Research on Large Language Models (LLMs) has often neglected subtle biases that, although less apparent, can significantly influence the models' outputs toward particular social narratives. This study addresses two such biases within LLMs: representative bias, which denotes a tendency of LLMs to generate outputs that mirror the experiences of certain identity groups, and affinity bias, reflecting the models' evaluative preferences for specific narratives or viewpoints. We introduce two novel metrics to measure these biases: the Representative Bias Score (RBS) and the Affinity Bias Score (ABS), and present the Creativity-Oriented Generation Suite (CoGS), a collection of open-ended tasks such as short story writing and poetry composition, designed with customized rubrics to detect these subtle biases. Our analysis uncovers marked representative biases in prominent LLMs, with a preference for identities associated with being white, straight, and men. Furthermore, our investigation of affinity bias reveals distinctive evaluative patterns within each model, akin to `bias fingerprints'. This trend is also seen in human evaluators, highlighting a complex interplay between human and machine bias perceptions.
相關內容
Large language models (LLMs) often exhibit undesirable behaviors, such as hallucinations and sequence repetitions. We propose to view these behaviors as fallbacks that models exhibit under uncertainty, and investigate the connection between them. We categorize fallback behaviors -- sequence repetitions, degenerate text, and hallucinations -- and extensively analyze them in models from the same family that differ by the amount of pretraining tokens, parameter count, or the inclusion of instruction-following training. Our experiments reveal a clear and consistent ordering of fallback behaviors, across all these axes: the more advanced an LLM is (i.e., trained on more tokens, has more parameters, or instruction-tuned), its fallback behavior shifts from sequence repetitions, to degenerate text, and then to hallucinations. Moreover, the same ordering is observed throughout a single generation, even for the best-performing models; as uncertainty increases, models shift from generating hallucinations to producing degenerate text and then sequence repetitions. Lastly, we demonstrate that while common decoding techniques, such as random sampling, might alleviate some unwanted behaviors like sequence repetitions, they increase harder-to-detect hallucinations.
Wireless Capsule Endoscopy (WCE) is highly valued for its non-invasive and painless approach, though its effectiveness is compromised by uneven illumination from hardware constraints and complex internal dynamics, leading to overexposed or underexposed images. While researchers have discussed the challenges of low-light enhancement in WCE, the issue of correcting for different exposure levels remains underexplored. To tackle this, we introduce EndoUIC, a WCE unified illumination correction solution using an end-to-end promptable diffusion transformer (DiT) model. In our work, the illumination prompt module shall navigate the model to adapt to different exposure levels and perform targeted image enhancement, in which the Adaptive Prompt Integration (API) and Global Prompt Scanner (GPS) modules shall further boost the concurrent representation learning between the prompt parameters and features. Besides, the U-shaped restoration DiT model shall capture the long-range dependencies and contextual information for unified illumination restoration. Moreover, we present a novel Capsule-endoscopy Exposure Correction (CEC) dataset, including ground-truth and corrupted image pairs annotated by expert photographers. Extensive experiments against a variety of state-of-the-art (SOTA) methods on four datasets showcase the effectiveness of our proposed method and components in WCE illumination restoration, and the additional downstream experiments further demonstrate its utility for clinical diagnosis and surgical assistance.
Ensuring that AI systems reliably and robustly avoid harmful or dangerous behaviours is a crucial challenge, especially for AI systems with a high degree of autonomy and general intelligence, or systems used in safety-critical contexts. In this paper, we will introduce and define a family of approaches to AI safety, which we will refer to as guaranteed safe (GS) AI. The core feature of these approaches is that they aim to produce AI systems which are equipped with high-assurance quantitative safety guarantees. This is achieved by the interplay of three core components: a world model (which provides a mathematical description of how the AI system affects the outside world), a safety specification (which is a mathematical description of what effects are acceptable), and a verifier (which provides an auditable proof certificate that the AI satisfies the safety specification relative to the world model). We outline a number of approaches for creating each of these three core components, describe the main technical challenges, and suggest a number of potential solutions to them. We also argue for the necessity of this approach to AI safety, and for the inadequacy of the main alternative approaches.
Unequal representation across cultures and socioeconomic groups in AI is a significant and challenging problem, often leading to uneven model performance. As a step toward addressing this issue, we formulate translated non-English, geographic, and socioeconomic integrated prompts and evaluate their impact on VL model performance for data from different countries and income groups. Our findings show that geographic and socioeconomic integrated prompts improve VL performance on lower-income data and favor the retrieval of topic appearances commonly found in data from low-income households. From our analyses, we identify and highlight contexts where these strategies yield the most improvements. Our model analysis code is publicly available at //github.com/Anniejoan/Uplifting-Lower-income-data .
ASRRL-TTS: Agile Speaker Representation Reinforcement Learning for Text-to-Speech Speaker Adaptation
Speaker adaptation, which involves cloning voices from unseen speakers in the Text-to-Speech task, has garnered significant interest due to its numerous applications in multi-media fields. Despite recent advancements, existing methods often struggle with inadequate speaker representation accuracy and overfitting, particularly in limited reference speeches scenarios. To address these challenges, we propose an Agile Speaker Representation Reinforcement Learning strategy to enhance speaker similarity in speaker adaptation tasks. ASRRL is the first work to apply reinforcement learning to improve the modeling accuracy of speaker embeddings in speaker adaptation, addressing the challenge of decoupling voice content and timbre. Our approach introduces two action strategies tailored to different reference speeches scenarios. In the single-sentence scenario, a knowledge-oriented optimal routine searching RL method is employed to expedite the exploration and retrieval of refinement information on the fringe of speaker representations. In the few-sentence scenario, we utilize a dynamic RL method to adaptively fuse reference speeches, enhancing the robustness and accuracy of speaker modeling. To achieve optimal results in the target domain, a multi-scale fusion scoring mechanism based reward model that evaluates speaker similarity, speech quality, and intelligibility across three dimensions is proposed, ensuring that improvements in speaker similarity do not compromise speech quality or intelligibility. The experimental results on the LibriTTS and VCTK datasets within mainstream TTS frameworks demonstrate the extensibility and generalization capabilities of the proposed ASRRL method. The results indicate that the ASRRL method significantly outperforms traditional fine-tuning approaches, achieving higher speaker similarity and better overall speech quality with limited reference speeches.
The recent emergence of Medical Large Vision Language Models (Med-LVLMs) has enhanced medical diagnosis. However, current Med-LVLMs frequently encounter factual issues, often generating responses that do not align with established medical facts. Retrieval-Augmented Generation (RAG), which utilizes external knowledge, can improve the factual accuracy of these models but introduces two major challenges. First, limited retrieved contexts might not cover all necessary information, while excessive retrieval can introduce irrelevant and inaccurate references, interfering with the model's generation. Second, in cases where the model originally responds correctly, applying RAG can lead to an over-reliance on retrieved contexts, resulting in incorrect answers. To address these issues, we propose RULE, which consists of two components. First, we introduce a provably effective strategy for controlling factuality risk through the calibrated selection of the number of retrieved contexts. Second, based on samples where over-reliance on retrieved contexts led to errors, we curate a preference dataset to fine-tune the model, balancing its dependence on inherent knowledge and retrieved contexts for generation. We demonstrate the effectiveness of RULE on three medical VQA datasets, achieving an average improvement of 20.8% in factual accuracy. We publicly release our benchmark and code in //github.com/richard-peng-xia/RULE.
We propose Waymo Open Motion Dataset-Reasoning (WOMD-Reasoning), a language annotation dataset built on WOMD, with a focus on describing and reasoning interactions and intentions in driving scenarios. Previous language datasets primarily captured interactions caused by close distances. However, interactions induced by traffic rules and human intentions, which can occur over long distances, are yet sufficiently covered, despite being very common and more challenging for prediction or planning models to understand. Therefore, our WOMD-Reasoning focuses extensively on these interactions, providing a total of 409k Q&As for varying types of interactions. Additionally, WOMD-Reasoning presents by far the largest Q&A dataset on real-world driving scenarios, with around 3 million Q&As covering various topics of autonomous driving from map descriptions, motion status descriptions, to narratives and analyses of agents' interactions, behaviors, and intentions. This extensive textual information enables fine-tuning driving-related Large Language Models (LLMs) for a wide range of applications like scene description, prediction, planning, etc. By incorporating interaction and intention language from WOMD-Reasoning, we see significant enhancements in the performance of the state-of-the-art trajectory prediction model, Multipath++, with improvements of 10.14% in $MR_6$ and 6.90% in $minFDE_6$, proving the effectiveness of WOMD-Reasoning. We hope WOMD-Reasoning would empower LLMs in driving to offer better interaction understanding and behavioral reasoning. The dataset is available on //waymo.com/open/download .
Group Equivariant CNNs (G-CNNs) have shown promising efficacy in various tasks, owing to their ability to capture hierarchical features in an equivariant manner. However, their equivariance is fixed to the symmetry of the whole group, limiting adaptability to diverse partial symmetries in real-world datasets, such as limited rotation symmetry of handwritten digit images and limited color-shift symmetry of flower images. Recent efforts address this limitation, one example being Partial G-CNN which restricts the output group space of convolution layers to break full equivariance. However, such an approach still fails to adjust equivariance levels across data. In this paper, we propose a novel approach, Variational Partial G-CNN (VP G-CNN), to capture varying levels of partial equivariance specific to each data instance. VP G-CNN redesigns the distribution of the output group elements to be conditioned on input data, leveraging variational inference to avoid overfitting. This enables the model to adjust its equivariance levels according to the needs of individual data points. Additionally, we address training instability inherent in discrete group equivariance models by redesigning the reparametrizable distribution. We demonstrate the effectiveness of VP G-CNN on both toy and real-world datasets, including MNIST67-180, CIFAR10, ColorMNIST, and Flowers102. Our results show robust performance, even in uncertainty metrics.
Commonsense reasoning is a difficult task for a computer, but a critical skill for an artificial intelligence (AI). It can enhance the explainability of AI models by enabling them to provide intuitive and human-like explanations for their decisions. This is necessary in many areas especially in question answering (QA), which is one of the most important tasks of natural language processing (NLP). Over time, a multitude of methods have emerged for solving commonsense reasoning problems such as knowledge-based approaches using formal logic or linguistic analysis. In this paper, we investigate the effectiveness of large language models (LLMs) on different QA tasks with a focus on their abilities in reasoning and explainability. We study three LLMs: GPT-3.5, Gemma and Llama 3. We further evaluate the LLM results by means of a questionnaire. We demonstrate the ability of LLMs to reason with commonsense as the models outperform humans on different datasets. While GPT-3.5's accuracy ranges from 56% to 93% on various QA benchmarks, Llama 3 achieved a mean accuracy of 90% on all eleven datasets. Thereby Llama 3 is outperforming humans on all datasets with an average 21% higher accuracy over ten datasets. Furthermore, we can appraise that, in the sense of explainable artificial intelligence (XAI), GPT-3.5 provides good explanations for its decisions. Our questionnaire revealed that 66% of participants rated GPT-3.5's explanations as either "good" or "excellent". Taken together, these findings enrich our understanding of current LLMs and pave the way for future investigations of reasoning and explainability.
Recently, Mutual Information (MI) has attracted attention in bounding the generalization error of Deep Neural Networks (DNNs). However, it is intractable to accurately estimate the MI in DNNs, thus most previous works have to relax the MI bound, which in turn weakens the information theoretic explanation for generalization. To address the limitation, this paper introduces a probabilistic representation of DNNs for accurately estimating the MI. Leveraging the proposed MI estimator, we validate the information theoretic explanation for generalization, and derive a tighter generalization bound than the state-of-the-art relaxations.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191