Post-training has emerged as a crucial paradigm for adapting large-scale pretrained models to various tasks, whose effects are fully reflected by delta parameters (i.e., the disparity between post-trained and pre-trained parameters). While numerous studies have explored delta parameter properties via operations like pruning, quantization, low-rank approximation, and extrapolation, a unified framework for systematically examining these characteristics has been lacking. In this paper, we propose a novel perspective based on Riemann sum approximation of the loss function to elucidate delta parameter editing operations. Our analysis categorizes existing methods into three classes based on their post-editing performance: competitive, decreased, and improved, explaining how they are expressed by the Riemann sum approximation term and how they alter the model performance. Extensive experiments on both visual and language models, including ViT, LLaMA 3, Qwen 2, and Mistral, corroborate our theoretical findings. Furthermore, we introduce extensions to existing techniques like DARE and BitDelta, highlighting their limitations in leveraging the properties of delta parameters and reorganizing them into general expressions to enhance the applicability and effectiveness of delta parameter editing in post-trained models. With the remarkable success of large-scale pre-trained models, post-training has emerged as the de facto standard paradigm for effective adaptations to various tasks (Han et al., 2024; Xin et al., 2024; Dodge et al., 2020; Zhao et al., 2023).

Traditional evaluation of information retrieval (IR) systems relies on human-annotated relevance labels, which can be both biased and costly at scale. In this context, large language models (LLMs) offer an alternative by allowing us to directly prompt them to assign relevance labels for passages associated with each query. In this study, we explore alternative methods to directly prompt LLMs for assigned relevance labels, by exploring two hypotheses: Hypothesis 1 assumes that it is helpful to break down "relevance" into specific criteria - exactness, coverage, topicality, and contextual fit. We explore different approaches that prompt large language models (LLMs) to obtain criteria-level grades for all passages, and we consider various ways to aggregate criteria-level grades into a relevance label. Hypothesis 2 assumes that differences in linguistic style between queries and passages may negatively impact the automatic relevance label prediction. We explore whether improvements can be achieved by first synthesizing a summary of the passage in the linguistic style of a query, and then using this summary in place of the passage to assess its relevance. We include an empirical evaluation of our approaches based on data from the LLMJudge challenge run in Summer 2024, where our "Four Prompts" approach obtained the highest scores in Kendall's tau.

Large Vision Language Models (LVLMs) such as LLaVA have demonstrated impressive capabilities as general-purpose chatbots that can engage in conversations about a provided input image. However, their responses are influenced by societal biases present in their training datasets, leading to undesirable differences in how the model responds when presented with images depicting people of different demographics. In this work, we propose a novel debiasing framework for LVLMs by directly ablating biased attributes during text generation to avoid generating text related to protected attributes, or even representing them internally. Our method requires no training and a relatively small amount of representative biased outputs ( 1000 samples). Our experiments show that not only can we can minimize the propensity of LVLMs to generate text related to protected attributes, but we can even use synthetic data to inform the ablation while retaining captioning performance on real data such as COCO. Furthermore, we find the resulting generations from a debiased LVLM exhibit similar accuracy as a baseline biased model, showing that debiasing effects can be achieved without sacrificing model performance.

Although many studies have investigated and reduced hallucinations in large language models (LLMs) for single-document tasks, research on hallucination in multi-document summarization (MDS) tasks remains largely unexplored. Specifically, it is unclear how the challenges arising from handling multiple documents (e.g., repetition and diversity of information) affect models outputs. In this work, we investigate how hallucinations manifest in LLMs when summarizing topic-specific information from multiple documents. Since no benchmarks exist for investigating hallucinations in MDS, we use existing news and conversation datasets, annotated with topic-specific insights, to create two novel multi-document benchmarks. When evaluating 5 LLMs on our benchmarks, we observe that on average, up to 75% of the content in LLM-generated summary is hallucinated, with hallucinations more likely to occur towards the end of the summaries. Moreover, when summarizing non-existent topic-related information, gpt-3.5-turbo and GPT-4o still generate summaries about 79.35% and 44% of the time, raising concerns about their tendency to fabricate content. To understand the characteristics of these hallucinations, we manually evaluate 700+ insights and find that most errors stem from either failing to follow instructions or producing overly generic insights. Motivated by these observations, we investigate the efficacy of simple post-hoc baselines in mitigating hallucinations but find them only moderately effective. Our results underscore the need for more effective approaches to systematically mitigate hallucinations in MDS. We release our dataset and code at github.com/megagonlabs/Hallucination_MDS.

Large language models (LLMs) achieve state-of-the-art performance on multiple language tasks, yet their safety guardrails can be circumvented, leading to harmful generations. In light of this, recent research on safety mechanisms has emerged, revealing that when safety representations or component are suppressed, the safety capability of LLMs are compromised. However, existing research tends to overlook the safety impact of multi-head attention mechanisms, despite their crucial role in various model functionalities. Hence, in this paper, we aim to explore the connection between standard attention mechanisms and safety capability to fill this gap in the safety-related mechanistic interpretability. We propose a novel metric which tailored for multi-head attention, the Safety Head ImPortant Score (Ships), to assess the individual heads' contributions to model safety. Based on this, we generalize Ships to the dataset level and further introduce the Safety Attention Head AttRibution Algorithm (Sahara) to attribute the critical safety attention heads inside the model. Our findings show that the special attention head has a significant impact on safety. Ablating a single safety head allows aligned model (e.g., Llama-2-7b-chat) to respond to 16 times more harmful queries, while only modifying 0.006% of the parameters, in contrast to the ~ 5% modification required in previous studies. More importantly, we demonstrate that attention heads primarily function as feature extractors for safety and models fine-tuned from the same base model exhibit overlapping safety heads through comprehensive experiments. Together, our attribution approach and findings provide a novel perspective for unpacking the black box of safety mechanisms within large models.

Retrieval-Augmented Generation (RAG) has proven its effectiveness in mitigating hallucinations in Large Language Models (LLMs) by retrieving knowledge from external resources. To adapt LLMs for RAG pipelines, current approaches use instruction tuning to optimize LLMs, improving their ability to utilize retrieved knowledge. This supervised fine-tuning (SFT) approach focuses on equipping LLMs to handle diverse RAG tasks using different instructions. However, it trains RAG modules to overfit training signals and overlooks the varying data preferences among agents within the RAG system. In this paper, we propose a Differentiable Data Rewards (DDR) method, which end-to-end trains RAG systems by aligning data preferences between different RAG modules. DDR works by collecting the rewards to optimize each agent with a rollout method. This method prompts agents to sample some potential responses as perturbations, evaluates the impact of these perturbations on the whole RAG system, and subsequently optimizes the agent to produce outputs that improve the performance of the RAG system. Our experiments on various knowledge-intensive tasks demonstrate that DDR significantly outperforms the SFT method, particularly for LLMs with smaller-scale parameters that depend more on the retrieved knowledge. Additionally, DDR exhibits a stronger capability to align the data preference between RAG modules. The DDR method makes generation module more effective in extracting key information from documents and mitigating conflicts between parametric memory and external knowledge. All codes are available at https://github.com/OpenMatch/RAG-DDR.

The prolific use of Large Language Models (LLMs) as an alternate knowledge base requires them to be factually consistent, necessitating both correctness and consistency traits for paraphrased queries. Recently, significant attempts have been made to benchmark datasets and metrics to evaluate LLMs for these traits. However, structural simplicity (subject-relation-object) and contemporary association in their query formulation limit the broader definition of factuality and consistency. In this study, we introduce TeCFaP, a novel Temporally Consistent Factuality Probe task to expand the consistent factuality probe in the temporal dimension. To this end, we propose TEMP-COFAC, a high-quality dataset of prefix-style English query paraphrases. Subsequently, we extend the definitions of existing metrics to represent consistent factuality across temporal dimension. We experiment with a diverse set of LLMs and find most of them performing poorly on TeCFaP. Next, we propose a novel solution CoTSeLF (Consistent-Time-Sensitive Learning Framework) combining multi-task instruction tuning (MT-IT) with consistent-time-sensitive reinforcement learning (CTSRL) to improve temporally consistent factuality in LLMs. Our experiments demonstrate the efficacy of CoTSeLF over several baselines.

This paper investigates the application of translation alignment algorithms in the creation of a Multilingual Digital Edition (MDE) of Alessandro Manzoni's Italian novel "I promessi sposi" ("The Betrothed"), with translations in eight languages (English, Spanish, French, German, Dutch, Polish, Russian and Chinese) from the 19th and 20th centuries. We identify key requirements for the MDE to improve both the reader experience and support for translation studies. Our research highlights the limitations of current state-of-the-art algorithms when applied to the translation of literary texts and outlines an automated pipeline for MDE creation. This pipeline transforms raw texts into web-based, side-by-side representations of original and translated texts with different rendering options. In addition, we propose new metrics for evaluating the alignment of literary translations and suggest visualization techniques for future analysis.

Fact-checking pipelines increasingly adopt the Decompose-Then-Verify paradigm, where texts are broken down into smaller claims for individual verification and subsequently combined for a veracity decision. While decomposition is widely-adopted in such pipelines, its effects on final fact-checking performance remain underexplored. Some studies have reported improvements from decompostition, while others have observed performance declines, indicating its inconsistent impact. To date, no comprehensive analysis has been conducted to understand this variability. To address this gap, we present an in-depth analysis that explicitly examines the impact of decomposition on downstream verification performance. Through error case inspection and experiments, we introduce a categorization of decomposition errors and reveal a trade-off between accuracy gains and the noise introduced through decomposition. Our analysis provides new insights into understanding current system's instability and offers guidance for future studies toward improving claim decomposition in fact-checking pipelines.

Predicting future human behavior is an increasingly popular topic in computer vision, driven by the interest in applications such as autonomous vehicles, digital assistants and human-robot interactions. The literature on behavior prediction spans various tasks, including action anticipation, activity forecasting, intent prediction, goal prediction, and so on. Our survey aims to tie together this fragmented literature, covering recent technical innovations as well as the development of new large-scale datasets for model training and evaluation. We also summarize the widely-used metrics for different tasks and provide a comprehensive performance comparison of existing approaches on eleven action anticipation datasets. This survey serves as not only a reference for contemporary methodologies in action anticipation, but also a guideline for future research direction of this evolving landscape.