Most existing social robot navigation techniques either leverage hand-crafted rules or human demonstrations to connect robot perception to socially compliant actions. However, there remains a significant gap in effectively translating perception into socially compliant actions, much like how human reasoning naturally occurs in dynamic environments. Considering the recent success of Vision-Language Models (VLMs), we propose using language to bridge the gap in human-like reasoning between perception and socially aware robot actions. We create a vision-language dataset, Social robot Navigation via Explainable Interactions (SNEI), featuring 40K human-annotated Visual Question Answers (VQAs) based on 2K human-robot social interactions in unstructured, crowded public spaces, spanning perception, prediction, chain-of-thought reasoning, action, and explanation. We fine-tune a VLM, Social-LLaVA, using SNEI to demonstrate the practical application of our dataset. Social-LLaVA outperforms state-of-the-art models like GPT-4V and Gemini, based on the average of fifteen different human-judge scores across 50 VQA. Deployed onboard a mobile robot, Social-LLaVA enables human-like reasoning, marking a promising step toward socially compliant robot navigation in dynamic public spaces through language reasoning.

Large Vision-Language Models (LVLMs) have demonstrated remarkable performance in performing complex multimodal tasks. However, they are still plagued by object hallucination: the misidentification or misclassification of objects present in images. To this end, we propose HALLUCINOGEN, a novel visual question answering (VQA) object hallucination attack benchmark that utilizes diverse contextual reasoning prompts to evaluate object hallucination in state-of-the-art LVLMs. We design a series of contextual reasoning hallucination prompts to evaluate LVLMs' ability to accurately identify objects in a target image while asking them to perform diverse visual-language tasks such as identifying, locating or performing visual reasoning around specific objects. Further, we extend our benchmark to high-stakes medical applications and introduce MED-HALLUCINOGEN, hallucination attacks tailored to the biomedical domain, and evaluate the hallucination performance of LVLMs on medical images, a critical area where precision is crucial. Finally, we conduct extensive evaluations of eight LVLMs and two hallucination mitigation strategies across multiple datasets to show that current generic and medical LVLMs remain susceptible to hallucination attacks.

With the widespread deployment of Multimodal Large Language Models (MLLMs) for visual-reasoning tasks, improving their safety has become crucial. Recent research indicates that despite training-time safety alignment, these models remain vulnerable to jailbreak attacks. In this work, we first highlight an important safety gap to describe that alignment achieved solely through safety training may be insufficient against jailbreak attacks. To address this vulnerability, we propose Immune, an inference-time defense framework that leverages a safe reward model through controlled decoding to defend against jailbreak attacks. Additionally, we provide a mathematical characterization of Immune, offering provable guarantees against jailbreaks. Extensive evaluations on diverse jailbreak benchmarks using recent MLLMs reveal that Immune effectively enhances model safety while preserving the model's original capabilities. For instance, against text-based jailbreak attacks on LLaVA-1.6, Immune reduces the attack success rate by 57.82% and 16.78% compared to the base MLLM and state-of-the-art defense strategy, respectively.

Understanding information from a collection of multiple documents, particularly those with visually rich elements, is important for document-grounded question answering. This paper introduces VisDoMBench, the first comprehensive benchmark designed to evaluate QA systems in multi-document settings with rich multimodal content, including tables, charts, and presentation slides. We propose VisDoMRAG, a novel multimodal Retrieval Augmented Generation (RAG) approach that simultaneously utilizes visual and textual RAG, combining robust visual retrieval capabilities with sophisticated linguistic reasoning. VisDoMRAG employs a multi-step reasoning process encompassing evidence curation and chain-of-thought reasoning for concurrent textual and visual RAG pipelines. A key novelty of VisDoMRAG is its consistency-constrained modality fusion mechanism, which aligns the reasoning processes across modalities at inference time to produce a coherent final answer. This leads to enhanced accuracy in scenarios where critical information is distributed across modalities and improved answer verifiability through implicit context attribution. Through extensive experiments involving open-source and proprietary large language models, we benchmark state-of-the-art document QA methods on VisDoMBench. Extensive results show that VisDoMRAG outperforms unimodal and long-context LLM baselines for end-to-end multimodal document QA by 12-20%.

Large Language Models (LLMs) have recently demonstrated impressive few-shot learning capabilities through in-context learning (ICL). However, ICL performance is highly dependent on the choice of few-shot demonstrations, making the selection of the most optimal examples a persistent research challenge. This issue is further amplified in low-resource Indic languages, where the scarcity of ground-truth data complicates the selection process. In this work, we propose PromptRefine, a novel Alternating Minimization approach for example selection that improves ICL performance on low-resource Indic languages. PromptRefine leverages auxiliary example banks from related high-resource Indic languages and employs multi-task learning techniques to align language-specific retrievers, enabling effective cross-language retrieval. Additionally, we incorporate diversity in the selected examples to enhance generalization and reduce bias. Through comprehensive evaluations on four text generation tasks -- Cross-Lingual Question Answering, Multilingual Question Answering, Machine Translation, and Cross-Lingual Summarization using state-of-the-art LLMs such as LLAMA-3.1-8B, LLAMA-2-7B, Qwen-2-7B, and Qwen-2.5-7B, we demonstrate that PromptRefine significantly outperforms existing frameworks for retrieving examples.

We present a generative adversarial network to synthesize 3D pose sequences of co-speech upper-body gestures with appropriate affective expressions. Our network consists of two components: a generator to synthesize gestures from a joint embedding space of features encoded from the input speech and the seed poses, and a discriminator to distinguish between the synthesized pose sequences and real 3D pose sequences. We leverage the Mel-frequency cepstral coefficients and the text transcript computed from the input speech in separate encoders in our generator to learn the desired sentiments and the associated affective cues. We design an affective encoder using multi-scale spatial-temporal graph convolutions to transform 3D pose sequences into latent, pose-based affective features. We use our affective encoder in both our generator, where it learns affective features from the seed poses to guide the gesture synthesis, and our discriminator, where it enforces the synthesized gestures to contain the appropriate affective expressions. We perform extensive evaluations on two benchmark datasets for gesture synthesis from the speech, the TED Gesture Dataset and the GENEA Challenge 2020 Dataset. Compared to the best baselines, we improve the mean absolute joint error by 10--33%, the mean acceleration difference by 8--58%, and the Fr\'echet Gesture Distance by 21--34%. We also conduct a user study and observe that compared to the best current baselines, around 15.28% of participants indicated our synthesized gestures appear more plausible, and around 16.32% of participants felt the gestures had more appropriate affective expressions aligned with the speech.

We present an autoencoder-based semi-supervised approach to classify perceived human emotions from walking styles obtained from videos or motion-captured data and represented as sequences of 3D poses. Given the motion on each joint in the pose at each time step extracted from 3D pose sequences, we hierarchically pool these joint motions in a bottom-up manner in the encoder, following the kinematic chains in the human body. We also constrain the latent embeddings of the encoder to contain the space of psychologically-motivated affective features underlying the gaits. We train the decoder to reconstruct the motions per joint per time step in a top-down manner from the latent embeddings. For the annotated data, we also train a classifier to map the latent embeddings to emotion labels. Our semi-supervised approach achieves a mean average precision of 0.84 on the Emotion-Gait benchmark dataset, which contains both labeled and unlabeled gaits collected from multiple sources. We outperform current state-of-art algorithms for both emotion recognition and action recognition from 3D gaits by 7%--23% on the absolute. More importantly, we improve the average precision by 10%--50% on the absolute on classes that each makes up less than 25% of the labeled part of the Emotion-Gait benchmark dataset.

We present a novel classifier network called STEP, to classify perceived human emotion from gaits, based on a Spatial Temporal Graph Convolutional Network (ST-GCN) architecture. Given an RGB video of an individual walking, our formulation implicitly exploits the gait features to classify the emotional state of the human into one of four emotions: happy, sad, angry, or neutral. We use hundreds of annotated real-world gait videos and augment them with thousands of annotated synthetic gaits generated using a novel generative network called STEP-Gen, built on an ST-GCN based Conditional Variational Autoencoder (CVAE). We incorporate a novel push-pull regularization loss in the CVAE formulation of STEP-Gen to generate realistic gaits and improve the classification accuracy of STEP. We also release a novel dataset (E-Gait), which consists of $2,177$ human gaits annotated with perceived emotions along with thousands of synthetic gaits. In practice, STEP can learn the affective features and exhibits classification accuracy of 89% on E-Gait, which is 14 - 30% more accurate over prior methods.

Text-based AI system optimization typically involves a feedback loop scheme where a single LLM generates an evaluation in natural language of the current output to improve the next iteration's output. However, in this work, we empirically demonstrate that for a practical and complex task (code generation) with multiple criteria to evaluate, utilizing only one LLM evaluator tends to let errors in generated code go undetected, thus leading to incorrect evaluations and ultimately suboptimal test case performance. Motivated by this failure case, we assume there exists an optimal evaluation policy that samples an evaluation between response and ground truth. We then theoretically prove that a linear combination of multiple evaluators can approximate this optimal policy. From this insight, we propose AI system optimization via Multiple LLM Evaluators (AIME). AIME is an evaluation protocol that utilizes multiple LLMs that each independently generate an evaluation on separate criteria and then combine them via concatenation. We provide an extensive empirical study showing AIME outperforming baseline methods in code generation tasks, with up to 62% higher error detection rate and up to 16% higher success rate than a single LLM evaluation protocol on LeetCodeHard and HumanEval datasets. We also show that the selection of the number of evaluators and which criteria to utilize is non-trivial as it can impact pact success rate by up to 12%. Pre-trained foundation models, such as Large Language Models (LLMs), have developed rapidly over the recent years (Achiam et al., 2023; Touvron et al., 2023). As the application complexity increases, the shift to AI systems containing multiple components such as LLM-based agents and web search (Xiong et al., 2024), will continue (Zaharia et al., 2024; Yuksekgonul et al., 2024).

The ability to comprehend audio--which includes speech, non-speech sounds, and music--is crucial for AI agents to interact effectively with the world. We present MMAU, a novel benchmark designed to evaluate multimodal audio understanding models on tasks requiring expert-level knowledge and complex reasoning. MMAU comprises 10k carefully curated audio clips paired with human-annotated natural language questions and answers spanning speech, environmental sounds, and music. It includes information extraction and reasoning questions, requiring models to demonstrate 27 distinct skills across unique and challenging tasks. Unlike existing benchmarks, MMAU emphasizes advanced perception and reasoning with domain-specific knowledge, challenging models to tackle tasks akin to those faced by experts. We assess 18 open-source and proprietary (Large) Audio-Language Models, demonstrating the significant challenges posed by MMAU. Notably, even the most advanced Gemini Pro v1.5 achieves only 52.97% accuracy, and the state-of-the-art open-source Qwen2-Audio achieves only 52.50%, highlighting considerable room for improvement. We believe MMAU will drive the audio and multimodal research community to develop more advanced audio understanding models capable of solving complex audio tasks.