Despite the remarkable success of Large Language Models (LLMs), evaluating their outputs' quality regarding *preference* remains a critical challenge. Existing works usually leverage a powerful LLM (e.g., GPT4) as the judge for comparing LLMs' output pairwisely, yet such model-based evaluator is vulnerable to *conflicting preference*, i.e., output A is better than B, B than C, but C than A, causing contradictory evaluation results. To improve model-based preference evaluation, we introduce GED (Preference Graph Ensemble and Denoise), a novel approach that leverages multiple model-based evaluators to construct preference graphs, and then ensemble and denoise these graphs for better, non-contradictory evaluation results. In particular, our method consists of two primary stages: aggregating evaluations into a unified graph and applying a denoising process to eliminate cyclic inconsistencies, ensuring a directed acyclic graph (DAG) structure. We provide theoretical guarantees for our framework, demonstrating its efficacy in recovering the ground truth preference structure. Extensive experiments across ten benchmark datasets show that GED outperforms baseline methods in model ranking, response selection, and model alignment tasks. Notably, GED combines weaker evaluators like Llama3-8B, Mistral-7B, and Qwen2-7B to surpass the performance of stronger evaluators like Qwen2-72B, highlighting its ability to enhance evaluation reliability and improve model performance.

With the rise of multimodal applications, instruction data has become critical for training multimodal language models capable of understanding complex image-based queries. Existing practices rely on powerful but costly large language models (LLMs) or multimodal language models (MLMs) to produce instruction data. These are often prone to hallucinations, licensing issues and the generation process is often hard to scale and interpret. In this work, we present a programmatic approach that employs scene graphs as symbolic representations of images and human-written programs to systematically synthesize vision-centric instruction data. Our approach ensures the interpretability and controllability of the data generation process and scales efficiently while maintaining factual accuracy. By implementing a suite of 24 single-image, 14 multi-image instruction generators, and a scene graph generation pipeline, we build a scalable, cost-effective system: ProVision which produces diverse question-answer pairs concerning objects, attributes, relations, depth, etc., for any given image. Applied to Visual Genome and DataComp datasets, we generate over 10 million instruction data points, ProVision-10M, and leverage them in both pretraining and instruction tuning stages of MLMs. When adopted in the instruction tuning stage, our single-image instruction data yields up to a 7% improvement on the 2D split and 8% on the 3D split of CVBench, along with a 3% increase in performance on QBench2, RealWorldQA, and MMMU. Our multi-image instruction data leads to an 8% improvement on Mantis-Eval. Incorporation of our data in both pre-training and fine-tuning stages of xGen-MM-4B leads to an averaged improvement of 1.6% across 11 benchmarks.

DALL-E and Sora have gained attention by producing implausible images, such as "astronauts riding a horse in space." Despite the proliferation of text-to-vision models that have inundated the internet with synthetic visuals, from images to 3D assets, current benchmarks predominantly evaluate these models on real-world scenes paired with captions. We introduce Generate Any Scene, a framework that systematically enumerates scene graphs representing a vast array of visual scenes, spanning realistic to imaginative compositions. Generate Any Scene leverages 'scene graph programming', a method for dynamically constructing scene graphs of varying complexity from a structured taxonomy of visual elements. This taxonomy includes numerous objects, attributes, and relations, enabling the synthesis of an almost infinite variety of scene graphs. Using these structured representations, Generate Any Scene translates each scene graph into a caption, enabling scalable evaluation of text-to-vision models through standard metrics. We conduct extensive evaluations across multiple text-to-image, text-to-video, and text-to-3D models, presenting key findings on model performance. We find that DiT-backbone text-to-image models align more closely with input captions than UNet-backbone models. Text-to-video models struggle with balancing dynamics and consistency, while both text-to-video and text-to-3D models show notable gaps in human preference alignment. We demonstrate the effectiveness of Generate Any Scene by conducting three practical applications leveraging captions generated by Generate Any Scene: 1) a self-improving framework where models iteratively enhance their performance using generated data, 2) a distillation process to transfer specific strengths from proprietary models to open-source counterparts, and 3) improvements in content moderation by identifying and generating challenging synthetic data.

Recent advancements have enabled Large Language Models (LLMs) to function as agents that can perform actions using external tools. This requires registering, i.e., integrating tool information into the LLM context prior to taking actions. Current methods indiscriminately incorporate all candidate tools into the agent's context and retain them across multiple reasoning steps. This process remains opaque to LLM agents and is not integrated into their reasoning procedures, leading to inefficiencies due to increased context length from irrelevant tools. To address this, we introduce EcoAct, a tool using algorithm that allows LLMs to selectively register tools as needed, optimizing context use. By integrating the tool registration process into the reasoning procedure, EcoAct reduces computational costs by over 50% in multiple steps reasoning tasks while maintaining performance, as demonstrated through extensive experiments. Moreover, it can be plugged into any reasoning pipeline with only minor modifications to the prompt, making it applicable to LLM agents now and future.

Recent advancements in multimodal foundation models have showcased impressive capabilities in understanding and reasoning with visual and textual information. Adapting these foundation models trained for general usage to specialized domains like biomedicine requires large-scale domain-specific instruction datasets. While existing works have explored curating such datasets automatically, the resultant datasets are not explicitly aligned with domain expertise. In this work, we propose a data-centric framework, Biomedical Visual Instruction Tuning with Clinician Preference Alignment (BioMed-VITAL), that incorporates clinician preferences into both stages of generating and selecting instruction data for tuning biomedical multimodal foundation models. First, during the generation stage, we prompt the GPT-4V generator with a diverse set of clinician-selected demonstrations for preference-aligned data candidate generation. Then, during the selection phase, we train a separate selection model, which explicitly distills clinician and policy-guided model preferences into a rating function to select high-quality data for medical instruction tuning. Results show that the model tuned with the instruction-following data from our method demonstrates a significant improvement in open visual chat (18.5% relatively) and medical VQA (win rate up to 81.73%). Our instruction-following data and models are available at BioMed-VITAL.github.io.

Recently, large language model (LLM)-based preference evaluation has been widely adopted to compare pairs of model responses. However, a severe bias towards lengthy responses has been observed, raising concerns about the reliability of this evaluation method. In this work, we designed a series of controlled experiments to study the major impacting factors of the metric of LLM-based preference evaluation, i.e., win rate, and conclude that the win rate is affected by two axes of model response: desirability and information mass, where the former is length-independent and related to trustworthiness, and the latter is length-dependent and can be represented by conditional entropy. We find that length impacts the existing evaluations by influencing information mass. However, a reliable evaluation metric should not only assess content quality but also ensure that the assessment is not confounded by extraneous factors such as response length. Therefore, we propose a simple yet effective adjustment, AdapAlpaca, to the existing practice of win rate measurement. Specifically, by adjusting the lengths of reference answers to match the test model's answers within the same interval, we debias information mass relative to length, ensuring a fair model evaluation.

Benchmarks for large multimodal language models (MLMs) now serve to simultaneously assess the general capabilities of models instead of evaluating for a specific capability. As a result, when a developer wants to identify which models to use for their application, they are overwhelmed by the number of benchmarks and remain uncertain about which benchmark's results are most reflective of their specific use case. This paper introduces Task-Me-Anything, a benchmark generation engine which produces a benchmark tailored to a user's needs. Task-Me-Anything maintains an extendable taxonomy of visual assets and can programmatically generate a vast number of task instances. Additionally, it algorithmically addresses user queries regarding MLM performance efficiently within a computational budget. It contains 113K images, 10K videos, 2K 3D object assets, over 365 object categories, 655 attributes, and 335 relationships. It can generate 750M image/video question-answering pairs, which focus on evaluating MLM perceptual capabilities. Task-Me-Anything reveals critical insights: open-source MLMs excel in object and attribute recognition but lack spatial and temporal understanding; each model exhibits unique strengths and weaknesses; larger models generally perform better, though exceptions exist; and GPT4o demonstrates challenges in recognizing rotating/moving objects and distinguishing colors.

Researchers and practitioners have recently reframed powerful Large Language Models (LLMs) as agents, enabling them to automate complex tasks largely via the use of specialized functions. To facilitate the development of LLM agents, we present a novel paradigm of training LLM agents without modifying the LLM weights, which is particularly useful when the LLMs are difficult or inaccessible for modifications. Inspired by how humans continuously forge tools to adapt to real-world tasks, rather than change our biological structure to fit a static set of tools, we propose to progressively forge agent's functions to better solve the downstream tasks instead of modifying the LLM weights. By treating the functions as learnable `agent parameters' and leveraging the fundamental idea of model training in artificial intelligence, we develop AgentOptimizer that employs the LLM to update agents' functions and devise an agent training algorithm with two strategies, roll-back, and early-stop, to streamline the training process. With extensive experiments, we showcase that the agent training paradigm could significantly improve the performance of representative LLM agents in various downstream tasks. We also study the behavior of the agent training regarding aspects like the learning curve and domain transferability.

Leveraging multiple large language model (LLM) agents has shown to be a promising approach for tackling complex tasks, while the effective design of multiple agents for a particular application remains an art. It is thus intriguing to answer a critical question: Given a task, how can we build a team of LLM agents to solve it effectively? Our new adaptive team-building paradigm offers a flexible solution, realized through a novel agent design named Captain Agent. It dynamically forms and manages teams for each step of a task-solving process, utilizing nested group conversations and reflection to ensure diverse expertise and prevent stereotypical outputs. It allows for a flexible yet structured approach to problem-solving and can help reduce redundancy and enhance output diversity. A comprehensive evaluation across six real-world scenarios demonstrates that Captain Agent significantly outperforms existing multi-agent methods with 21.94% improvement in average accuracy, providing outstanding performance without requiring task-specific prompt engineering.

Real-world multi-modal problems are rarely solved by a single machine learning model, and often require multi-step computational plans that involve stitching several models. Tool-augmented LLMs hold tremendous promise for automating the generation of such computational plans. However, the lack of standardized benchmarks for evaluating LLMs as planners for multi-step multi-modal tasks has prevented a systematic study of planner design decisions. Should LLMs generate a full plan in a single shot or step-by-step? Should they invoke tools directly with Python code or through structured data formats like JSON? Does feedback improve planning? To answer these questions and more, we introduce m&m's: a benchmark containing 4K+ multi-step multi-modal tasks involving 33 tools that include multi-modal models, (free) public APIs, and image processing modules. For each of these task queries, we provide automatically generated plans using this realistic toolset. We further provide a high-quality subset of 1,565 task plans that are human-verified and correctly executable. With m&m's, we evaluate 6 popular LLMs with 2 planning strategies (multi-step vs. step-by-step planning), 2 plan formats (JSON vs. code), and 3 types of feedback (parsing/verification/execution). Finally, we summarize takeaways from our extensive experiments. Our dataset and code are available on HuggingFace (https://huggingface.co/datasets/zixianma/mnms) and Github (https://github.com/RAIVNLab/mnms).