knowledge to generate descriptions of human body part states for various interactions. Then we integrate the generalizable and fine-grained semantics of human body parts to improve interaction recognition. Experimental results on two datasets, SWIG-HOI and HICO-DET, demonstrate that our proposed method achieves state-of-the-art results in open vocabulary HOI detection. The code and models are available at https://github.com/ltttpku/CMD-SE-release.
While many have shown how Large Language Models (LLMs) can be applied to a diverse set of tasks, the critical issues of data contamination and memorization are often glossed over. In this work, we address this concern for tabular data. Specifically, we introduce a variety of different techniques to assess whether a language model has seen a tabular dataset during training. This investigation reveals that LLMs have memorized many popular tabular datasets verbatim. We then compare the few-shot learning performance of LLMs on datasets that were seen during training to the performance on datasets released after training. We find that LLMs perform better on datasets seen during training, indicating that memorization leads to overfitting. At the same time, LLMs show non-trivial performance on novel datasets and are surprisingly robust to data transformations. We then investigate the in-context statistical learning abilities of LLMs. While LLMs are significantly better than random at solving statistical classification problems, the sample efficiency of few-shot learning lags behind traditional statistical learning algorithms, especially as the dimension of the problem increases. This suggests that much of the observed few-shot performance on novel real-world datasets is due to the LLM's world knowledge. Overall, our results highlight the importance of testing whether an LLM has seen an evaluation dataset during pre-training. We release the https://github.com/interpretml/LLM-Tabular-Memorization-Checker Python package to test LLMs for memorization of tabular datasets.
Generative models have shown a giant leap in synthesizing photo-realistic images with minimal expertise, sparking concerns about the authenticity of online information. This study aims to develop a universal AI-generated image detector capable of identifying images from diverse sources. Existing methods struggle to generalize across unseen generative models when provided with limited sample sources. Inspired by the zero-shot transferability of pre-trained vision-language models, we seek to harness the nontrivial visual-world knowledge and descriptive proficiency of CLIP-ViT to generalize over unknown domains. This paper presents a novel parameter-efficient fine-tuning approach, mixture of low-rank experts, to fully exploit CLIP-ViT's potential while preserving knowledge and expanding capacity for transferable detection. We adapt only the MLP layers of deeper ViT blocks via an integration of shared and separate LoRAs within an MoE-based structure. Extensive experiments on public benchmarks show that our method achieves superiority over state-of-the-art approaches in cross-generator generalization and robustness to perturbations. Remarkably, our best-performing ViT-L/14 variant requires training only 0.08% of its parameters to surpass the leading baseline by +3.64% mAP and +12.72% avg.Acc across unseen diffusion and autoregressive models. This even outperforms the baseline with just 0.28% of the training data. Our code and pre-trained models will be available at https://github.com/zhliuworks/CLIPMoLE.
Sentences containing multiple semantic operators with overlapping scope often create ambiguities in interpretation, known as scope ambiguities. These ambiguities offer rich insights into the interaction between semantic structure and world knowledge in language processing. Despite this, there has been little research into how modern large language models treat them. In this paper, we investigate how different versions of certain autoregressive language models -- GPT-2, GPT-3/3.5, Llama 2 and GPT-4 -- treat scope ambiguous sentences, and compare this with human judgments. We introduce novel datasets that contain a joint total of almost 1,000 unique scope-ambiguous sentences, containing interactions between a range of semantic operators, and annotated for human judgments. Using these datasets, we find evidence that several models (i) are sensitive to the meaning ambiguity in these sentences, in a way that patterns well with human judgments, and (ii) can successfully identify human-preferred readings at a high level of accuracy (over 90% in some cases).
Knowledge probing assesses to which degree a language model (LM) has successfully learned relational knowledge during pre-training. Probing is an inexpensive way to compare LMs of different sizes and training configurations. However, previous approaches rely on the objective function used in pre-training LMs and are thus applicable only to masked or causal LMs. As a result, comparing different types of LMs becomes impossible. To address this, we propose an approach that uses an LM's inherent ability to estimate the log-likelihood of any given textual statement. We carefully design an evaluation dataset of 7,731 instances (40,916 in a larger variant) from which we produce alternative statements for each relational fact, one of which is correct. We then evaluate whether an LM correctly assigns the highest log-likelihood to the correct statement. Our experimental evaluation of 22 common LMs shows that our proposed framework, BEAR, can effectively probe for knowledge across different LM types. We release the BEAR datasets and an open-source framework that implements the probing approach to the research community to facilitate the evaluation and development of LMs.
This paper introduces PRobELM (Plausibility Ranking Evaluation for Language Models), a benchmark designed to assess language models' ability to discern more plausible from less plausible scenarios through their parametric knowledge. While benchmarks such as TruthfulQA emphasise factual accuracy or truthfulness, and others such as COPA explore plausible scenarios without explicitly incorporating world knowledge, PRobELM seeks to bridge this gap by evaluating models' capabilities to prioritise plausible scenarios that leverage world knowledge over less plausible alternatives. This design allows us to assess the potential of language models for downstream use cases such as literature-based discovery where the focus is on identifying information that is likely but not yet known. Our benchmark is constructed from a dataset curated from Wikidata edit histories, tailored to align the temporal bounds of the training data for the evaluated models. PRobELM facilitates the evaluation of language models across multiple prompting types, including statement, text completion, and question-answering. Experiments with 10 models of various sizes and architectures on the relationship between model scales, training recency, and plausibility performance, reveal that factual accuracy does not directly correlate with plausibility performance and that up-to-date training data enhances plausibility assessment across different model architectures.
Recent advancements in 3D perception systems have significantly improved their ability to perform visual recognition tasks such as segmentation. However, these systems still heavily rely on explicit human instruction to identify target objects or categories, lacking the capability to actively reason and comprehend implicit user intentions. We introduce a novel segmentation task known as reasoning part segmentation for 3D objects, aiming to output a segmentation mask based on complex and implicit textual queries about specific parts of a 3D object. To facilitate evaluation and benchmarking, we present a large 3D dataset comprising over 60k instructions paired with corresponding ground-truth part segmentation annotations specifically curated for reasoning-based 3D part segmentation. We propose a model that is capable of segmenting parts of 3D objects based on implicit textual queries and generating natural language explanations corresponding to 3D object segmentation requests. Experiments show that our method achieves competitive performance to models that use explicit queries, with the additional abilities to identify part concepts, reason about them, and complement them with world knowledge. Our source code, dataset, and trained models are available at https://github.com/AmrinKareem/PARIS3D.
Unsupervised pre-training methods utilizing large and diverse datasets have achieved tremendous success across a range of domains. Recent work has investigated such unsupervised pre-training methods for model-based reinforcement learning (MBRL) but is limited to domain-specific or simulated data. In this paper, we study the problem of pre-training world models with abundant in-the-wild videos for efficient learning of downstream visual control tasks. However, in-the-wild videos are complicated with various contextual factors, such as intricate backgrounds and textured appearance, which precludes a world model from extracting shared world knowledge to generalize better. To tackle this issue, we introduce Contextualized World Models (ContextWM) that explicitly separate context and dynamics modeling to overcome the complexity and diversity of in-the-wild videos and facilitate knowledge transfer between distinct scenes. Specifically, a contextualized extension of the latent dynamics model is elaborately realized by incorporating a context encoder to retain contextual information and empower the image decoder, which encourages the latent dynamics model to concentrate on essential temporal variations. Our experiments show that in-the-wild video pre-training equipped with ContextWM can significantly improve the sample efficiency of MBRL in various domains, including robotic manipulation, locomotion, and autonomous driving. Code is available at this repository: https://github.com/thuml/ContextWM.
The lack of training data gives rise to the system cold-start problem in recommendation systems, making them struggle to provide effective recommendations. To address this problem, Large Language Models (LLMs) can model recommendation tasks as language analysis tasks and provide zero-shot results based on their vast open-world knowledge. However, the large scale of the item corpus poses a challenge to LLMs, leading to substantial token consumption that makes it impractical to deploy in real-world recommendation systems. To tackle this challenge, we introduce a tree-based LLM recommendation framework LLMTreeRec, which structures all items into an item tree to improve the efficiency of LLM's item retrieval. LLMTreeRec achieves state-of-the-art performance under the system cold-start setting in two widely used datasets, which is even competitive with conventional deep recommendation systems that use substantial training data. Furthermore, LLMTreeRec outperforms the baseline model in A/B testing on Huawei industrial systems. Consequently, LLMTreeRec demonstrates its effectiveness as an industry-friendly solution that has been successfully deployed online. Our code is available at: https://github.com/Applied-Machine-Learning-Lab/LLMTreeRec.
Narrative reasoning relies on the understanding of eventualities in story contexts, which requires a wealth of background world knowledge. To help machines leverage such knowledge, existing solutions can be categorized into two groups. Some focus on implicitly modeling eventuality knowledge by pretraining language models (LMs) with eventuality-aware objectives. However, this approach breaks down knowledge structures and lacks interpretability. Others explicitly collect world knowledge of eventualities into structured eventuality-centric knowledge graphs (KGs). However, existing research on leveraging these knowledge sources for free-texts is limited. In this work, we propose an initial comprehensive framework called EventGround, which aims to tackle the problem of grounding free-texts to eventuality-centric KGs for contextualized narrative reasoning. We identify two critical problems in this direction: the event representation and sparsity problems. We provide simple yet effective parsing and partial information extraction methods to tackle these problems. Experimental results demonstrate that our approach consistently outperforms baseline models when combined with graph neural network (GNN) or large language model (LLM) based graph reasoning models. Our framework, incorporating grounded knowledge, achieves state-of-the-art performance while providing interpretable evidence.
In real-world applications, users express different behaviors when they interact with different items, including implicit click/like interactions, and explicit comments/reviews interactions. Nevertheless, almost all recommender works are focused on how to describe user preferences by the implicit click/like interactions, to find the synergy of people. For the content-based explicit comments/reviews interactions, some works attempt to utilize them to mine the semantic knowledge to enhance recommender models. However, they still neglect the following two points: (1) The content semantic is a universal world knowledge; how do we extract the multi-aspect semantic information to empower different domains? (2) The user/item ID feature is a fundamental element for recommender models; how do we align the ID and content semantic feature space? In this paper, we propose a `plugin' semantic knowledge transferring method \textbf{LoID}, which includes two major components: (1) LoRA-based large language model pretraining to extract multi-aspect semantic information; (2) ID-based contrastive objective to align their feature spaces. We conduct extensive experiments with SOTA baselines on real-world datasets, the detailed results demonstrating significant improvements of our method LoID.
Recently, 3D Gaussian Splatting (3DGS) has emerged as an efficient approach for accurately representing scenes. However, despite its superior novel view synthesis capabilities, extracting the geometry of the scene directly from the Gaussian properties remains a challenge, as those are optimized based on a photometric loss. While some concurrent models have tried adding geometric constraints during the Gaussian optimization process, they still produce noisy, unrealistic surfaces. We propose a novel approach for bridging the gap between the noisy 3DGS representation and the smooth 3D mesh representation, by injecting real-world knowledge into the depth extraction process. Instead of extracting the geometry of the scene directly from the Gaussian properties, we instead extract the geometry through a pre-trained stereo-matching model. We render stereo-aligned pairs of images corresponding to the original training poses, feed the pairs into a stereo model to get a depth profile, and finally fuse all of the profiles together to get a single mesh. The resulting reconstruction is smoother, more accurate and shows more intricate details compared to other methods for surface reconstruction from Gaussian Splatting, while only requiring a small overhead on top of the fairly short 3DGS optimization process. We performed extensive testing of the proposed method on in-the-wild scenes, obtained using a smartphone, showcasing its superior reconstruction abilities. Additionally, we tested the method on the Tanks and Temples and DTU benchmarks, achieving state-of-the-art results.
Open-domain Question Answering (OpenQA) aims at answering factual questions with an external large-scale knowledge corpus. However, real-world knowledge is not static; it updates and evolves continually. Such a dynamic characteristic of knowledge poses a vital challenge for these models, as the trained models need to constantly adapt to the latest information to make sure that the answers remain accurate. In addition, it is still unclear how well an OpenQA model can transfer to completely new knowledge domains. In this paper, we investigate the generalization performance of a retrieval-augmented QA model in two specific scenarios: 1) adapting to updated versions of the same knowledge corpus; 2) switching to completely different knowledge domains. We observe that the generalization challenges of OpenQA models stem from the reader's over-reliance on memorizing the knowledge from the external corpus, which hinders the model from generalizing to a new knowledge corpus. We introduce Corpus-Invariant Tuning (CIT), a simple but effective training strategy, to mitigate the knowledge over-memorization by controlling the likelihood of retrieved contexts during training. Extensive experimental results on multiple OpenQA benchmarks show that CIT achieves significantly better generalizability without compromising the model's performance in its original corpus and domain.
Large vision-language models (LVLMs) have recently achieved rapid progress, sparking numerous studies to evaluate their multi-modal capabilities. However, we dig into current evaluation works and identify two primary issues: 1) Visual content is unnecessary for many samples. The answers can be directly inferred from the questions and options, or the world knowledge embedded in LLMs. This phenomenon is prevalent across current benchmarks. For instance, GeminiPro achieves 42.9% on the MMMU benchmark without any visual input, and outperforms the random choice baseline across six benchmarks over 24% on average. 2) Unintentional data leakage exists in LLM and LVLM training. LLM and LVLM could still answer some visual-necessary questions without visual content, indicating the memorizing of these samples within large-scale training data. For example, Sphinx-X-MoE gets 43.6% on MMMU without accessing images, surpassing its LLM backbone with 17.9%. Both problems lead to misjudgments of actual multi-modal gains and potentially misguide the study of LVLM. To this end, we present MMStar, an elite vision-indispensable multi-modal benchmark comprising 1,500 samples meticulously selected by humans. MMStar benchmarks 6 core capabilities and 18 detailed axes, aiming to evaluate LVLMs' multi-modal capacities with carefully balanced and purified samples. These samples are first roughly selected from current benchmarks with an automated pipeline, human review is then involved to ensure each curated sample exhibits visual dependency, minimal data leakage, and requires advanced multi-modal capabilities. Moreover, two metrics are developed to measure data leakage and actual performance gain in multi-modal training. We evaluate 16 leading LVLMs on MMStar to assess their multi-modal capabilities, and on 7 benchmarks with the proposed metrics to investigate their data leakage and actual multi-modal gain.
Most studies on machine learning in sensing systems focus on low-level perception tasks that process raw sensory data within a short time window. However, many practical applications, such as human routine modeling and occupancy tracking, require high-level reasoning abilities to comprehend concepts and make inferences based on long-term sensor traces. Existing machine learning-based approaches for handling such complex tasks struggle to generalize due to the limited training samples and the high dimensionality of sensor traces, necessitating the integration of human knowledge for designing first-principle models or logic reasoning methods. We pose a fundamental question: Can we harness the reasoning capabilities and world knowledge of Large Language Models (LLMs) to recognize complex events from long-term spatiotemporal sensor traces? To answer this question, we design an effective prompting framework for LLMs on high-level reasoning tasks, which can handle traces from the raw sensor data as well as the low-level perception results. We also design two strategies to enhance performance with long sensor traces, including summarization before reasoning and selective inclusion of historical traces. Our framework can be implemented in an edge-cloud setup, running small LLMs on the edge for data summarization and performing high-level reasoning on the cloud for privacy preservation. The results show that LLMSense can achieve over 80\% accuracy on two high-level reasoning tasks such as dementia diagnosis with behavior traces and occupancy tracking with environmental sensor traces. This paper provides a few insights and guidelines for leveraging LLM for high-level reasoning on sensor traces and highlights several directions for future work.
Constructing personalized and anthropomorphic agents holds significant importance in the simulation of social networks. However, there are still two key problems in existing works: the agent possesses world knowledge that does not belong to its personas, and it cannot eliminate the interference of diverse persona information on current actions, which reduces the personalization and anthropomorphism of the agent. To solve the above problems, we construct the social media agent based on personalized knowledge and dynamic persona information. For personalized knowledge, we add external knowledge sources and match them with the persona information of agents, thereby giving the agent personalized world knowledge. For dynamic persona information, we use current action information to internally retrieve the persona information of the agent, thereby reducing the interference of diverse persona information on the current action. To make the agent suitable for social media, we design five basic modules for it: persona, planning, action, memory and reflection. To provide an interaction and verification environment for the agent, we build a social media simulation sandbox. In the experimental verification, automatic and human evaluations demonstrated the effectiveness of the agent we constructed.