The various limitations of Generative AI, such as hallucinations and model failures, have made it crucial to understand the role of different modalities in Visual Language Model (VLM) predictions. Our work investigates how the integration of information from image and text modalities influences the performance and behavior of VLMs in visual question answering (VQA) and reasoning tasks. We measure this effect through answer accuracy, reasoning quality, model uncertainty, and modality relevance. We study the interplay between text and image modalities in different configurations where visual content is essential for solving the VQA task. Our contributions include (1) the Semantic Interventions (SI)-VQA dataset, (2) a benchmark study of various VLM architectures under different modality configurations, and (3) the Interactive Semantic Interventions (ISI) tool. The SI-VQA dataset serves as the foundation for the benchmark, while the ISI tool provides an interface to test and apply semantic interventions in image and text inputs, enabling more fine-grained analysis. Our results show that complementary information between modalities improves answer and reasoning quality, while contradictory information harms model performance and confidence. Image text annotations have minimal impact on accuracy and uncertainty, slightly increasing image relevance. Attention analysis confirms the dominant role of image inputs over text in VQA tasks. In this study, we evaluate state-of-the-art VLMs that allow us to extract attention coefficients for each modality. A key finding is PaliGemma's harmful overconfidence, which poses a higher risk of silent failures compared to the LLaVA models. This work sets the foundation for rigorous analysis of modality integration, supported by datasets specifically designed for this purpose.

Large language models (LLMs) are increasingly applied to complex reasoning tasks that require executing several complex steps before receiving any reward. Properly assigning credit to these steps is essential for enhancing model performance. Proximal Policy Optimization (PPO), a state-of-the-art reinforcement learning (RL) algorithm used for LLM finetuning, employs value networks to tackle credit assignment. However, value networks face challenges in predicting the expected cumulative rewards accurately in complex reasoning tasks, often leading to high-variance updates and suboptimal performance. In this work, we systematically evaluate the efficacy of value networks and reveal their significant shortcomings in reasoning-heavy LLM tasks, showing that they barely outperform a random baseline when comparing alternative steps. To address this, we propose VinePPO, a straightforward approach that leverages the flexibility of language environments to compute unbiased Monte Carlo-based estimates, bypassing the need for large value networks. Our method consistently outperforms PPO and other RL-free baselines across MATH and GSM8K datasets with fewer gradient updates (up to 9x), less wall-clock time (up to 3.0x). These results emphasize the importance of accurate credit assignment in RL finetuning of LLM and demonstrate VinePPO's potential as a superior alternative.

Differentiating between generated and human-written content is important for navigating the modern world. Large language models (LLMs) are crucial drivers behind the increased quality of computer-generated content. Reportedly, humans find it increasingly difficult to identify whether an AI model generated a piece of text. Our work tests how two important factors contribute to the human vs AI race: empathy and an incentive to appear human. We address both aspects in two experiments: human participants and a state-of-the-art LLM wrote relationship advice (Study 1, n=530) or mere descriptions (Study 2, n=610), either instructed to be as human as possible or not. New samples of humans (n=428 and n=408) then judged the texts' source. Our findings show that when empathy is required, humans excel. Contrary to expectations, instructions to appear human were only effective for the LLM, so the human advantage diminished. Computational text analysis revealed that LLMs become more human because they may have an implicit representation of what makes a text human and effortlessly apply these heuristics. The model resorts to a conversational, self-referential, informal tone with a simpler vocabulary to mimic stochastic empathy. We discuss these findings in light of recent claims on the on-par performance of LLMs.

Large language models (LLMs) have shown remarkable capabilities in natural language processing; however, they still face difficulties when tasked with understanding lengthy contexts and executing effective question answering. These challenges often arise due to the complexity and ambiguity present in longer texts. To enhance the performance of LLMs in such scenarios, we introduce the Long Question Coreference Adaptation (LQCA) method. This innovative framework focuses on coreference resolution tailored to long contexts, allowing the model to identify and manage references effectively. The LQCA method encompasses four key steps: resolving coreferences within sub-documents, computing the distances between mentions, defining a representative mention for coreference, and answering questions through mention replacement. By processing information systematically, the framework provides easier-to-handle partitions for LLMs, promoting better understanding. Experimental evaluations on a range of LLMs and datasets have yielded positive results, with a notable improvements on OpenAI-o1-mini and GPT-4o models, highlighting the effectiveness of leveraging coreference resolution to bridge context gaps in question answering.

Recently, retrieval-based language models (RLMs) have received much attention. However, most of them leverage a pre-trained retriever with fixed parameters, which may not adapt well to causal language models. In this work, we propose Grouped Cross-Attention, a novel module enabling joint pre-training of the retriever and causal LM, and apply it to long-context modeling. For a given input sequence, we split it into chunks and use the current chunk to retrieve past chunks for subsequent text generation. Our innovation allows the retriever to learn how to retrieve past chunks that better minimize the auto-regressive loss of subsequent tokens in an end-to-end manner. By integrating top-$k$ retrieval, our model can be pre-trained efficiently from scratch with context lengths up to 64K tokens. Our experiments show our model, compared with long-range LM baselines, can achieve lower perplexity with comparable or lower pre-training and inference costs.

One of the prominent issues stifling the current generation of large language models is their limited context length. Recent proprietary models such as GPT-4 and Claude 2 have introduced longer context lengths, 8k/32k and 100k, respectively; however, despite the efforts in the community, most common models, such as LLama-2, have a context length of 4k or less. Unlimiformer (Bertsch et al., 2023) is a recently popular vector-retrieval augmentation method that offloads cross-attention computations to a kNN index. However, its main limitation is incompatibility with decoder-only transformers out of the box. In this work, we explore practical considerations of adapting Unlimiformer to decoder-only transformers and introduce a series of modifications to overcome this limitation. Moreover, we expand the original experimental setup on summarization to include a new task (i.e., free-form Q&A) and an instruction-tuned model (i.e., a custom 6.7B GPT model). Our results showcase the effectiveness of these modifications on summarization, performing on par with a model with 2x the context length. Moreover, we discuss limitations and future directions for free-form Q&A and instruction-tuned models.

Intent detection is a critical component of task-oriented dialogue systems (TODS) which enables the identification of suitable actions to address user utterances at each dialog turn. Traditional approaches relied on computationally efficient supervised sentence transformer encoder models, which require substantial training data and struggle with out-of-scope (OOS) detection. The emergence of generative large language models (LLMs) with intrinsic world knowledge presents new opportunities to address these challenges. In this work, we adapt 7 SOTA LLMs using adaptive in-context learning and chain-of-thought prompting for intent detection, and compare their performance with contrastively fine-tuned sentence transformer (SetFit) models to highlight prediction quality and latency tradeoff. We propose a hybrid system using uncertainty based routing strategy to combine the two approaches that along with negative data augmentation results in achieving the best of both worlds ( i.e. within 2% of native LLM accuracy with 50% less latency). To better understand LLM OOS detection capabilities, we perform controlled experiments revealing that this capability is significantly influenced by the scope of intent labels and the size of the label space. We also introduce a two-step approach utilizing internal LLM representations, demonstrating empirical gains in OOS detection accuracy and F1-score by >5% for the Mistral-7B model.

Mixture-of-Experts (MoE) shines brightly in large language models (LLMs) and demonstrates outstanding performance in plentiful natural language processing tasks. However, existing methods transforming LLMs from dense to MoE face significant data requirements and typically rely on large-scale post-training. In this paper, we propose Upcycling Instruction Tuning (UpIT), a data-efficient approach for tuning a dense pre-trained model into a MoE instruction model. Specifically, we first point out that intermediate checkpoints during instruction tuning of the dense model are naturally suitable for specialized experts, and then propose an expert expansion stage to flexibly achieve models with flexible numbers of experts, where genetic algorithm and parameter merging are introduced to ensure sufficient diversity of new extended experts. To ensure that each specialized expert in the MoE model works as expected, we select a small amount of seed data that each expert excels to pre-optimize the router. Extensive experiments with various data scales and upcycling settings demonstrate the outstanding performance and data efficiency of UpIT, as well as stable improvement in expert or data scaling. Further analysis reveals the importance of ensuring expert diversity in upcycling. Large Language Models (LLMs) have demonstrated remarkable performance on various NLP tasks and are gradually becoming part of our daily lives through chatbot applications such as ChatGPT, Copilot, etc (Ouyang et al., 2022; Touvron et al., 2023; OpenAI, 2024). As LLMs become increasingly prevalent, the high computational of traditional dense architecture with high computational costs in the inference phase poses significant obstacles to downstream deployment. How to improve the model performance without proportionally increasing computing resources become a hot topic in the field (Muennighoff et al., 2024; Xue et al., 2024).

Red teaming assesses how large language models (LLMs) can produce content that violates norms, policies, and rules set during their safety training. However, most existing automated methods in the literature are not representative of the way humans tend to interact with AI models. Common users of AI models may not have advanced knowledge of adversarial machine learning methods or access to model internals, and they do not spend a lot of time crafting a single highly effective adversarial prompt. Instead, they are likely to make use of techniques commonly shared online and exploit the multiturn conversational nature of LLMs. While manual testing addresses this gap, it is an inefficient and often expensive process. To address these limitations, we introduce the Generative Offensive Agent Tester (GOAT), an automated agentic red teaming system that simulates plain language adversarial conversations while leveraging multiple adversarial prompting techniques to identify vulnerabilities in LLMs. We instantiate GOAT with 7 red teaming attacks by prompting a general-purpose model in a way that encourages reasoning through the choices of methods available, the current target model's response, and the next steps. Our approach is designed to be extensible and efficient, allowing human testers to focus on exploring new areas of risk while automation covers the scaled adversarial stress-testing of known risk territory. We present the design and evaluation of GOAT, demonstrating its effectiveness in identifying vulnerabilities in state-of-the-art LLMs, with an ASR@10 of 97% against Llama 3.1 and 88% against GPT-4 on the JailbreakBench dataset.

In Query-driven Travel Recommender Systems (RSs), it is crucial to understand the user intent behind challenging natural language (NL) destination queries such as the broadly worded "youth-friendly activities" or the indirect description "a high school graduation trip". Such queries are challenging due to the wide scope and subtlety of potential user intents that confound the ability of retrieval methods to infer relevant destinations from available textual descriptions such as WikiVoyage. While query reformulation (QR) has proven effective in enhancing retrieval by addressing user intent, existing QR methods tend to focus only on expanding the range of potentially matching query subtopics (breadth) or elaborating on the potential meaning of a query (depth), but not both. In this paper, we introduce Elaborative Subtopic Query Reformulation (EQR), a large language model-based QR method that combines both breadth and depth by generating potential query subtopics with information-rich elaborations. We also release TravelDest, a novel dataset for query-driven travel destination RSs. Experiments on TravelDest show that EQR achieves significant improvements in recall and precision over existing state-of-the-art QR methods.