Zixian Ma

Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma et al. · uw

In the last year alone, a surge of new benchmarks to measure compositional understanding of vision-language models have permeated the machine learning ecosystem. Given an image, these benchmarks probe a model's ability to identify its associated caption amongst a set of compositional distractors. Surprisingly, we find significant biases in all these benchmarks rendering them hackable. This hackability is so dire that blind models with no access to the image outperform state-of-the-art vision-language models. To remedy this rampant vulnerability, we introduce SugarCrepe, a new benchmark for vision-language compositionality evaluation. We employ large language models, instead of rule-based templates used in previous benchmarks, to generate fluent and sensical hard negatives, and utilize an adversarial refinement mechanism to maximally reduce biases. We re-evaluate state-of-the-art models and recently proposed compositionality inducing strategies, and find that their improvements were hugely overestimated, suggesting that more innovation is needed in this important direction. We release SugarCrepe and the code for evaluation at: https://github.com/RAIVNLab/sugar-crepe.

Zixian Ma, Jerry Hong, Mustafa Omer Gul et al. · uw

A fundamental characteristic common to both human vision and natural language is their compositional nature. Yet, despite the performance gains contributed by large vision and language pretraining, we find that: across 7 architectures trained with 4 algorithms on massive datasets, they struggle at compositionality. To arrive at this conclusion, we introduce a new compositionality evaluation benchmark, CREPE, which measures two important aspects of compositionality identified by cognitive science literature: systematicity and productivity. To measure systematicity, CREPE consists of a test dataset containing over $370K$ image-text pairs and three different seen-unseen splits. The three splits are designed to test models trained on three popular training datasets: CC-12M, YFCC-15M, and LAION-400M. We also generate $325K$, $316K$, and $309K$ hard negative captions for a subset of the pairs. To test productivity, CREPE contains $17K$ image-text pairs with nine different complexities plus $183K$ hard negative captions with atomic, swapping and negation foils. The datasets are generated by repurposing the Visual Genome scene graphs and region descriptions and applying handcrafted templates and GPT-3. For systematicity, we find that model performance decreases consistently when novel compositions dominate the retrieval set, with Recall@1 dropping by up to $12\%$. For productivity, models' retrieval success decays as complexity increases, frequently nearing random chance at high complexity. These results hold regardless of model and training dataset size.

Haolin Chen, Deon Metelski, Leon Qi et al.

End-to-end automation of realistic healthcare operations stresses three capabilities underrepresented in current benchmarks: policy density, decisions must be grounded in a large library of medical, insurance, and operational rules; Multi-role composition: a single task requires the agent to play multiple roles with handoffs; and multilateral interaction: intermediate workflow steps are multi-turn dialogs, such as peer-to-peer review and patient outreach. We introduce $χ$-Bench, a benchmark of long-horizon healthcare workflows across three domains: provider prior authorization, payer utilization management, and care management. Each task hands the agent a clinical case in a high-fidelity simulator of 20 healthcare apps exposed via 87 MCP tools, which it must drive to a terminal status through tool calls and writing the role's artifacts, guided by a 1,290+ document managed-care operations handbook skill. Across 30 agent harness/models configurations, the best agent resolves only 28.0% of tasks, no agent clears 20% on strict pass^3, and executing all tasks in a single session slumps the performance to 3.8%. These results raise the hypothesis that similar gaps are likely to surface in other policy-dense, role-composed, irreversible enterprise domains.

Jitesh Jain, Jialuo Li, Zixian Ma et al. · gatech

As humans, we are natural any-horizon reasoners, i.e., we can decide whether to iteratively skim long videos or watch short ones in full when necessary for a given task. With this in mind, one would expect video reasoning models to reason flexibly across different durations. However, SOTA models are still trained to predict answers in a single turn while processing a large number of frames, akin to watching an entire long video, requiring significant resources. This raises the question: Is it possible to develop performant any-horizon video reasoning systems? Inspired by human behavior, we first propose SAGE, an agent system that performs multi-turn reasoning on long videos while handling simpler problems in a single turn. Secondly, we introduce an easy synthetic data generation pipeline using Gemini-2.5-Flash to train the orchestrator, SAGE-MM, which lies at the core of SAGE. We further propose an effective RL post-training recipe essential for instilling any-horizon reasoning ability in SAGE-MM. Thirdly, we curate SAGE-Bench with an average duration of greater than 700 seconds for evaluating video reasoning ability in real-world entertainment use cases. Lastly, we empirically validate the effectiveness of our system, data, and RL recipe, observing notable improvements of up to 6.1% on open-ended video reasoning tasks, as well as an impressive 8.2% improvement on videos longer than 10 minutes.

Zixian Ma, Weikai Huang, Jieyu Zhang et al. · uw

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 10 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).

Zetong Zhou, Dongping Chen, Zixian Ma et al.

Visual reasoning, a cornerstone of human intelligence, encompasses complex perceptual and logical processes essential for solving diverse visual problems. While advances in computer vision have produced powerful models for various perceptual tasks, leveraging these for general visual reasoning remains challenging. Prior work demonstrates that augmenting LLMs with vision models via supervised finetuning improves performance, but faces key limitations such as expensive data generation, reliance on careful data filtering, and poor generalization. To address these issues, we propose ReVPT to enhance multi-modal LLMs' abilities to reason about and use visual tools through reinforcement learning. We introduce a novel RL algorithm based on GRPO, designed to train models to reason with a suite of four visual tools. Through extensive experiments, we show that our method achieves state-of-the-art performance on several perception-heavy benchmarks, including SAT, CV-Bench, BLINK and MMStar, significantly outperforming the supervised and text-based RL finetuning baselines. Notably, Our ReVPT-3B and ReVPT-7B outperform the instruct models by 9.03% and 9.44% on CV-Bench. Finally, we bring to the community new insights on RL-based visual tool-usage through extensive ablations. Our code is available at https://github.com/ls-kelvin/REVPT.

Zixian Ma, Jianguo Zhang, Zhiwei Liu et al. · salesforce, stanford

While open-source vision-language models perform well on simple question-answering, they still struggle with complex questions that require both perceptual and reasoning capabilities. We propose LATTE, a family of vision-language models that have LeArned to Think wiTh vision spEcialists. By offloading perception to state-of-the-art vision models, our approach enables vision-language models to focus solely on reasoning over high-quality perceptual information. To train LATTE, we synthesize and filter a large dataset of 293K multi-modal reasoning traces over perceptual outputs of vision specialists. LATTE trained on this data achieves significant 4-5% gains over baselines across 6 benchmarks covering both perception and reasoning abilities. Ablation studies reveal that the effectiveness of multi-modal reasoning traces depends on the data sources, formats, and quality of thoughts.

Baiqi Li, Zhiqiu Lin, Wenxuan Peng et al.

Vision-language models (VLMs) have made significant progress in recent visual-question-answering (VQA) benchmarks that evaluate complex visio-linguistic reasoning. However, are these models truly effective? In this work, we show that VLMs still struggle with natural images and questions that humans can easily answer, which we term natural adversarial samples. We also find it surprisingly easy to generate these VQA samples from natural image-text corpora using off-the-shelf models like CLIP and ChatGPT. We propose a semi-automated approach to collect a new benchmark, NaturalBench, for reliably evaluating VLMs with 10,000 human-verified VQA samples. Crucially, we adopt a $\textbf{vision-centric}$ design by pairing each question with two images that yield different answers, preventing blind solutions from answering without using the images. This makes NaturalBench more challenging than previous benchmarks that can be solved with commonsense priors. We evaluate 53 state-of-the-art VLMs on NaturalBench, showing that models like LLaVA-OneVision, Cambrian-1, Llama3.2-Vision, Molmo, Qwen2-VL, and even GPT-4o lag 50%-70% behind human performance (over 90%). We analyze why NaturalBench is hard from two angles: (1) Compositionality: Solving NaturalBench requires diverse visio-linguistic skills, including understanding attribute bindings, object relationships, and advanced reasoning like logic and counting. To this end, unlike prior work that uses a single tag per sample, we tag each NaturalBench sample with 1 to 8 skill tags for fine-grained evaluation. (2) Biases: NaturalBench exposes severe biases in VLMs, as models often choose the same answer regardless of the image. Lastly, we apply our benchmark curation method to diverse data sources, including long captions (over 100 words) and non-English languages like Chinese and Hindi, highlighting its potential for dynamic evaluations of VLMs.

Jieyu Zhang, Le Xue, Linxin Song et al. · salesforce, stanford

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.

Linjie Li, Mahtab Bigverdi, Jiawei Gu et al.

Spatial cognition is essential for human intelligence, enabling problem-solving through visual simulations rather than solely relying on verbal reasoning. However, existing AI benchmarks primarily assess verbal reasoning, neglecting the complexities of non-verbal, multi-step visual simulation. We introduce STARE(Spatial Transformations and Reasoning Evaluation), a benchmark designed to rigorously evaluate multimodal large language models on tasks better solved through multi-step visual simulation. STARE features 4K tasks spanning foundational geometric transformations (2D and 3D), integrated spatial reasoning (cube net folding and tangram puzzles), and real-world spatial reasoning (perspective and temporal reasoning), reflecting practical cognitive challenges like object assembly, mechanical diagram interpretation, and everyday spatial navigation. Our evaluations show that models excel at reasoning over simpler 2D transformations, but perform close to random chance on more complex tasks like 3D cube net folding and tangram puzzles that require multi-step visual simulations. Humans achieve near-perfect accuracy but take considerable time (up to 28.9s) on complex tasks, significantly speeding up (down by 7.5 seconds on average) with intermediate visual simulations. In contrast, models exhibit inconsistent performance gains from visual simulations, improving on most tasks but declining in specific cases like tangram puzzles (GPT-4o, o1) and cube net folding (Claude-3.5, Gemini-2.0 Flash), indicating that models may not know how to effectively leverage intermediate visual information.

Jiyue Jiang, Pengan Chen, Jiuming Wang et al.

Large language models (LLMs) have become important tools in solving biological problems, offering improvements in accuracy and adaptability over conventional methods. Several benchmarks have been proposed to evaluate the performance of these LLMs. However, current benchmarks can hardly evaluate the performance of these models across diverse tasks effectively. In this paper, we introduce a comprehensive prompting-based benchmarking framework, termed Bio-benchmark, which includes 30 key bioinformatics tasks covering areas such as proteins, RNA, drugs, electronic health records, and traditional Chinese medicine. Using this benchmark, we evaluate six mainstream LLMs, including GPT-4o and Llama-3.1-70b, etc., using 0-shot and few-shot Chain-of-Thought (CoT) settings without fine-tuning to reveal their intrinsic capabilities. To improve the efficiency of our evaluations, we demonstrate BioFinder, a new tool for extracting answers from LLM responses, which increases extraction accuracy by round 30% compared to existing methods. Our benchmark results show the biological tasks suitable for current LLMs and identify specific areas requiring enhancement. Furthermore, we propose targeted prompt engineering strategies for optimizing LLM performance in these contexts. Based on these findings, we provide recommendations for the development of more robust LLMs tailored for various biological applications. This work offers a comprehensive evaluation framework and robust tools to support the application of LLMs in bioinformatics.

Ziang Li, Manasi Ganti, Zixian Ma et al.

Large language models (LLMs) often generate natural language rationales -- free-form explanations that help improve performance on complex reasoning tasks and enhance interpretability for human users. However, evaluating these rationales remains challenging. While recent work has relied on binary preference judgments from humans or LLM judges, such evaluations are often opaque and coarse-grained, offering limited insight into what makes one rationale better than another. In this work, we rethink preference evaluation for LLM-generated rationales by asking: (1) What attributes define good rationales? (2) Can human preferences be explained by these attributes? (3) Can attribute-based evaluation overcome the limitations of binary comparisons? We identify a set of key rationale attributes from prior literature and assess them using automatic metrics, LLM judgments, and human annotations. We then analyze two standard human preference datasets MT Bench and Chatbot Arena using SHAP to identify which attributes best explain human preference outcomes. Finally, we re-evaluate model-generated rationales using attribute-specific ELO scores, revealing more nuanced model comparisons and insights. Our findings suggest that fine-grained attribute evaluations can better characterize rationale quality and guide future research toward more interpretable and reliable evaluation practices.

Xin Liu, Yuntao Wang, Sinan Xie et al.

Camera-based contactless photoplethysmography refers to a set of popular techniques for contactless physiological measurement. The current state-of-the-art neural models are typically trained in a supervised manner using videos accompanied by gold standard physiological measurements. However, they often generalize poorly out-of-domain examples (i.e., videos that are unlike those in the training set). Personalizing models can help improve model generalizability, but many personalization techniques still require some gold standard data. To help alleviate this dependency, in this paper, we present a novel mobile sensing system called MobilePhys, the first mobile personalized remote physiological sensing system, that leverages both front and rear cameras on a smartphone to generate high-quality self-supervised labels for training personalized contactless camera-based PPG models. To evaluate the robustness of MobilePhys, we conducted a user study with 39 participants who completed a set of tasks under different mobile devices, lighting conditions/intensities, motion tasks, and skin types. Our results show that MobilePhys significantly outperforms the state-of-the-art on-device supervised training and few-shot adaptation methods. Through extensive user studies, we further examine how does MobilePhys perform in complex real-world settings. We envision that calibrated or personalized camera-based contactless PPG models generated from our proposed dual-camera mobile sensing system will open the door for numerous future applications such as smart mirrors, fitness and mobile health applications.