Concept erasure in language models has traditionally lacked a comprehensive evaluation framework, leading to incomplete assessments of effectiveness of erasure methods. We propose an evaluation paradigm centered on three critical criteria: innocence (complete knowledge removal), seamlessness (maintaining conditional fluent generation), and specificity (preserving unrelated task performance). Our evaluation metrics naturally motivate the development of Erasure of Language Memory (ELM), a new method designed to address all three dimensions. ELM employs targeted low-rank updates to alter output distributions for erased concepts while preserving overall model capabilities including fluency when prompted for an erased concept. We demonstrate ELM's efficacy on biosecurity, cybersecurity, and literary domain erasure tasks. Comparative analysis shows that ELM achieves superior performance across our proposed metrics, including near-random scores on erased topic assessments, generation fluency, maintained accuracy on unrelated benchmarks, and robustness under adversarial attacks. Our code, data, and trained models are available at https://elm.baulab.info

Contrastive Language-Image Pre-training (CLIP) has been a celebrated method for training vision encoders to generate image/text representations facilitating various applications. Recently, CLIP has been widely adopted as the vision backbone of multimodal large language models (MLLMs) to connect image inputs for language interactions. The success of CLIP as a vision-language foundation model relies on aligning web-crawled noisy text annotations at image levels. Nevertheless, such criteria may become insufficient for downstream tasks in need of fine-grained vision representations, especially when region-level understanding is demanding for MLLMs. In this paper, we improve the localization capability of CLIP with several advances. We propose a pre-training method called Contrastive Localized Language-Image Pre-training (CLOC) by complementing CLIP with region-text contrastive loss and modules. We formulate a new concept, promptable embeddings, of which the encoder produces image embeddings easy to transform into region representations given spatial hints. To support large-scale pre-training, we design a visually-enriched and spatially-localized captioning framework to effectively generate region-text pseudo-labels at scale. By scaling up to billions of annotated images, CLOC enables high-quality regional embeddings for image region recognition and retrieval tasks, and can be a drop-in replacement of CLIP to enhance MLLMs, especially on referring and grounding tasks.

Recent advancements in multimodal models highlight the value of rewritten captions for improving performance, yet key challenges remain. For example, while synthetic captions often provide superior quality and image-text alignment, it is not clear whether they can fully replace AltTexts: the role of synthetic captions and their interaction with original web-crawled AltTexts in pre-training is still not well understood. Moreover, different multimodal foundation models may have unique preferences for specific caption formats, but efforts to identify the optimal captions for each model remain limited. In this work, we propose a novel, controllable, and scalable captioning pipeline designed to generate diverse caption formats tailored to various multimodal models. By examining Short Synthetic Captions (SSC) towards Dense Synthetic Captions (DSC+) as case studies, we systematically explore their effects and interactions with AltTexts across models such as CLIP, multimodal LLMs, and diffusion models. Our findings reveal that a hybrid approach that keeps both synthetic captions and AltTexts can outperform the use of synthetic captions alone, improving both alignment and performance, with each model demonstrating preferences for particular caption formats. This comprehensive analysis provides valuable insights into optimizing captioning strategies, thereby advancing the pre-training of multimodal foundation models.

Inference-time computation is a powerful paradigm to enhance the performance of large language models (LLMs), with Best-of-N sampling being a widely used technique. However, this method is computationally expensive, requiring both (1) an external reward model and (2) the generation of multiple samples. In this work, we introduce a new generative self-evaluation scheme designed to adaptively reduce the number of generated samples while maintaining or even improving performance. We use a generative reward model formulation, allowing the LLM to predict mid-generation the probability that restarting the generation will yield a better response. These predictions are obtained without an external reward model and can be used to decide whether or not to generate more samples, prune unpromising samples early on, or to pick the best sample. This capability is very inexpensive as it involves generating a single predefined token. Trained using a dataset constructed with real unfiltered LMSYS user prompts, Llama 3.1 8B's win rate against GPT-4 on AlpacaEval increases from 21% to 34% with 16 samples and math performance on GSM8K improves from 84% to 91%. By sampling only when the LLM determines that it is beneficial to do so and adaptively adjusting temperature annealing, we demonstrate that 74% of the improvement from using 16 samples can be achieved with only 1.2 samples on average. We further demonstrate that 50-75% of samples can be pruned early in generation with minimal degradation in performance. Overall, our methods enable more efficient and scalable compute utilization during inference for LLMs.

Large Language Models (LLMs) are pre-trained on large-scale corpora and excel in numerous general natural language processing (NLP) tasks, such as question answering (QA). Despite their advanced language capabilities, when it comes to domain-specific and knowledge-intensive tasks, LLMs suffer from hallucinations, knowledge cut-offs, and lack of knowledge attributions. Additionally, fine tuning LLMs' intrinsic knowledge to highly specific domains is an expensive and time consuming process. The retrieval-augmented generation (RAG) process has recently emerged as a method capable of optimization of LLM responses, by referencing them to a predetermined ontology. It was shown that using a Knowledge Graph (KG) ontology for RAG improves the QA accuracy, by taking into account relevant sub-graphs that preserve the information in a structured manner. In this paper, we introduce SMART-SLIC, a highly domain-specific LLM framework, that integrates RAG with KG and a vector store (VS) that store factual domain specific information. Importantly, to avoid hallucinations in the KG, we build these highly domain-specific KGs and VSs without the use of LLMs, but via NLP, data mining, and nonnegative tensor factorization with automatic model selection. Pairing our RAG with a domain-specific: (i) KG (containing structured information), and (ii) VS (containing unstructured information) enables the development of domain-specific chat-bots that attribute the source of information, mitigate hallucinations, lessen the need for fine-tuning, and excel in highly domain-specific question answering tasks. We pair SMART-SLIC with chain-of-thought prompting agents. The framework is designed to be generalizable to adapt to any specific or specialized domain. In this paper, we demonstrate the question answering capabilities of our framework on a corpus of scientific publications on malware analysis and anomaly detection.

We present UncertaintyRAG, a novel approach for long-context Retrieval-Augmented Generation (RAG) that utilizes Signal-to-Noise Ratio (SNR)-based span uncertainty to estimate similarity between text chunks. This span uncertainty enhances model calibration, improving robustness and mitigating semantic inconsistencies introduced by random chunking. Leveraging this insight, we propose an efficient unsupervised learning technique to train the retrieval model, alongside an effective data sampling and scaling strategy. UncertaintyRAG outperforms baselines by 2.03% on LLaMA-2-7B, achieving state-of-the-art results while using only 4% of the training data compared to other advanced open-source retrieval models under distribution shift settings. Our method demonstrates strong calibration through span uncertainty, leading to improved generalization and robustness in long-context RAG tasks. Additionally, UncertaintyRAG provides a lightweight retrieval model that can be integrated into any large language model with varying context window lengths, without the need for fine-tuning, showcasing the flexibility of our approach.

We introduce LLaVA-Critic, the first open-source large multimodal model (LMM) designed as a generalist evaluator to assess performance across a wide range of multimodal tasks. LLaVA-Critic is trained using a high-quality critic instructionfollowing dataset that incorporates diverse evaluation criteria and scenarios. Our experiments demonstrate the model's effectiveness in two key areas: (i) LMMas-a-Judge, where LLaVA-Critic provides reliable evaluation scores, performing on par with or surpassing GPT models on multiple evaluation benchmarks; and (ii) Preference Learning, where it generates reward signals for preference learning, enhancing model alignment capabilities. This work underscores the potential of open-source LMMs in self-critique and evaluation, setting the stage for future research into scalable, superhuman alignment feedback mechanisms for LMMs. The ability of learning to evaluate is increasingly taking on a pivotal role in the development of modern large multimodal models (LMMs), as pre-training on existing web data reaches maturity and the focus is shifting towards post-training with AI-enhanced synthetic data, which shows growing potential. Reliable AI evaluation is essential, not only for offering scalable solutions to reduce human labor in complex task assessments, but also for generating effective reward signals in reinforcement learning and guiding inference-time search (Ouyang et al., 2022; OpenAI, 2024a; Snell et al., 2024). It remains unexplored to develop open LMMs to play the role of a judge and evaluate the performance of multimodal models. For instance, a model can follow a well-designed, itemized evaluation criterion to provide a score between 1 and 10 for rating different model responses in a visual chat task (Liu et al., 2023b). Along with the score, it would also offer the associated reasoning behind the evaluation, ensuring transparency and consistency in assessing model performance. In this paper, we present the first attempt to curate the instruction-following data particularly for evaluation, based on which we develop a LMM, LLaVA-Critic. Two primary scenarios/goals of building LLaVA-Critic are highlighted: Scenario 1: LMM-as-a-Judge. Open-source LMMs that can deliver reliable evaluation scores, comparable to or surpassing proprietary models like GPT-4V (OpenAI, 2023)/GPT-4o (OpenAI, 2024b). These models can serve as a free alternative to replace commercial GPT models in various evaluation benchmarks. This approach enhances preference alignment with AI-generated feedback. In summary, our contributions are as follows: Critic Instruction-Following Data: We present a high-quality dataset tailored to follow instructions in complex evaluation setting to provide quantitative judgment and the corresponding reasoning process.

LLM watermarks stand out as a promising way to attribute ownership of LLMgenerated text. One threat to watermark credibility comes from spoofing attacks, where an unauthorized third party forges the watermark, enabling it to falsely attribute arbitrary texts to a particular LLM. While recent works have demonstrated that state-of-the-art schemes are in fact vulnerable to spoofing, they lack deeper qualitative analysis of the texts produced by spoofing methods. In this work, we for the first time reveal that there are observable differences between genuine and spoofed watermark texts. Namely, we show that regardless of their underlying approach, all current spoofing methods consistently leave observable artifacts in spoofed texts, indicative of watermark forgery. We build upon these findings to propose rigorous statistical tests that reliably reveal the presence of such artifacts, effectively discovering that a watermark was spoofed. Our experimental evaluation shows high test power across all current spoofing methods, providing insights into their fundamental limitations, and suggesting a way to mitigate this threat. The improving abilities of large language models (LLMs) to generate human-like text at scale (Bubeck et al., 2023; Dubey et al., 2024) come with a growing risk of potential misuse. Hence, reliable detection of machine-generated text becomes increasingly important. Researchers have proposed the concept of watermarking: augmenting generated text with an imperceptible signal that can later be detected to attribute ownership of a text to a specific LLM (Kirchenbauer et al., 2023; Kuditipudi et al., 2023; Christ et al., 2024). Major LLM companies have pledged to watermark their models (Bartz & Hu, 2023), and regulators actively advocate for their use (Biden, 2023; CEU, 2024).

As Large Language Models (LLMs) grow increasingly powerful, ensuring their safety and alignment with human values remains a critical challenge. Ideally, LLMs should provide informative responses while avoiding the disclosure of harmful or sensitive information. However, current alignment approaches, which rely heavily on refusal strategies, such as training models to completely reject harmful prompts or applying coarse filters are limited by their binary nature. These methods either fully deny access to information or grant it without sufficient nuance, leading to overly cautious responses or failures to detect subtle harmful content. For example, LLMs may refuse to provide basic, public information about medication due to misuse concerns. Moreover, these refusal-based methods struggle to handle mixed-content scenarios and lack the ability to adapt to context-dependent sensitivities, which can result in over-censorship of benign content. To overcome these challenges, we introduce HiddenGuard, a novel framework for fine-grained, safe generation in LLMs. HiddenGuard incorporates Prism (rePresentation Router for In-Stream Moderation), which operates alongside the LLM to enable real-time, token-level detection and redaction of harmful content by leveraging intermediate hidden states. This fine-grained approach allows for more nuanced, context-aware moderation, enabling the model to generate informative responses while selectively redacting or replacing sensitive information, rather than outright refusal. We also contribute a comprehensive dataset with token-level fine-grained annotations of potentially harmful information across diverse contexts. Our experiments demonstrate that HiddenGuard achieves over 90% in F1 score for detecting and redacting harmful content while preserving the overall utility and informativeness of the model's responses.

As we increasingly seek guidance from LLMs for decision-making in daily life, many of these decisions are not clear-cut and depend significantly on the personal values and ethical standards of the users. We present DailyDilemmas, a dataset of 1,360 moral dilemmas encountered in everyday life. Each dilemma includes two possible actions and with each action, the affected parties and human values invoked. Based on these dilemmas, we consolidated a set of human values across everyday topics e.g., interpersonal relationships, workplace, and environmental issues. We evaluated LLMs on these dilemmas to determine what action they will take and the values represented by these actions. Then, we analyzed these values through the lens of five popular theories inspired by sociology, psychology and philosophy. These theories are: World Value Survey, Moral Foundation Theory, Maslow's Hierarchy of Needs, Aristotle's Virtues, and Plutchik Wheel of Emotion. We find that LLMs are most aligned with the self-expression over survival values in terms of World Value Survey, care over loyalty in Moral Foundation Theory. Interestingly, we find large preferences differences in models for some core values such as truthfulness e.g., Mixtral-8x7B model tends to neglect it by 9.7% while GPT-4-turbo model tends to select it by 9.4%. We also study the recent guidance released by OpenAI (ModelSpec), and Anthropic (Constitutional AI) to understand how their released principles reflect their actual value prioritization when facing nuanced moral reasoning in daily-life settings. We find that end users cannot effectively steer such prioritization using system prompts.