We introduce VideoLISA, a video-based multimodal large language model designed to tackle the problem of language-instructed reasoning segmentation in videos. Leveraging the reasoning capabilities and world knowledge of large language models, and augmented by the Segment Anything Model, VideoLISA generates temporally consistent segmentation masks in videos based on language instructions. Existing image-based methods, such as LISA, struggle with video tasks due to the additional temporal dimension, which requires temporal dynamic understanding and consistent segmentation across frames. VideoLISA addresses these challenges by integrating a Sparse Dense Sampling strategy into the video-LLM, which balances temporal context and spatial detail within computational constraints. Additionally, we propose a One-Token-Seg-All approach using a specially designed <TRK> token, enabling the model to segment and track objects across multiple frames. Extensive evaluations on diverse benchmarks, including our newly introduced ReasonVOS benchmark, demonstrate VideoLISA's superior performance in video object segmentation tasks involving complex reasoning, temporal understanding, and object tracking. While optimized for videos, VideoLISA also shows promising generalization to image segmentation, revealing its potential as a unified foundation model for language-instructed object segmentation. Code and model will be available at: https://github.com/showlab/VideoLISA.
Training large language models (LLMs) from scratch is an expensive endeavor, particularly as world knowledge continually evolves. To maintain relevance and accuracy of LLMs, model editing has emerged as a pivotal research area. While these methods hold promise, they can also produce unintended side effects. Their underlying factors and causes remain largely unexplored. This paper delves into a critical factor-question type-by categorizing model editing questions. Our findings reveal that the extent of performance degradation varies significantly across different question types, providing new insights for experimental design in knowledge editing. Furthermore, we investigate whether insights from smaller models can be extrapolated to larger models. Our results indicate discrepancies in findings between models of different sizes, suggesting that insights from smaller models may not necessarily apply to larger models. Additionally, we examine the impact of batch size on side effects, discovering that increasing the batch size can mitigate performance drops.
Curriculum learning is a training mechanism in reinforcement learning (RL) that facilitates the achievement of complex policies by progressively increasing the task difficulty during training. However, designing effective curricula for a specific task often requires extensive domain knowledge and human intervention, which limits its applicability across various domains. Our core idea is that large language models (LLMs), with their extensive training on diverse language data and ability to encapsulate world knowledge, present significant potential for efficiently breaking down tasks and decomposing skills across various robotics environments. Additionally, the demonstrated success of LLMs in translating natural language into executable code for RL agents strengthens their role in generating task curricula. In this work, we propose CurricuLLM, which leverages the high-level planning and programming capabilities of LLMs for curriculum design, thereby enhancing the efficient learning of complex target tasks. CurricuLLM consists of: (Step 1) Generating sequence of subtasks that aid target task learning in natural language form, (Step 2) Translating natural language description of subtasks in executable task code, including the reward code and goal distribution code, and (Step 3) Evaluating trained policies based on trajectory rollout and subtask description. We evaluate CurricuLLM in various robotics simulation environments, ranging from manipulation, navigation, and locomotion, to show that CurricuLLM can aid learning complex robot control tasks. In addition, we validate humanoid locomotion policy learned through CurricuLLM in real-world. Project website is https://iconlab.negarmehr.com/CurricuLLM/
Large language models (LLMs) provide powerful means to leverage prior knowledge for predictive modeling when data is limited. In this work, we demonstrate how LLMs can use their compressed world knowledge to generate intrinsically interpretable machine learning models, i.e., decision trees, without any training data. We find that these zero-shot decision trees can surpass data-driven trees on some small-sized tabular datasets and that embeddings derived from these trees perform on par with data-driven tree-based embeddings on average. Our knowledge-driven decision tree induction and embedding approaches therefore serve as strong new baselines for data-driven machine learning methods in the low-data regime.
During pre-training, the Text-to-Image (T2I) diffusion models encode factual knowledge into their parameters. These parameterized facts enable realistic image generation, but they may become obsolete over time, thereby misrepresenting the current state of the world. Knowledge editing techniques aim to update model knowledge in a targeted way. However, facing the dual challenges posed by inadequate editing datasets and unreliable evaluation criterion, the development of T2I knowledge editing encounter difficulties in effectively generalizing injected knowledge. In this work, we design a T2I knowledge editing framework by comprehensively spanning on three phases: First, we curate a dataset \textbf{CAKE}, comprising paraphrase and multi-object test, to enable more fine-grained assessment on knowledge generalization. Second, we propose a novel criterion, \textbf{adaptive CLIP threshold}, to effectively filter out false successful images under the current criterion and achieve reliable editing evaluation. Finally, we introduce \textbf{MPE}, a simple but effective approach for T2I knowledge editing. Instead of tuning parameters, MPE precisely recognizes and edits the outdated part of the conditioning text-prompt to accommodate the up-to-date knowledge. A straightforward implementation of MPE (Based on in-context learning) exhibits better overall performance than previous model editors. We hope these efforts can further promote faithful evaluation of T2I knowledge editing methods.
Who is the US President? The answer changes depending on when the question is asked. While large language models (LLMs) are evaluated on various reasoning tasks, they often miss a crucial dimension: time. In real-world scenarios, the correctness of answers is frequently tied to temporal context. To address this gap, we present a novel framework and dataset spanning over 8,000 events from 2018 to 2024, annotated with day-level granularity and sourced globally across domains such as politics, science, and business. Our TimeShift evaluation method systematically probes LLMs for temporal reasoning, revealing that base models often outperform instruction-tuned and synthetic-trained counterparts on time-sensitive recall. Additionally, we find that even large-scale models exhibit brittleness in handling paraphrased facts, highlighting unresolved challenges in temporal consistency. By identifying these limitations, our work provides a significant step toward advancing time-aware language models capable of adapting to the dynamic nature of real-world knowledge.
Large language models (LLMs) encode extensive world knowledge through pre-training on massive datasets, which can then be fine-tuned for the question-answering (QA) task. However, effective strategies for fine-tuning LLMs for the QA task remain largely unexplored. To address this gap, we categorize supervised fine-tuning (SFT) data based on the extent of knowledge memorized by the pretrained LLMs and conduct a series of empirical analyses. Our experiments, involving four LLMs from three different model families, focus on three key factors: the amount of data required for SFT, the impact of different SFT datasets on model performance, and how data requirements vary across LLMs. The results show that as few as 60 data points during the SFT stage can activate the knowledge encoded during pre-training, enabling LLMs to perform the QA task. Additionally, SFT with data of varying memory levels has a significant impact on LLM performance, with the optimal dataset differing based on the specific model being fine-tuned. Future research will delve deeper into the mechanisms underlying these phenomena.
The release of ChatGPT in November 2022 sparked an explosion of interest in post-training and an avalanche of new preference optimization (PO) methods. These methods claim superior alignment by virtue of better correspondence with human pairwise preferences, often measured by LLM-judges. In this work, we attempt to answer the following question -- do LLM-judge preferences translate to progress on other, more concrete metrics for alignment, and if not, why not? We define a concrete metric for alignment, and introduce SOS-Bench (Substance Outweighs Style Benchmark), which is to the best of our knowledge the largest standardized, reproducible LLM meta-benchmark to date. We find that (1) LLM-judge preferences do not correlate with concrete measures of safety, world knowledge, and instruction following; (2) LLM-judges have powerful implicit biases, prioritizing style over factuality and safety; and (3) the supervised fine-tuning (SFT) stage of post-training, and not the PO stage, has the greatest impact on alignment, with data scaling and prompt diversity as the driving factors. Our codebase and complete results can be found at https://github.com/penfever/sos-bench.
Human brain possesses the ability to effectively focus on important environmental components, which enhances perception, learning, reasoning, and decision-making. Inspired by this cognitive mechanism, we introduced a novel concept termed relevance for Human-Robot Collaboration (HRC). Relevance is a dimensionality reduction process that incorporates a continuously operating perception module, evaluates cue sufficiency within the scene, and applies a flexible formulation and computation framework. In this paper, we present an enhanced two-loop framework that integrates real-time and asynchronous processing to quantify relevance and leverage it for safer and more efficient human-robot collaboration (HRC). The two-loop framework integrates an asynchronous loop, which leverages LLM world knowledge to quantify relevance, and a real-time loop, which performs scene understanding, human intent prediction, and decision-making based on relevance. HRC decision-making is enhanced by a relevance-based task allocation method, as well as a motion generation and collision avoidance approach that incorporates human trajectory prediction. Simulations and experiments show that our methodology for relevance quantification can accurately and robustly predict the human objective and relevance, with an average accuracy of up to 0.90 for objective prediction and up to 0.96 for relevance prediction. Moreover, our motion generation methodology reduces collision cases by 63.76% and collision frames by 44.74% when compared with a state-of-the-art (SOTA) collision avoidance method. Our framework and methodologies, with relevance, guide the robot on how to best assist humans and generate safer and more efficient actions for HRC.
Large multimodal models have demonstrated impressive problem-solving abilities in vision and language tasks, and have the potential to encode extensive world knowledge. However, it remains an open challenge for these models to perceive, reason, plan, and act in realistic environments. In this work, we introduce Can-Do, a benchmark dataset designed to evaluate embodied planning abilities through more diverse and complex scenarios than previous datasets. Our dataset includes 400 multimodal samples, each consisting of natural language user instructions, visual images depicting the environment, state changes, and corresponding action plans. The data encompasses diverse aspects of commonsense knowledge, physical understanding, and safety awareness. Our fine-grained analysis reveals that state-of-the-art models, including GPT-4V, face bottlenecks in visual perception, comprehension, and reasoning abilities. To address these challenges, we propose NeuroGround, a neurosymbolic framework that first grounds the plan generation in the perceived environment states and then leverages symbolic planning engines to augment the model-generated plans. Experimental results demonstrate the effectiveness of our framework compared to strong baselines. Our code and dataset are available at https://embodied-planning.github.io.
Social networks form a valuable source of world knowledge, where influential entities correspond to popular accounts. Unlike factual knowledge bases (KBs), which maintain a semantic ontology, structured semantic information is not available on social media. In this work, we consider a social KB of roughly 200K popular Twitter accounts, which denotes entities of interest. We elicit semantic information about those entities. In particular, we associate them with a fine-grained set of 136 semantic types, e.g., determine whether a given entity account belongs to a politician, or a musical artist. In the lack of explicit type information in Twitter, we obtain semantic labels for a subset of the accounts via alignment with the KBs of DBpedia and Wikidata. Given the labeled dataset, we finetune a transformer-based text encoder to generate semantic embeddings of the entities based on the contents of their accounts. We then exploit this evidence alongside network-based embeddings to predict the entities semantic types. In our experiments, we show high type prediction performance on the labeled dataset. Consequently, we apply our type classification model to all of the entity accounts in the social KB. Our analysis of the results offers insights about the global semantics of the Twitter sphere. We discuss downstream applications that should benefit from semantic type information and the semantic embeddings of social entities generated in this work. In particular, we demonstrate enhanced performance on the key task of entity similarity assessment using this information.
Large Language Models (LLMs) have achieved remarkable success in various fields, prompting several studies to explore their potential in recommendation systems. However, these attempts have so far resulted in only modest improvements over traditional recommendation models. Moreover, three critical questions remain under-explored: firstly, the real value of LLMs' pre-trained weights, often considered to encapsulate world knowledge; secondly, the necessity of fine-tuning for recommendation tasks; lastly, whether LLMs can exhibit the same scalability benefits in recommendation systems as they do in other domains. In this paper, we propose a novel Hierarchical Large Language Model (HLLM) architecture designed to enhance sequential recommendation systems. Our approach employs a two-tier model: the first Item LLM extracts rich content features from the detailed text description of the item, while the second User LLM utilizes these features to predict users' future interests based on their interaction history. Extensive experiments demonstrate that our method effectively leverages the pre-trained capabilities of open-source LLMs, and further fine-tuning leads to significant performance boosts. Additionally, HLLM achieves excellent scalability, with the largest configuration utilizing 7B parameters for both item feature extraction and user interest modeling. Moreover, HLLM offers excellent training and serving efficiency, making it practical in real-world applications. Evaluations on two large-scale datasets, PixelRec and Amazon Reviews, show that HLLM achieves state-of-the-art results, outperforming traditional ID-based models by a wide margin. In online A/B testing, HLLM showcases notable gains, validating its practical impact in real-world recommendation scenarios. Codes are available at https://github.com/bytedance/HLLM.
Finetuning large language models on instruction data is crucial for enhancing pre-trained knowledge and improving instruction-following capabilities. As instruction datasets proliferate, selecting optimal data for effective training becomes increasingly important. This work addresses the question: How can we determine the optimal subset of data for effective training? While existing research often emphasizes local criteria like instance quality for subset selection, we argue that a global approach focused on data diversity is more critical. Our method employs k-means clustering to ensure the selected subset effectively represents the full dataset. We propose an iterative refinement method inspired by active learning techniques to resample instances from clusters, reassessing each cluster's importance and sampling weight in every training iteration. This approach reduces the effect of outliers and automatically filters out clusters containing low-quality data. Through extensive evaluation across natural language reasoning, general world knowledge, code and math reasoning tasks, and by fine-tuning models from various families, we observe consistent improvements, achieving a 7% increase over random selection and a 3.8% improvement over state-of-the-art sampling methods. Our work highlights the significance of diversity-first sampling when finetuning LLMs to enhance performance across a broad array of evaluation tasks. Our code is available at https://github.com/for-ai/iterative-data-selection.
Visual Language Tracking (VLT) enhances tracking by mitigating the limitations of relying solely on the visual modality, utilizing high-level semantic information through language. This integration of the language enables more advanced human-machine interaction. The essence of interaction is cognitive alignment, which typically requires multiple information exchanges, especially in the sequential decision-making process of VLT. However, current VLT benchmarks do not account for multi-round interactions during tracking. They provide only an initial text and bounding box (bbox) in the first frame, with no further interaction as tracking progresses, deviating from the original motivation of the VLT task. To address these limitations, we propose a novel and robust benchmark, VLT-MI (Visual Language Tracking with Multi-modal Interaction), which introduces multi-round interaction into the VLT task for the first time. (1) We generate diverse, multi-granularity texts for multi-round, multi-modal interaction based on existing mainstream VLT benchmarks using DTLLM-VLT, leveraging the world knowledge of LLMs. (2) We propose a new VLT interaction paradigm that achieves multi-round interaction through text updates and object recovery. When multiple tracking failures occur, we provide the tracker with more aligned texts and corrected bboxes through interaction, thereby expanding the scope of VLT downstream tasks. (3) We conduct comparative experiments on both traditional VLT benchmarks and VLT-MI, evaluating and analyzing the accuracy and robustness of trackers under the interactive paradigm. This work offers new insights and paradigms for the VLT task, enabling a fine-grained evaluation of multi-modal trackers. We believe this approach can be extended to additional datasets in the future, supporting broader evaluations and comparisons of video-language model capabilities.
Pretraining on large-scale, unstructured internet text enables language models to acquire a significant amount of world knowledge. However, this knowledge acquisition is data-inefficient--to learn a given fact, models must be trained on hundreds to thousands of diverse representations of it. This poses a challenge when adapting a pretrained model to a small corpus of domain-specific documents, where each fact may appear rarely or only once. We propose to bridge this gap with synthetic continued pretraining: using the small domain-specific corpus to synthesize a large corpus more amenable to learning, and then performing continued pretraining on the synthesized corpus. We instantiate this proposal with EntiGraph, a synthetic data augmentation algorithm that extracts salient entities from the source documents and then generates diverse text by drawing connections between the sampled entities. Synthetic continued pretraining with EntiGraph enables a language model to answer questions and follow generic instructions related to the source documents without access to them. If, instead, the source documents are available at inference time, we show that the knowledge acquired through our approach compounds with retrieval-augmented generation. To better understand these results, we build a simple mathematical model of EntiGraph, and show how synthetic data augmentation can "rearrange" knowledge to enable more data-efficient learning.
The generation of corner cases has become increasingly crucial for efficiently testing autonomous vehicles prior to road deployment. However, existing methods struggle to accommodate diverse testing requirements and often lack the ability to generalize to unseen situations, thereby reducing the convenience and usability of the generated scenarios. A method that facilitates easily controllable scenario generation for efficient autonomous vehicles (AV) testing with realistic and challenging situations is greatly needed. To address this, we proposed OmniTester: a multimodal Large Language Model (LLM) based framework that fully leverages the extensive world knowledge and reasoning capabilities of LLMs. OmniTester is designed to generate realistic and diverse scenarios within a simulation environment, offering a robust solution for testing and evaluating AVs. In addition to prompt engineering, we employ tools from Simulation of Urban Mobility to simplify the complexity of codes generated by LLMs. Furthermore, we incorporate Retrieval-Augmented Generation and a self-improvement mechanism to enhance the LLM's understanding of scenarios, thereby increasing its ability to produce more realistic scenes. In the experiments, we demonstrated the controllability and realism of our approaches in generating three types of challenging and complex scenarios. Additionally, we showcased its effectiveness in reconstructing new scenarios described in crash report, driven by the generalization capability of LLMs.
Large language models are increasingly trained on corpora containing both natural language and non-linguistic data like source code. Aside from aiding programming-related tasks, anecdotal evidence suggests that including code in pretraining corpora may improve performance on other, unrelated tasks, yet to date no work has been able to establish a causal connection by controlling between language and code data. Here we do just this. We pretrain language models on datasets which interleave natural language and code in two different settings: additive, in which the total volume of data seen during pretraining is held constant; and competitive, in which the volume of language data is held constant. We study how the pretraining mixture affects performance on (a) a diverse collection of tasks included in the BigBench benchmark, and (b) compositionality, measured by generalization accuracy on semantic parsing and syntactic transformations. We find that pretraining on higher proportions of code improves performance on compositional tasks involving structured output (like semantic parsing), and mathematics. Conversely, increase code mixture can harm performance on other tasks, including on tasks that requires sensitivity to linguistic structure such as syntax or morphology, and tasks measuring real-world knowledge.
Out-of-distribution (OOD) object detection is a challenging task due to the absence of open-set OOD data. Inspired by recent advancements in text-to-image generative models, such as Stable Diffusion, we study the potential of generative models trained on large-scale open-set data to synthesize OOD samples, thereby enhancing OOD object detection. We introduce SyncOOD, a simple data curation method that capitalizes on the capabilities of large foundation models to automatically extract meaningful OOD data from text-to-image generative models. This offers the model access to open-world knowledge encapsulated within off-the-shelf foundation models. The synthetic OOD samples are then employed to augment the training of a lightweight, plug-and-play OOD detector, thus effectively optimizing the in-distribution (ID)/OOD decision boundaries. Extensive experiments across multiple benchmarks demonstrate that SyncOOD significantly outperforms existing methods, establishing new state-of-the-art performance with minimal synthetic data usage.
Vision-language models (VLMs) have shown impressive zero- and few-shot performance on real-world visual question answering (VQA) benchmarks, alluding to their capabilities as visual reasoning engines. However, the benchmarks being used conflate "pure" visual reasoning with world knowledge, and also have questions that involve a limited number of reasoning steps. Thus, it remains unclear whether a VLM's apparent visual reasoning performance is due to its world knowledge, or due to actual visual reasoning capabilities. To clarify this ambiguity, we systematically benchmark and dissect the zero-shot visual reasoning capabilities of VLMs through synthetic datasets that require minimal world knowledge, and allow for analysis over a broad range of reasoning steps. We focus on two novel aspects of zero-shot visual reasoning: i) evaluating the impact of conveying scene information as either visual embeddings or purely textual scene descriptions to the underlying large language model (LLM) of the VLM, and ii) comparing the effectiveness of chain-of-thought prompting to standard prompting for zero-shot visual reasoning. We find that the underlying LLMs, when provided textual scene descriptions, consistently perform better compared to being provided visual embeddings. In particular, 18% higher accuracy is achieved on the PTR dataset. We also find that CoT prompting performs marginally better than standard prompting only for the comparatively large GPT-3.5-Turbo (175B) model, and does worse for smaller-scale models. This suggests the emergence of CoT abilities for visual reasoning in LLMs at larger scales even when world knowledge is limited. Overall, we find limitations in the abilities of VLMs and LLMs for more complex visual reasoning, and highlight the important role that LLMs can play in visual reasoning.
This study investigates the ability of various vision-language (VL) models to ground context-dependent and non-context-dependent verb phrases. To do that, we introduce the CV-Probes dataset, designed explicitly for studying context understanding, containing image-caption pairs with context-dependent verbs (e.g., "beg") and non-context-dependent verbs (e.g., "sit"). We employ the MM-SHAP evaluation to assess the contribution of verb tokens towards model predictions. Our results indicate that VL models struggle to ground context-dependent verb phrases effectively. These findings highlight the challenges in training VL models to integrate context accurately, suggesting a need for improved methodologies in VL model training and evaluation.