The correct specification of reward models is a well-known challenge in reinforcement learning. Hand-crafted reward functions often lead to inefficient or suboptimal policies and may not be aligned with user values. Reinforcement learning from human feedback is a successful technique that can mitigate such issues, however, the collection of human feedback can be laborious. Recent works have solicited feedback from pre-trained large language models rather than humans to reduce or eliminate human effort, however, these approaches yield poor performance in the presence of hallucination and other errors. This paper studies the advantages and limitations of reinforcement learning from large language model feedback and proposes a simple yet effective method for soliciting and applying feedback as a potential-based shaping function. We theoretically show that inconsistent rankings, which approximate ranking errors, lead to uninformative rewards with our approach. Our method empirically improves convergence speed and policy returns over commonly used baselines even with significant ranking errors, and eliminates the need for complex post-processing of reward functions.
Delirium is an acute confusional state that has been shown to affect up to 31% of patients in the intensive care unit (ICU). Early detection of this condition could lead to more timely interventions and improved health outcomes. While artificial intelligence (AI) models have shown great potential for ICU delirium prediction using structured electronic health records (EHR), most of them have not explored the use of state-of-the-art AI models, have been limited to single hospitals, or have been developed and validated on small cohorts. The use of large language models (LLM), models with hundreds of millions to billions of parameters, with structured EHR data could potentially lead to improved predictive performance. In this study, we propose DeLLiriuM, a novel LLM-based delirium prediction model using EHR data available in the first 24 hours of ICU admission to predict the probability of a patient developing delirium during the rest of their ICU admission. We develop and validate DeLLiriuM on ICU admissions from 104,303 patients pertaining to 195 hospitals across three large databases: the eICU Collaborative Research Database, the Medical Information Mart for Intensive Care (MIMIC)-IV, and the University of Florida Health's Integrated Data Repository. The performance measured by the area under the receiver operating characteristic curve (AUROC) showed that DeLLiriuM outperformed all baselines in two external validation sets, with 0.77 (95% confidence interval 0.76-0.78) and 0.84 (95% confidence interval 0.83-0.85) across 77,543 patients spanning 194 hospitals. To the best of our knowledge, DeLLiriuM is the first LLM-based delirium prediction tool for the ICU based on structured EHR data, outperforming deep learning baselines which employ structured features and can provide helpful information to clinicians for timely interventions.
Graphs are widely used for modeling relational data in real-world scenarios, such as social networks and urban computing. Existing LLM-based graph analysis approaches either integrate graph neural networks (GNNs) for specific machine learning tasks, limiting their transferability, or rely solely on LLMs' internal reasoning ability, resulting in suboptimal performance. To address these limitations, we take advantage of recent advances in LLM-based agents, which have shown capabilities of utilizing external knowledge or tools for problem solving. By simulating human problem-solving strategies such as analogy and collaboration, we propose a multi-agent system based on LLMs named GraphTeam, for graph analysis. GraphTeam consists of five LLM-based agents from three modules, and the agents with different specialities can collaborate with each other to address complex problems. Specifically, (1) input-output normalization module: the question agent extracts and refines four key arguments from the original question, facilitating the problem understanding, and the answer agent organizes the results to meet the output requirement; (2) external knowledge retrieval module: we first build a knowledge base consisting of relevant documentation and experience information, and then the search agent retrieves the most relevant entries for each question. (3) problem-solving module: given the retrieved information from search agent, the coding agent uses established algorithms via programming to generate solutions, and in case the coding agent does not work, the reasoning agent will directly compute the results without programming. Extensive experiments on six graph analysis benchmarks demonstrate that GraphTeam achieves state-of-the-art performance with an average 25.85% improvement over the best baseline in terms of accuracy. The code and data are available at https://github.com/BUPT-GAMMA/GraphTeam.
This paper introduces a novel mobile phone control architecture, termed ``app agents", for efficient interactions and controls across various Android apps. The proposed Lightweight Multi-modal App Control (LiMAC) takes as input a textual goal and a sequence of past mobile observations, such as screenshots and corresponding UI trees, to generate precise actions. To address the computational constraints inherent to smartphones, within LiMAC, we introduce a small Action Transformer (AcT) integrated with a fine-tuned vision-language model (VLM) for real-time decision-making and task execution. We evaluate LiMAC on two open-source mobile control datasets, demonstrating the superior performance of our small-form-factor approach against fine-tuned versions of open-source VLMs, such as Florence2 and Qwen2-VL. It also significantly outperforms prompt engineering baselines utilising closed-source foundation models like GPT-4o. More specifically, LiMAC increases the overall action accuracy by up to 19% compared to fine-tuned VLMs, and up to 42% compared to prompt-engineering baselines.
Large Language Models (LLMs) are increasingly used in everyday life and research. One of the most common use cases is conversational interactions, enabled by the language generation capabilities of LLMs. Just as between two humans, a conversation between an LLM-powered entity and a human depends on the personality of the conversants. However, measuring the personality of a given LLM is currently a challenge. This paper introduces the Language Model Linguistic Personality Assessment (LMLPA), a system designed to evaluate the linguistic personalities of LLMs. Our system helps to understand LLMs' language generation capabilities by quantitatively assessing the distinct personality traits reflected in their linguistic outputs. Unlike traditional human-centric psychometrics, the LMLPA adapts a personality assessment questionnaire, specifically the Big Five Inventory, to align with the operational capabilities of LLMs, and also incorporates the findings from previous language-based personality measurement literature. To mitigate sensitivity to the order of options, our questionnaire is designed to be open-ended, resulting in textual answers. Thus, the AI rater is needed to transform ambiguous personality information from text responses into clear numerical indicators of personality traits. Utilising Principal Component Analysis and reliability validations, our findings demonstrate that LLMs possess distinct personality traits that can be effectively quantified by the LMLPA. This research contributes to Human-Computer Interaction and Human-Centered AI, providing a robust framework for future studies to refine AI personality assessments and expand their applications in multiple areas, including education and manufacturing.
We present a novel soft prompt based framework, SoftSRV, that leverages a frozen pre-trained large language model (LLM) to generate targeted synthetic text sequences. Given a sample from the target distribution, our proposed framework uses data-driven loss minimization to train a parameterized "contextual" soft prompt. This soft prompt is then used to steer the frozen LLM to generate synthetic sequences that are similar to the target distribution. We argue that SoftSRV provides a practical improvement over common hard-prompting approaches that rely on human-curated prompt-templates, which can be idiosyncratic, labor-intensive to craft, and may need to be specialized per domain. We empirically evaluate SoftSRV and hard-prompting baselines by generating synthetic data to fine-tune a small Gemma model on three different domains (coding, math, reasoning). To stress the generality of SoftSRV, we perform these evaluations without any particular specialization of the framework to each domain. We find that SoftSRV significantly improves upon hard-prompting baselines, generating data with superior fine-tuning performance and that better matches the target distribution according to the MAUVE similarity metric.
Large Language Models (LLMs) have shown great potential in automatically generating and optimizing (meta)heuristics, making them valuable tools in heuristic optimization tasks. However, LLMs are generally inefficient when it comes to fine-tuning hyper-parameters of the generated algorithms, often requiring excessive queries that lead to high computational and financial costs. This paper presents a novel hybrid approach, LLaMEA-HPO, which integrates the open source LLaMEA (Large Language Model Evolutionary Algorithm) framework with a Hyper-Parameter Optimization (HPO) procedure in the loop. By offloading hyper-parameter tuning to an HPO procedure, the LLaMEA-HPO framework allows the LLM to focus on generating novel algorithmic structures, reducing the number of required LLM queries and improving the overall efficiency of the optimization process. We empirically validate the proposed hybrid framework on benchmark problems, including Online Bin Packing, Black-Box Optimization, and the Traveling Salesperson Problem. Our results demonstrate that LLaMEA-HPO achieves superior or comparable performance compared to existing LLM-driven frameworks while significantly reducing computational costs. This work highlights the importance of separating algorithmic innovation and structural code search from parameter tuning in LLM-driven code optimization and offers a scalable approach to improve the efficiency and effectiveness of LLM-based code generation.
Human commonsense understanding of the physical and social world is organized around intuitive theories. These theories support making causal and moral judgments. When something bad happens, we naturally ask: who did what, and why? A rich literature in cognitive science has studied people's causal and moral intuitions. This work has revealed a number of factors that systematically influence people's judgments, such as the violation of norms and whether the harm is avoidable or inevitable. We collected a dataset of stories from 24 cognitive science papers and developed a system to annotate each story with the factors they investigated. Using this dataset, we test whether large language models (LLMs) make causal and moral judgments about text-based scenarios that align with those of human participants. On the aggregate level, alignment has improved with more recent LLMs. However, using statistical analyses, we find that LLMs weigh the different factors quite differently from human participants. These results show how curated, challenge datasets combined with insights from cognitive science can help us go beyond comparisons based merely on aggregate metrics: we uncover LLMs implicit tendencies and show to what extent these align with human intuitions.
Knowledge distillation (KD) is widely used to train small, high-performing student language models (LMs) using large teacher LMs. While effective in fine-tuning, KD during pre-training faces challenges in efficiency, flexibility, and effectiveness. Existing methods either incur high computational costs due to online teacher inference, require tokenization matching between teacher and student LMs, or risk losing the difficulty and diversity of the teacher-generated training data. To address these issues, we propose MiniPLM, a KD framework for pre-training LMs by refining the training data distribution with the teacher's knowledge. For efficiency, MiniPLM performs offline teacher LM inference, allowing KD for multiple student LMs without adding training-time costs. For flexibility, MiniPLM operates solely on the training corpus, enabling KD across model families. For effectiveness, MiniPLM leverages the differences between large and small LMs to enhance the difficulty and diversity of the training data, helping student LMs acquire versatile and sophisticated knowledge. Extensive experiments demonstrate that MiniPLM boosts the student LMs' performance on 9 widely used downstream tasks, improves the language modeling capabilities, and reduces pre-training computation. The benefit of MiniPLM extends to large pre-training scales, evidenced by the extrapolation of the scaling curves. Further analysis reveals that MiniPLM supports KD across model families and enhances the utilization of pre-training data. Our model, code, and data are available at https://github.com/thu-coai/MiniPLM.
Quantization and pruning are fundamental approaches for model compression, enabling efficient inference for language models. In a post-training setting, state-of-the-art quantization and pruning methods require calibration data, a small set of unlabeled examples. Conventionally, randomly sampled web text is used, aiming to reflect the model training data. However, this poses two key problems: (1) unrepresentative calibration examples can harm model performance, and (2) organizations increasingly avoid releasing model training data. In this paper, we propose self-calibration as a solution. Our approach requires no external data, instead leveraging the model itself to generate synthetic calibration data as a better approximation of the pre-training data distribution. We extensively compare the performance of self-calibration with several baselines, across a variety of models, compression methods, and tasks. Our approach proves consistently competitive in maximizing downstream task performance, frequently outperforming even using real data.
To improve relevance scoring on Pinterest Search, we integrate Large Language Models (LLMs) into our search relevance model, leveraging carefully designed text representations to predict the relevance of Pins effectively. Our approach uses search queries alongside content representations that include captions extracted from a generative visual language model. These are further enriched with link-based text data, historically high-quality engaged queries, user-curated boards, Pin titles and Pin descriptions, creating robust models for predicting search relevance. We use a semi-supervised learning approach to efficiently scale up the amount of training data, expanding beyond the expensive human labeled data available. By utilizing multilingual LLMs, our system extends training data to include unseen languages and domains, despite initial data and annotator expertise being confined to English. Furthermore, we distill from the LLM-based model into real-time servable model architectures and features. We provide comprehensive offline experimental validation for our proposed techniques and demonstrate the gains achieved through the final deployed system at scale.
Users can divulge sensitive information to proprietary LLM providers, raising significant privacy concerns. While open-source models, hosted locally on the user's machine, alleviate some concerns, models that users can host locally are often less capable than proprietary frontier models. Toward preserving user privacy while retaining the best quality, we propose Privacy-Conscious Delegation, a novel task for chaining API-based and local models. We utilize recent public collections of user-LLM interactions to construct a natural benchmark called PUPA, which contains personally identifiable information (PII). To study potential approaches, we devise PAPILLON, a multi-stage LLM pipeline that uses prompt optimization to address a simpler version of our task. Our best pipeline maintains high response quality for 85.5% of user queries while restricting privacy leakage to only 7.5%. We still leave a large margin to the generation quality of proprietary LLMs for future work. Our data and code is available at https://github.com/siyan-sylvia-li/PAPILLON.
The fusion of speech and language in the era of large language models has garnered significant attention. Discrete speech token is often utilized in text-to-speech tasks for speech compression and portability, which is convenient for joint training with text and have good compression efficiency. However, we found that the discrete speech tokenizer still suffers from information loss. Therefore, we propose a simple yet effective continuous speech tokenizer named Cont-SPT, and a text-to-speech model based on continuous speech tokens. Our results show that the speech language model based on the continuous speech tokenizer has better continuity and higher estimated Mean Opinion Scores (MoS). This enhancement is attributed to better information preservation rate of the continuous speech tokenizer across both low and high frequencies in the frequency domain. The code and resources for Cont-SPT can be found in https://github.com/Yixing-Li/Continuous-Speech-Tokenizer
Sharing protected health information (PHI) is critical for furthering biomedical research. Before data can be distributed, practitioners often perform deidentification to remove any PHI contained in the text. Contemporary deidentification methods are evaluated on highly saturated datasets (tools achieve near-perfect accuracy) which may not reflect the full variability or complexity of real-world clinical text and annotating them is resource intensive, which is a barrier to real-world applications. To address this gap, we developed an adversarial approach using a large language model (LLM) to re-identify the patient corresponding to a redacted clinical note and evaluated the performance with a novel De-Identification/Re-Identification (DIRI) method. Our method uses a large language model to reidentify the patient corresponding to a redacted clinical note. We demonstrate our method on medical data from Weill Cornell Medicine anonymized with three deidentification tools: rule-based Philter and two deep-learning-based models, BiLSTM-CRF and ClinicalBERT. Although ClinicalBERT was the most effective, masking all identified PII, our tool still reidentified 9% of clinical notes Our study highlights significant weaknesses in current deidentification technologies while providing a tool for iterative development and improvement.
Catastrophic forgetting remains a formidable obstacle to building an omniscient model in large language models (LLMs). Despite the pioneering research on task-level forgetting in LLM fine-tuning, there is scant focus on forgetting during pre-training. We systematically explored the existence and measurement of forgetting in pre-training, questioning traditional metrics such as perplexity (PPL) and introducing new metrics to better detect entity memory retention. Based on our revised assessment of forgetting metrics, we explored low-cost, straightforward methods to mitigate forgetting during the pre-training phase. Further, we carefully analyzed the learning curves, offering insights into the dynamics of forgetting. Extensive evaluations and analyses on forgetting of pre-training could facilitate future research on LLMs.
There are already many DNA large language models, but most of them still follow traditional uses, such as extracting sequence features for classification tasks. More innovative applications of large language models, such as prompt engineering, RAG, and zero-shot or few-shot prediction, remain challenging for DNA-based models. The key issue lies in the fact that DNA models and human natural language models are entirely separate; however, techniques like prompt engineering require the use of natural language, thereby significantly limiting the application of DNA large language models. This paper introduces a pre-trained model trained on the GPT-2 network, combining DNA sequences and English text, and uses a unified BPE tokenization method. We then convert classification and other downstream tasks into Alpaca format instruction data, and perform instruction fine-tuning on this pre-trained model to create a fine-tuned model capable of handling multiple tasks. The model has demonstrated its effectiveness in DNA related zero-shot prediction and multitask application. This research provides a highly promising direction for building a unified DNA sequence task framework.
Access to real clinical data is often restricted due to privacy obligations, creating significant barriers for healthcare research. Synthetic datasets provide a promising solution, enabling secure data sharing and model development. However, most existing approaches focus on data realism rather than utility -- ensuring that models trained on synthetic data yield clinically meaningful insights comparable to those trained on real data. In this paper, we present Masked Clinical Modelling (MCM), a framework inspired by masked language modelling, designed for both data synthesis and conditional data augmentation. We evaluate this prototype on the WHAS500 dataset using Cox Proportional Hazards models, focusing on the preservation of hazard ratios as key clinical metrics. Our results show that data generated using the MCM framework improves both discrimination and calibration in survival analysis, outperforming existing methods. MCM demonstrates strong potential to support survival data analysis and broader healthcare applications.
Self-play methods have demonstrated remarkable success in enhancing model capabilities across various domains. In the context of Reinforcement Learning from Human Feedback (RLHF), self-play not only boosts Large Language Model (LLM) performance but also overcomes the limitations of traditional Bradley-Terry (BT) model assumptions by finding the Nash equilibrium (NE) of a preference-based, two-player constant-sum game. However, existing methods either guarantee only average-iterate convergence, incurring high storage and inference costs, or converge to the NE of a regularized game, failing to accurately reflect true human preferences. In this paper, we introduce Magnetic Preference Optimization (MPO), a novel approach capable of achieving last-iterate convergence to the NE of the original game, effectively overcoming the limitations of existing methods. Building upon Magnetic Mirror Descent (MMD), MPO attains a linear convergence rate, making it particularly suitable for fine-tuning LLMs. To ensure our algorithm is both theoretically sound and practically viable, we present a simple yet effective implementation that adapts the theoretical insights to the RLHF setting. Empirical results demonstrate that MPO can significantly enhance the performance of LLMs, highlighting the potential of self-play methods in alignment.
Adsorption energy is a key reactivity descriptor in catalysis, enabling efficient screening for optimal catalysts. However, determining adsorption energy typically requires evaluating numerous adsorbate-catalyst configurations. Current algorithmic approaches rely on exhaustive enumeration of adsorption sites and configurations, which makes the process computationally intensive and does not inherently guarantee the identification of the global minimum energy. In this work, we introduce Adsorb-Agent, a Large Language Model (LLM) agent designed to efficiently identify system-specific stable adsorption configurations corresponding to the global minimum adsorption energy. Adsorb-Agent leverages its built-in knowledge and emergent reasoning capabilities to strategically explore adsorption configurations likely to hold adsorption energy. By reducing the reliance on exhaustive sampling, it significantly decreases the number of initial configurations required while improving the accuracy of adsorption energy predictions. We evaluate Adsorb-Agent's performance across twenty representative systems encompassing a range of complexities. The Adsorb-Agent successfully identifies comparable adsorption energies for 83.7% of the systems and achieves lower energies, closer to the actual global minimum, for 35% of the systems, while requiring significantly fewer initial configurations than conventional methods. Its capability is particularly evident in complex systems, where it identifies lower adsorption energies for 46.7% of systems involving intermetallic surfaces and 66.7% of systems with large adsorbate molecules. These results demonstrate the potential of Adsorb-Agent to accelerate catalyst discovery by reducing computational costs and improving the reliability of adsorption energy predictions.
Travel planning is a challenging and time-consuming task that aims to find an itinerary which satisfies multiple, interdependent constraints regarding flights, accommodations, attractions, and other travel arrangements. In this paper, we propose To the Globe (TTG), a real-time demo system that takes natural language requests from users, translates it to symbolic form via a fine-tuned Large Language Model, and produces optimal travel itineraries with Mixed Integer Linear Programming solvers. The overall system takes ~5 seconds to reply to the user request with guaranteed itineraries. To train TTG, we develop a synthetic data pipeline that generates user requests, flight and hotel information in symbolic form without human annotations, based on the statistics of real-world datasets, and fine-tune an LLM to translate NL user requests to their symbolic form, which is sent to the symbolic solver to compute optimal itineraries. Our NL-symbolic translation achieves ~91% exact match in a backtranslation metric (i.e., whether the estimated symbolic form of generated natural language matches the groundtruth), and its returned itineraries have a ratio of 0.979 compared to the optimal cost of the ground truth user request. When evaluated by users, TTG achieves consistently high Net Promoter Scores (NPS) of 35-40% on generated itinerary.