Pelican: Correcting Hallucination in Vision-LLMs via Claim Decomposition and Program of Thought Verification

Large Visual Language Models (LVLMs) struggle with hallucinations in visual instruction following task(s), limiting their trustworthiness and real-world applicability. We propose Pelican -- a novel framework designed to detect and mitigate hallucinations through claim verification. Pelican first decomposes the visual claim into a chain of sub-claims based on first-order predicates. These sub-claims consist of (predicate, question) pairs and can be conceptualized as nodes of a computational graph. We then use Program-of-Thought prompting to generate Python code for answering these questions through flexible composition of external tools. Pelican improves over prior work by introducing (1) intermediate variables for precise grounding of object instances, and (2) shared computation for answering the sub-question to enable adaptive corrections and inconsistency identification. We finally use reasoning abilities of LLM to verify the correctness of the the claim by considering the consistency and confidence of the (question, answer) pairs from each sub-claim. Our experiments reveal a drop in hallucination rate by $\sim$8%-32% across various baseline LVLMs and a 27% drop compared to approaches proposed for hallucination mitigation on MMHal-Bench. Results on two other benchmarks further corroborate our results.

Unpacking Large Language Models with Conceptual Consistency

If a Large Language Model (LLM) answers "yes" to the question "Are mountains tall?" then does it know what a mountain is? Can you rely on it responding correctly or incorrectly to other questions about mountains? The success of Large Language Models (LLMs) indicates they are increasingly able to answer queries like these accurately, but that ability does not necessarily imply a general understanding of concepts relevant to the anchor query. We propose conceptual consistency to measure a LLM's understanding of relevant concepts. This novel metric measures how well a model can be characterized by finding out how consistent its responses to queries about conceptually relevant background knowledge are. To compute it we extract background knowledge by traversing paths between concepts in a knowledge base and then try to predict the model's response to the anchor query from the background knowledge. We investigate the performance of current LLMs in a commonsense reasoning setting using the CSQA dataset and the ConceptNet knowledge base. While conceptual consistency, like other metrics, does increase with the scale of the LLM used, we find that popular models do not necessarily have high conceptual consistency. Our analysis also shows significant variation in conceptual consistency across different kinds of relations, concepts, and prompts. This serves as a step toward building models that humans can apply a theory of mind to, and thus interact with intuitively.

SAViR-T: Spatially Attentive Visual Reasoning with Transformers

We present a novel computational model, "SAViR-T", for the family of visual reasoning problems embodied in the Raven's Progressive Matrices (RPM). Our model considers explicit spatial semantics of visual elements within each image in the puzzle, encoded as spatio-visual tokens, and learns the intra-image as well as the inter-image token dependencies, highly relevant for the visual reasoning task. Token-wise relationship, modeled through a transformer-based SAViR-T architecture, extract group (row or column) driven representations by leveraging the group-rule coherence and use this as the inductive bias to extract the underlying rule representations in the top two row (or column) per token in the RPM. We use this relation representations to locate the correct choice image that completes the last row or column for the RPM. Extensive experiments across both synthetic RPM benchmarks, including RAVEN, I-RAVEN, RAVEN-FAIR, and PGM, and the natural image-based "V-PROM" demonstrate that SAViR-T sets a new state-of-the-art for visual reasoning, exceeding prior models' performance by a considerable margin.

Challenges in Procedural Multimodal Machine Comprehension:A Novel Way To Benchmark

We focus on Multimodal Machine Reading Comprehension (M3C) where a model is expected to answer questions based on given passage (or context), and the context and the questions can be in different modalities. Previous works such as RecipeQA have proposed datasets and cloze-style tasks for evaluation. However, we identify three critical biases stemming from the question-answer generation process and memorization capabilities of large deep models. These biases makes it easier for a model to overfit by relying on spurious correlations or naive data patterns. We propose a systematic framework to address these biases through three Control-Knobs that enable us to generate a test bed of datasets of progressive difficulty levels. We believe that our benchmark (referred to as Meta-RecipeQA) will provide, for the first time, a fine grained estimate of a model's generalization capabilities. We also propose a general M3C model that is used to realize several prior SOTA models and motivate a novel hierarchical transformer based reasoning network (HTRN). We perform a detailed evaluation of these models with different language and visual features on our benchmark. We observe a consistent improvement with HTRN over SOTA (~18% in Visual Cloze task and ~13% in average over all the tasks). We also observe a drop in performance across all the models when testing on RecipeQA and proposed Meta-RecipeQA (e.g. 83.6% versus 67.1% for HTRN), which shows that the proposed dataset is relatively less biased. We conclude by highlighting the impact of the control knobs with some quantitative results.

DAReN: A Collaborative Approach Towards Reasoning And Disentangling

Computational learning approaches to solving visual reasoning tests, such as Raven's Progressive Matrices (RPM),critically depend on the ability of the computational approach to identify the visual concepts used in the test (i.e., the representation) as well as the latent rules based on those concepts (i.e., the reasoning). However, learning of representation and reasoning is a challenging and ill-posed task,often approached in a stage-wise manner (first representation, then reasoning). In this work, we propose an end-to-end joint representation-reasoning learning framework, which leverages a weak form of inductive bias to improve both tasks together. Specifically, we propose a general generative graphical model for RPMs, GM-RPM, and apply it to solve the reasoning test. We accomplish this using a novel learning framework Disentangling based Abstract Reasoning Network (DAReN) based on the principles of GM-RPM. We perform an empirical evaluation of DAReN over several benchmark datasets. DAReN shows consistent improvement over state-of-the-art (SOTA) models on both the reasoning and the disentanglement tasks. This demonstrates the strong correlation between disentangled latent representation and the ability to solve abstract visual reasoning tasks.

Comprehension Based Question Answering using Bloom's Taxonomy

Current pre-trained language models have lots of knowledge, but a more limited ability to use that knowledge. Bloom's Taxonomy helps educators teach children how to use knowledge by categorizing comprehension skills, so we use it to analyze and improve the comprehension skills of large pre-trained language models. Our experiments focus on zero-shot question answering, using the taxonomy to provide proximal context that helps the model answer questions by being relevant to those questions. We show targeting context in this manner improves performance across 4 popular common sense question answer datasets.

Towards Solving Multimodal Comprehension

This paper targets the problem of procedural multimodal machine comprehension (M3C). This task requires an AI to comprehend given steps of multimodal instructions and then answer questions. Compared to vanilla machine comprehension tasks where an AI is required only to understand a textual input, procedural M3C is more challenging as the AI needs to comprehend both the temporal and causal factors along with multimodal inputs. Recently Yagcioglu et al. [35] introduced RecipeQA dataset to evaluate M3C. Our first contribution is the introduction of two new M3C datasets- WoodworkQA and DecorationQA with 16K and 10K instructional procedures, respectively. We then evaluate M3C using a textual cloze style question-answering task and highlight an inherent bias in the question answer generation method from [35] that enables a naive baseline to cheat by learning from only answer choices. This naive baseline performs similar to a popular method used in question answering- Impatient Reader [6] that uses attention over both the context and the query. We hypothesized that this naturally occurring bias present in the dataset affects even the best performing model. We verify our proposed hypothesis and propose an algorithm capable of modifying the given dataset to remove the bias elements. Finally, we report our performance on the debiased dataset with several strong baselines. We observe that the performance of all methods falls by a margin of 8% - 16% after correcting for the bias. We hope these datasets and the analysis will provide valuable benchmarks and encourage further research in this area.

Zero-Shot Learning with Knowledge Enhanced Visual Semantic Embeddings

We improve zero-shot learning (ZSL) by incorporating common-sense knowledge in DNNs. We propose Common-Sense based Neuro-Symbolic Loss (CSNL) that formulates prior knowledge as novel neuro-symbolic loss functions that regularize visual-semantic embedding. CSNL forces visual features in the VSE to obey common-sense rules relating to hypernyms and attributes. We introduce two key novelties for improved learning: (1) enforcement of rules for a group instead of a single concept to take into account class-wise relationships, and (2) confidence margins inside logical operators that enable implicit curriculum learning and prevent premature overfitting. We evaluate the advantages of incorporating each knowledge source and show consistent gains over prior state-of-art methods in both conventional and generalized ZSL e.g. 11.5%, 5.5%, and 11.6% improvements on AWA2, CUB, and Kinetics respectively.

Task-Discriminative Domain Alignment for Unsupervised Domain Adaptation

Domain Adaptation (DA), the process of effectively adapting task models learned on one domain, the source, to other related but distinct domains, the targets, with no or minimal retraining, is typically accomplished using the process of source-to-target manifold alignment. However, this process often leads to unsatisfactory adaptation performance, in part because it ignores the task-specific structure of the data. In this paper, we improve the performance of DA by introducing a discriminative discrepancy measure which takes advantage of auxiliary information available in the source and the target domains to better align the source and target distributions. Specifically, we leverage the cohesive clustering structure within individual data manifolds, associated with different tasks, to improve the alignment. This structure is explicit in the source, where the task labels are available, but is implicit in the target, making the problem challenging. We address the challenge by devising a deep DA framework, which combines a new task-driven domain alignment discriminator with domain regularizers that encourage the shared features as task-specific and domain invariant, and prompt the task model to be data structure preserving, guiding its decision boundaries through the low density data regions. We validate our framework on standard benchmarks, including Digits (MNIST, USPS, SVHN, MNIST-M), PACS, and VisDA. Our results show that our proposed model consistently outperforms the state-of-the-art in unsupervised domain adaptation.

Bayes-Factor-VAE: Hierarchical Bayesian Deep Auto-Encoder Models for Factor Disentanglement

We propose a family of novel hierarchical Bayesian deep auto-encoder models capable of identifying disentangled factors of variability in data. While many recent attempts at factor disentanglement have focused on sophisticated learning objectives within the VAE framework, their choice of a standard normal as the latent factor prior is both suboptimal and detrimental to performance. Our key observation is that the disentangled latent variables responsible for major sources of variability, the relevant factors, can be more appropriately modeled using long-tail distributions. The typical Gaussian priors are, on the other hand, better suited for modeling of nuisance factors. Motivated by this, we extend the VAE to a hierarchical Bayesian model by introducing hyper-priors on the variances of Gaussian latent priors, mimicking an infinite mixture, while maintaining tractable learning and inference of the traditional VAEs. This analysis signifies the importance of partitioning and treating in a different manner the latent dimensions corresponding to relevant factors and nuisances. Our proposed models, dubbed Bayes-Factor-VAEs, are shown to outperform existing methods both quantitatively and qualitatively in terms of latent disentanglement across several challenging benchmark tasks.