Devils in Middle Layers of Large Vision-Language Models: Interpreting, Detecting and Mitigating Object Hallucinations via Attention Lens

Hallucinations in Large Vision-Language Models (LVLMs) significantly undermine their reliability, motivating researchers to explore the causes of hallucination. However, most studies primarily focus on the language aspect rather than the visual. In this paper, we address how LVLMs process visual information and whether this process causes hallucination. Firstly, we use the attention lens to identify the stages at which LVLMs handle visual data, discovering that the middle layers are crucial. Moreover, we find that these layers can be further divided into two stages: "visual information enrichment" and "semantic refinement" which respectively propagate visual data to object tokens and interpret it through text. By analyzing attention patterns during the visual information enrichment stage, we find that real tokens consistently receive higher attention weights than hallucinated ones, serving as a strong indicator of hallucination. Further examination of multi-head attention maps reveals that hallucination tokens often result from heads interacting with inconsistent objects. Based on these insights, we propose a simple inference-time method that adjusts visual attention by integrating information across various heads. Extensive experiments demonstrate that this approach effectively mitigates hallucinations in mainstream LVLMs without additional training costs.

Look Twice Before You Answer: Memory-Space Visual Retracing for Hallucination Mitigation in Multimodal Large Language Models

Despite their impressive capabilities, Multimodal Large Language Models (MLLMs) are susceptible to hallucinations, especially assertively fabricating content not present in the visual inputs. To address the aforementioned challenge, we follow a common cognitive process - when one's initial memory of critical on-sight details fades, it is intuitive to look at them a second time to seek a factual and accurate answer. Therefore, we introduce Memory-space Visual Retracing (MemVR), a novel hallucination mitigation paradigm that without the need for external knowledge retrieval or additional fine-tuning. In particular, we treat visual prompts as supplementary evidence to be reinjected into MLLMs via Feed Forward Network (FFN) as key-value memory, when the model is uncertain or even amnesic about question-relevant visual memories. Comprehensive experimental evaluations demonstrate that MemVR significantly mitigates hallucination issues across various MLLMs and excels in general benchmarks without incurring added time overhead, thus emphasizing its potential for widespread applicability.

Reefknot: A Comprehensive Benchmark for Relation Hallucination Evaluation, Analysis and Mitigation in Multimodal Large Language Models

Hallucination issues persistently plagued current multimodal large language models (MLLMs). While existing research primarily focuses on object-level or attribute-level hallucinations, sidelining the more sophisticated relation hallucinations that necessitate advanced reasoning abilities from MLLMs. Besides, recent benchmarks regarding relation hallucinations lack in-depth evaluation and effective mitigation. Moreover, their datasets are typically derived from a systematic annotation process, which could introduce inherent biases due to the predefined process. To handle the aforementioned challenges, we introduce Reefknot, a comprehensive benchmark specifically targeting relation hallucinations, consisting of over 20,000 samples derived from real-world scenarios. Specifically, we first provide a systematic definition of relation hallucinations, integrating perspectives from perceptive and cognitive domains. Furthermore, we construct the relation-based corpus utilizing the representative scene graph dataset Visual Genome (VG), from which semantic triplets follow real-world distributions. Our comparative evaluation across three distinct tasks revealed a substantial shortcoming in the capabilities of current MLLMs to mitigate relation hallucinations. Finally, we advance a novel confidence-based mitigation strategy tailored to tackle the relation hallucinations problem. Across three datasets, including Reefknot, we observed an average reduction of 9.75% in the hallucination rate. We believe our paper sheds valuable insights into achieving trustworthy multimodal intelligence. Our dataset and code will be released upon paper acceptance.

NLPerturbator: Studying the Robustness of Code LLMs to Natural Language Variations

Large language models (LLMs) achieve promising results in code generation based on a given natural language description. They have been integrated into open-source projects and commercial products to facilitate daily coding activities. The natural language description in the prompt is crucial for LLMs to comprehend users' requirements. Prior studies uncover that LLMs are sensitive to the changes in the prompts, including slight changes that look inconspicuous. However, the natural language descriptions often vary in real-world scenarios (e.g., different formats, grammar, and wording). Prior studies on the robustness of LLMs are often based on random perturbations and such perturbations may not actually happen. In this paper, we conduct a comprehensive study to investigate how are code LLMs robust to variations of natural language description in real-world scenarios. We summarize 18 categories of perturbations of natural language and 3 combinations of co-occurred categories based on our literature review and an online survey with practitioners. We propose an automated framework, NLPerturbator, which can perform perturbations of each category given a set of prompts. Through a series of experiments on code generation using six code LLMs, we find that the perturbed prompts can decrease the performance of code generation by a considerable margin (e.g., up to 21.2%, and 4.8% to 6.1% on average). Our study highlights the importance of enhancing the robustness of LLMs to real-world variations in the prompts, as well as the essentiality of attentively constructing the prompts.

Evaluating Large Language Models with Runtime Behavior of Program Execution

Large language models for code (i.e., code LLMs) have shown strong code understanding and generation capabilities. To evaluate the capabilities of code LLMs in various aspects, many benchmarks have been proposed (e.g., HumanEval and ClassEval). Code reasoning is one of the most essential abilities of code LLMs, but existing benchmarks for code reasoning are not sufficient. Typically, they focus on predicting the input and output of a program, ignoring the evaluation of the intermediate behavior during program execution, as well as the logical consistency (e.g., the model should not give the correct output if the prediction of execution path is wrong) when performing the reasoning. To address these problems, in this paper, we propose a framework, namely REval, for evaluating code reasoning abilities and consistency of code LLMs with program execution. We utilize existing code benchmarks and adapt them to new benchmarks within our framework. A large-scale empirical study is conducted and most LLMs show unsatisfactory performance on both Runtime Behavior Reasoning (i.e., an average accuracy of 44.4%) and Incremental Consistency Evaluation (i.e., an average IC score of 10.3). Evaluation results of current code LLMs reflect the urgent need for the community to strengthen the code reasoning capability of code LLMs.