As data retrieval demands become increasingly complex, traditional search methods often fall short in addressing nuanced and conceptual queries. Vector similarity search has emerged as a promising technique for finding semantically similar information efficiently. However, its effectiveness diminishes when handling intricate queries with contextual nuances. This paper explores a hybrid approach combining vector similarity search with Large Language Models (LLMs) to enhance search accuracy and relevance. The proposed two-step solution first employs vector similarity search to shortlist potential matches, followed by an LLM for context-aware ranking of the results. Experiments on structured datasets demonstrate that while vector similarity search alone performs well for straightforward queries, the LLM-assisted approach excels in processing complex queries involving constraints, negations, or conceptual requirements. By leveraging the natural language understanding capabilities of LLMs, this method improves the accuracy of search results for complex tasks without sacrificing efficiency. We also discuss real-world applications and propose directions for future research to refine and scale this technique for diverse datasets and use cases. Original article: https://engineering.grab.com/llm-assisted-vector-similarity-search

The growing ubiquity of Retrieval-Augmented Generation (RAG) systems in several real-world services triggers severe concerns about their security. A RAG system improves the generative capabilities of a Large Language Models (LLM) by a retrieval mechanism which operates on a private knowledge base, whose unintended exposure could lead to severe consequences, including breaches of private and sensitive information. This paper presents a black-box attack to force a RAG system to leak its private knowledge base which, differently from existing approaches, is adaptive and automatic. A relevance-based mechanism and an attacker-side open-source LLM favor the generation of effective queries to leak most of the (hidden) knowledge base. Extensive experimentation proves the quality of the proposed algorithm in different RAG pipelines and domains, comparing to very recent related approaches, which turn out to be either not fully black-box, not adaptive, or not based on open-source models. The findings from our study remark the urgent need for more robust privacy safeguards in the design and deployment of RAG systems.

While our previous study examined four prominent LLMs using the Moral Machine experimental framework, the dynamic landscape of LLM development demands a more comprehensive analysis. Here, we evaluate moral judgments across 52 different LLMs, including multiple versions of proprietary models (GPT, Claude, Gemini) and open-source alternatives (Llama, Gemma), to assess their alignment with human moral preferences in autonomous driving scenarios. Using a conjoint analysis framework, we evaluated how closely LLM responses aligned with human preferences in ethical dilemmas and examined the effects of model size, updates, and architecture. Results showed that proprietary models and open-source models exceeding 10 billion parameters demonstrated relatively close alignment with human judgments, with a significant negative correlation between model size and distance from human judgments in open-source models. However, model updates did not consistently improve alignment with human preferences, and many LLMs showed excessive emphasis on specific ethical principles. These findings suggest that while increasing model size may naturally lead to more human-like moral judgments, practical implementation in autonomous driving systems requires careful consideration of the trade-off between judgment quality and computational efficiency. Our comprehensive analysis provides crucial insights for the ethical design of autonomous systems and highlights the importance of considering cultural contexts in AI moral decision-making.

Despite their tremendous success and versatility, Large Language Models (LLMs) suffer from inference inefficiency while relying on advanced computational infrastructure. To address these challenges and make LLMs more accessible and cost-effective, in this paper, we propose algorithms to improve the inference time and memory efficiency of 1.58-bit LLMs with ternary weight matrices. Particularly focusing on matrix multiplication as the bottle-neck operation of inference, we observe that, once trained, the weight matrices of a model no longer change. This allows us to preprocess these matrices and create indices that help reduce the storage requirements by a logarithmic factor while enabling our efficient inference algorithms. Specifically, for a $n$ by $n$ weight matrix, our efficient algorithm guarantees a time complexity of $O(\frac{n^2}{\log n})$, a logarithmic factor improvement over the standard $O(n^2)$ vector-matrix multiplication. Besides theoretical analysis, we conduct extensive experiments to evaluate the practical efficiency of our algorithms. Our results confirm the superiority of the approach both with respect to time and memory, as we observed a reduction in inference time up to 29x and memory usage up to 6x.

We present a simple way to merge masked language modeling with causal language modeling. This hybrid training objective results in a model that combines the strengths of both modeling paradigms within a single transformer stack: GPT-BERT can be transparently used like any standard causal or masked language model. We test the pretraining process that enables this flexible behavior on the BabyLM Challenge 2024. The results show that the hybrid pretraining outperforms masked-only or causal-only models. We openly release the models, training corpora and code.

Long-context LLMs are increasingly in demand for applications such as retrieval-augmented generation. To defray the cost of pretraining LLMs over long contexts, recent work takes an approach of synthetic context extension: fine-tuning LLMs with synthetically generated long-context data in a post-training stage. However, it remains unclear how and why this synthetic context extension imparts abilities for downstream long-context tasks. In this paper, we investigate fine-tuning on synthetic data for three long-context tasks that require retrieval and reasoning. We vary the realism of "needle" concepts to be retrieved and diversity of the surrounding "haystack" context, from using LLMs to construct synthetic documents to using templated relations and creating symbolic datasets. We find that models trained on synthetic data fall short of the real data, but surprisingly, the mismatch can be interpreted and even predicted in terms of a special set of attention heads that are responsible for retrieval over long context, retrieval heads (Wu et al., 2024). The retrieval heads learned on synthetic data have high overlap with retrieval heads learned on real data, and there is a strong correlation between the recall of heads learned and the downstream performance of a model. Furthermore, with attention knockout and activation patching, we mechanistically show that retrieval heads are necessary and explain model performance, although they are not totally sufficient. Our results shed light on how to interpret synthetic data fine-tuning performance and how to approach creating better data for learning real-world capabilities over long contexts.

We show that existing evaluations for fake news detection based on conventional sources, such as claims on fact-checking websites, result in high accuracies over time for LLM-based detectors -- even after their knowledge cutoffs. This suggests that recent popular fake news from such sources can be easily detected due to pre-training and retrieval corpus contamination or increasingly salient shallow patterns. Instead, we argue that a proper fake news detection dataset should test a model's ability to reason factually about the current world by retrieving and reading related evidence. To this end, we develop a novel pipeline that leverages natural language feedback from a RAG-based detector to iteratively modify real-time news into deceptive fake news that challenges LLMs. Our iterative rewrite decreases the binary classification ROC-AUC by an absolute 17.5 percent for a strong RAG-based GPT-4o detector. Our experiments reveal the important role of RAG in both detecting and generating fake news, as retrieval-free LLM detectors are vulnerable to unseen events and adversarial attacks, while feedback from RAG detection helps discover more deceitful patterns in fake news.

If 2024 was the year of large language models (LLMs), then 2025 looks like the year of AI "agents". These are quasi-intelligent systems that harness LLMs to go beyond their usual tricks of generating plausible text or responding to prompts. The idea is that an agent can be given a high-level – possibly even vague – goal and break it down into a series of actionable steps. Once it "understands" the goal, it can devise a plan to achieve it, much as a human would. OpenAI's chief financial officer, Sarah Friar, recently explained it thus to the Financial Times: "It could be a researcher, a helpful assistant for everyday people, working moms like me. In 2025, we will see the first very successful agents deployed that help people in their day to day."

Large language models (LLMs) have demonstrated emergent abilities in text generation, question answering, and reasoning, facilitating various tasks and domains. Despite their proficiency in various tasks, LLMs like PaLM 540B and Llama-3.1 405B face limitations due to large parameter sizes and computational demands, often requiring cloud API use which raises privacy concerns, limits real-time applications on edge devices, and increases fine-tuning costs. Additionally, LLMs often underperform in specialized domains such as healthcare and law due to insufficient domain-specific knowledge, necessitating specialized models. Therefore, Small Language Models (SLMs) are increasingly favored for their low inference latency, cost-effectiveness, efficient development, and easy customization and adaptability. These models are particularly well-suited for resource-limited environments and domain knowledge acquisition, addressing LLMs' challenges and proving ideal for applications that require localized data handling for privacy, minimal inference latency for efficiency, and domain knowledge acquisition through lightweight fine-tuning. The rising demand for SLMs has spurred extensive research and development. However, a comprehensive survey investigating issues related to the definition, acquisition, application, enhancement, and reliability of SLM remains lacking, prompting us to conduct a detailed survey on these topics. The definition of SLMs varies widely, thus to standardize, we propose defining SLMs by their capability to perform specialized tasks and suitability for resource-constrained settings, setting boundaries based on the minimal size for emergent abilities and the maximum size sustainable under resource constraints. For other aspects, we provide a taxonomy of relevant models/methods and develop general frameworks for each category to enhance and utilize SLMs effectively.

Complex table question answering (TQA) aims to answer questions that require complex reasoning, such as multi-step or multi-category reasoning, over data represented in tabular form. Previous approaches demonstrated notable performance by leveraging either closed-source large language models (LLMs) or fine-tuned open-weight LLMs. However, fine-tuning LLMs requires high-quality training data, which is costly to obtain, and utilizing closed-source LLMs poses accessibility challenges and leads to reproducibility issues. In this paper, we propose Multi-Agent Collaboration with Tool use (MACT), a framework that requires neither closed-source models nor fine-tuning. In MACT, a planning agent and a coding agent that also make use of tools collaborate to answer questions. Our experiments on four TQA benchmarks show that MACT outperforms previous SoTA systems on three out of four benchmarks and that it performs comparably to the larger and more expensive closed-source model GPT-4 on two benchmarks, even when using only open-weight models without any fine-tuning. We conduct extensive analyses to prove the effectiveness of MACT's multi-agent collaboration in TQA.