Large language models (LLMs) can solve complex multi-step problems, but little is known about how these computations are implemented internally. Motivated by this, we study how LLMs answer multi-hop queries such as "The spouse of the performer of Imagine is". These queries require two information extraction steps: a latent one for resolving the first hop ("the performer of Imagine") into the bridge entity (John Lennon), and one for resolving the second hop ("the spouse of John Lennon") into the target entity (Yoko Ono). Understanding how the latent step is computed internally is key to understanding the overall computation. By carefully analyzing the internal computations of transformer-based LLMs, we discover that the bridge entity is resolved in the early layers of the model. Then, only after this resolution, the two-hop query is solved in the later layers. Because the second hop commences in later layers, there could be cases where these layers no longer encode the necessary knowledge for correctly predicting the answer. Motivated by this, we propose a novel "back-patching" analysis method whereby a hidden representation from a later layer is patched back to an earlier layer. We find that in up to 57% of previously incorrect cases there exists a back-patch that results in the correct generation of the answer, showing that the later layers indeed sometimes lack the needed functionality. Overall our methods and findings open further opportunities for understanding and improving latent reasoning in transformer-based LLMs.

The parametric knowledge memorized by large language models (LLMs) becomes outdated quickly. In-context editing (ICE) is currently the most effective method for updating the knowledge of LLMs. Recent advancements involve enhancing ICE by modifying the decoding strategy, obviating the need for altering internal model structures or adjusting external prompts. However, this enhancement operates across the entire sequence generation, encompassing a plethora of non-critical tokens. In this work, we introduce $\textbf{A}$daptive $\textbf{T}$oken $\textbf{Bias}$er ($\textbf{ATBias}$), a new decoding technique designed to enhance ICE. It focuses on the tokens that are mostly related to knowledge during decoding, biasing their logits by matching key entities related to new and parametric knowledge. Experimental results show that ATBias significantly enhances ICE performance, achieving up to a 32.3% improvement over state-of-the-art ICE methods while incurring only half the latency. ATBias not only improves the knowledge editing capabilities of ICE but can also be widely applied to LLMs with negligible cost.

Reinforcement learning from human feedback (RLHF) has emerged as the primary method for aligning large language models (LLMs) with human preferences. The RLHF process typically starts by training a reward model (RM) using human preference data. Conventional RMs are trained on pairwise responses to the same user request, with relative ratings indicating which response humans prefer. The trained RM serves as a proxy for human preferences. However, due to the black-box nature of RMs, their outputs lack interpretability, as humans cannot intuitively understand why an RM thinks a response is good or not. As RMs act as human preference proxies, we believe they should be human-interpretable to ensure that their internal decision processes are consistent with human preferences and to prevent reward hacking in LLM alignment. To build RMs with interpretable preferences, we propose a two-stage approach: i) train an Absolute-Rating Multi-Objective Reward Model (ArmoRM) with multi-dimensional absolute-rating data, each dimension corresponding to a human-interpretable objective (e.g., honesty, verbosity, safety); ii) employ a Mixture-of-Experts (MoE) strategy with a gating network that automatically selects the most suitable reward objectives based on the context. We efficiently trained an ArmoRM with Llama-3 8B and a gating network consisting of a shallow MLP on top of the ArmoRM. Our trained model, ArmoRM-Llama3-8B, obtains state-of-the-art performance on RewardBench, a benchmark evaluating RMs for language modeling. Notably, the performance of our model surpasses the LLM-as-a-judge method with GPT-4 judges by a margin, and approaches the performance of the much larger Nemotron-4 340B reward model.

Assessing the capabilities of large multimodal models (LMMs) often requires the creation of ad-hoc evaluations. Currently, building new benchmarks requires tremendous amounts of manual work for each specific analysis. This makes the evaluation process tedious and costly. In this paper, we present APEx, Automatic Programming of Experiments, the first framework for automatic benchmarking of LMMs. Given a research question expressed in natural language, APEx leverages a large language model (LLM) and a library of pre-specified tools to generate a set of experiments for the model at hand, and progressively compile a scientific report. The report drives the testing procedure: based on the current status of the investigation, APEx chooses which experiments to perform and whether the results are sufficient to draw conclusions. Finally, the LLM refines the report, presenting the results to the user in natural language. Thanks to its modularity, our framework is flexible and extensible as new tools become available. Empirically, APEx reproduces the findings of existing studies while allowing for arbitrary analyses and hypothesis testing.

Long prompt leads to huge hardware costs when using transformer-based Large Language Models (LLMs). Unfortunately, many tasks, such as summarization, inevitably introduce long documents, and the wide application of in-context learning easily makes the prompt length explode. This paper proposes a Self-Compressor (SelfCP), which employs the target LLM itself to compress over-limit prompts into dense vectors while keeping the allowed prompts unmodified. Dense vectors are then projected into dense tokens via a learnable connector to make the same LLM unburden to understand. The connector is supervised-tuned under the language modeling objective of the LLM on relatively long texts selected from publicly accessed datasets, involving an instruction dataset to make SelfCP respond to various prompts, while the target LLM keeps frozen during training. We build the lightweight SelfCP upon 2 different backbones with merely 17M learnable parameters originating from the connector and a learnable embedding. Evaluation on both English and Chinese benchmarks demonstrate that SelfCP effectively substitutes 12$\times$ over-limit prompts with dense tokens to reduce memory costs and booster inference throughputs, yet improving response quality. The outstanding performance brings an efficient solution for LLMs to tackle long prompts without training LLMs from scratch.

In the context of real-world applications, leveraging large language models (LLMs) for domain-specific tasks often faces two major challenges: domain-specific knowledge privacy and constrained resources. To address these issues, we propose PDSS, a privacy-preserving framework for step-by-step distillation of LLMs. PDSS works on a server-client architecture, wherein client transmits perturbed prompts to the server's LLM for rationale generation. The generated rationales are then decoded by the client and used to enrich the training of task-specific small language model(SLM) within a multi-task learning paradigm. PDSS introduces two privacy protection strategies: the Exponential Mechanism Strategy and the Encoder-Decoder Strategy, balancing prompt privacy and rationale usability. Experiments demonstrate the effectiveness of PDSS in various text generation tasks, enabling the training of task-specific SLM with enhanced performance while prioritizing data privacy protection.

Ultrasonography has revolutionized non-invasive diagnostic methodologies, significantly enhancing patient outcomes across various medical domains. Despite its advancements, integrating ultrasound technology with robotic systems for automated scans presents challenges, including limited command understanding and dynamic execution capabilities. To address these challenges, this paper introduces a novel Ultrasound Embodied Intelligence system that synergistically combines ultrasound robots with large language models (LLMs) and domain-specific knowledge augmentation, enhancing ultrasound robots' intelligence and operational efficiency. Our approach employs a dual strategy: firstly, integrating LLMs with ultrasound robots to interpret doctors' verbal instructions into precise motion planning through a comprehensive understanding of ultrasound domain knowledge, including APIs and operational manuals; secondly, incorporating a dynamic execution mechanism, allowing for real-time adjustments to scanning plans based on patient movements or procedural errors. We demonstrate the effectiveness of our system through extensive experiments, including ablation studies and comparisons across various models, showcasing significant improvements in executing medical procedures from verbal commands. Our findings suggest that the proposed system improves the efficiency and quality of ultrasound scans and paves the way for further advancements in autonomous medical scanning technologies, with the potential to transform non-invasive diagnostics and streamline medical workflows.

Large Language Models (LLMs) have become increasingly capable of handling diverse tasks with the aid of well-crafted prompts and integration of external tools, but as task complexity rises, the workflow involving LLMs can be complicated and thus challenging to implement and maintain. To address this challenge, we propose APPL, A Prompt Programming Language that acts as a bridge between computer programs and LLMs, allowing seamless embedding of prompts into Python functions, and vice versa. APPL provides an intuitive and Python-native syntax, an efficient parallelized runtime with asynchronous semantics, and a tracing module supporting effective failure diagnosis and replaying without extra costs. We demonstrate that APPL programs are intuitive, concise, and efficient through three representative scenarios: Chain-of-Thought with self-consistency (CoT-SC), ReAct tool use agent, and multi-agent chat. Experiments on three parallelizable workflows further show that APPL can effectively parallelize independent LLM calls, with a significant speedup ratio that almost matches the estimation.

Sequential recommendation (SR) tasks enhance recommendation accuracy by capturing the connection between users' past interactions and their changing preferences. Conventional models often focus solely on capturing sequential patterns within the training data, neglecting the broader context and semantic information embedded in item titles from external sources. This limits their predictive power and adaptability. Recently, large language models (LLMs) have shown promise in SR tasks due to their advanced understanding capabilities and strong generalization abilities. Researchers have attempted to enhance LLMs' recommendation performance by incorporating information from SR models. However, previous approaches have encountered problems such as 1) only influencing LLMs at the result level; 2) increased complexity of LLMs recommendation methods leading to reduced interpretability; 3) incomplete understanding and utilization of SR models information by LLMs. To address these problems, we proposes a novel framework, DELRec, which aims to extract knowledge from SR models and enable LLMs to easily comprehend and utilize this supplementary information for more effective sequential recommendations. DELRec consists of two main stages: 1) SR Models Pattern Distilling, focusing on extracting behavioral patterns exhibited by SR models using soft prompts through two well-designed strategies; 2) LLMs-based Sequential Recommendation, aiming to fine-tune LLMs to effectively use the distilled auxiliary information to perform SR tasks. Extensive experimental results conducted on three real datasets validate the effectiveness of the DELRec framework.

Location-based services play an critical role in improving the quality of our daily lives. Despite the proliferation of numerous specialized AI models within spatio-temporal context of location-based services, these models struggle to autonomously tackle problems regarding complex urban planing and management. To bridge this gap, we introduce UrbanLLM, a fine-tuned large language model (LLM) designed to tackle diverse problems in urban scenarios. UrbanLLM functions as a problem-solver by decomposing urban-related queries into manageable sub-tasks, identifying suitable spatio-temporal AI models for each sub-task, and generating comprehensive responses to the given queries. Our experimental results indicate that UrbanLLM significantly outperforms other established LLMs, such as Llama and the GPT series, in handling problems concerning complex urban activity planning and management. UrbanLLM exhibits considerable potential in enhancing the effectiveness of solving problems in urban scenarios, reducing the workload and reliance for human experts.