Carnegie Mellon University
Abstract:Large language models (LLMs) have shown promising capabilities in using external tools to solve complex problems. However, existing approaches either involve fine-tuning on tool demonstrations, which do not generalize to new tools without additional training, or providing tool documentation in context, limiting the number of tools. Both approaches often generate syntactically invalid tool calls. In this paper, we propose ToolDec, a finite-state machine-guided decoding algorithm for tool-augmented LLMs. ToolDec eliminates tool-related errors for any tool-augmented LLMs by ensuring valid tool names and type-conforming arguments. Furthermore, ToolDec enables LLM to effectively select tools using only the information contained in their names, with no need for fine-tuning or in-context documentation. We evaluated multiple prior methods and their ToolDec-enhanced versions on a variety of tasks involving tools like math functions, knowledge graph relations, and complex real-world RESTful APIs. Our experiments show that ToolDec reduces syntactic errors to zero, consequently achieving significantly better performance and as much as a 2x speedup. We also show that ToolDec achieves superior generalization performance on unseen tools, performing up to 8x better than the baselines.
Abstract:While large language models (LLMs) have shown impressive results for more objective tasks such as QA and retrieval, it remains nontrivial to evaluate their performance on open-ended text generation for reasons including (1) data contamination; (2) multi-dimensional evaluation criteria; and (3) subjectiveness stemming from reviewers' personal preferences. To address such issues, we propose to model personalization in an uncontaminated open-ended generation assessment. We create two new datasets Per-MPST and Per-DOC for personalized story evaluation, by re-purposing existing datasets with proper anonymization and new personalized labels. We further develop a personalized story evaluation model PERSE to infer reviewer preferences and provide a personalized evaluation. Specifically, given a few exemplary reviews from a particular reviewer, PERSE predicts either a detailed review or fine-grained comparison in several aspects (such as interestingness and surprise) for that reviewer on a new text input. Experimental results show that PERSE outperforms GPT-4 by 15.8% on Kendall correlation of story ratings, and by 13.7% on pairwise preference prediction accuracy. Both datasets and code will be released.
Abstract:Proteins are macromolecules responsible for essential functions in almost all living organisms. Designing reasonable proteins with desired functions is crucial. A protein's sequence and structure are strongly correlated and they together determine its function. In this paper, we propose NAEPro, a model to jointly design Protein sequence and structure based on automatically detected functional sites. NAEPro is powered by an interleaving network of attention and equivariant layers, which can capture global correlation in a whole sequence and local influence from nearest amino acids in three dimensional (3D) space. Such an architecture facilitates effective yet economic message passing at two levels. We evaluate our model and several strong baselines on two protein datasets, $\beta$-lactamase and myoglobin. Experimental results show that our model consistently achieves the highest amino acid recovery rate, TM-score, and the lowest RMSD among all competitors. These findings prove the capability of our model to design protein sequences and structures that closely resemble their natural counterparts. Furthermore, in-depth analysis further confirms our model's ability to generate highly effective proteins capable of binding to their target metallocofactors. We provide code, data and models in Github.
Abstract:Designing novel proteins with desired functions is crucial in biology and chemistry. However, most existing work focus on protein sequence design, leaving protein sequence and structure co-design underexplored. In this paper, we propose GeoPro, a method to design protein backbone structure and sequence jointly. Our motivation is that protein sequence and its backbone structure constrain each other, and thus joint design of both can not only avoid nonfolding and misfolding but also produce more diverse candidates with desired functions. To this end, GeoPro is powered by an equivariant encoder for three-dimensional (3D) backbone structure and a protein sequence decoder guided by 3D geometry. Experimental results on two biologically significant metalloprotein datasets, including $\beta$-lactamases and myoglobins, show that our proposed GeoPro outperforms several strong baselines on most metrics. Remarkably, our method discovers novel $\beta$-lactamases and myoglobins which are not present in protein data bank (PDB) and UniProt. These proteins exhibit stable folding and active site environments reminiscent of those of natural proteins, demonstrating their excellent potential to be biologically functional.
Abstract:Reward modeling (a.k.a., preference modeling) is instrumental for aligning large language models with human preferences, particularly within the context of reinforcement learning from human feedback (RLHF). While conventional reward models (RMs) have exhibited remarkable scalability, they oft struggle with fundamental functionality such as arithmetic computation, code execution, and factual lookup. In this paper, we propose a tool-augmented preference modeling approach, named \name, to address these limitations by empowering RMs with access to external environments, including calculators and search engines. This approach not only fosters synergy between tool utilization and reward grading but also enhances interpretive capacity and scoring reliability. Our study delves into the integration of external tools into RMs, enabling them to interact with diverse external sources and construct task-specific tool engagement and reasoning traces in an autoregressive manner. We validate our approach across a wide range of domains, incorporating seven distinct external tools. Our experimental results demonstrate a noteworthy overall improvement of 17.7% across eight tasks in preference ranking. Furthermore, our approach outperforms Gopher 280B by 7.3% on TruthfulQA task in zero-shot evaluation. In human evaluations, RLHF trained with Themis attains an average win rate of 32% when compared to baselines across four distinct tasks. Additionally, we provide a comprehensive collection of tool-related RM datasets, incorporating data from seven distinct tool APIs, totaling 15,000 instances. We anticipate that this publicly available dataset will facilitate and inspire further research advancements in the field.
Abstract:This paper proposes an innovative approach to Hierarchical Edge Aware 3D Point Cloud Learning (HEA-Net) that seeks to address the challenges of noise in point cloud data, and improve object recognition and segmentation by focusing on edge features. In this study, we present an innovative edge-aware learning methodology, specifically designed to enhance point cloud classification and segmentation. Drawing inspiration from the human visual system, the concept of edge-awareness has been incorporated into this methodology, contributing to improved object recognition while simultaneously reducing computational time. Our research has led to the development of an advanced 3D point cloud learning framework that effectively manages object classification and segmentation tasks. A unique fusion of local and global network learning paradigms has been employed, enriched by edge-focused local and global embeddings, thereby significantly augmenting the model's interpretative prowess. Further, we have applied a hierarchical transformer architecture to boost point cloud processing efficiency, thus providing nuanced insights into structural understanding. Our approach demonstrates significant promise in managing noisy point cloud data and highlights the potential of edge-aware strategies in 3D point cloud learning. The proposed approach is shown to outperform existing techniques in object classification and segmentation tasks, as demonstrated by experiments on ModelNet40 and ShapeNet datasets.
Abstract:Reasoning is a cognitive process of using evidence to reach a sound conclusion. The reasoning capability is essential for large language models (LLMs) to serve as the brain of the artificial general intelligence agent. Recent studies reveal that fine-tuning LLMs on data with the chain of thought (COT) reasoning process can significantly enhance their reasoning capabilities. However, we find that the fine-tuned LLMs suffer from an \textit{Assessment Misalignment} problem, i.e., they frequently assign higher scores to subpar COTs, leading to potential limitations in their reasoning abilities. To address this problem, we introduce an \textit{Alignment Fine-Tuning (AFT)} paradigm, which involves three steps: 1) fine-tuning LLMs with COT training data; 2) generating multiple COT responses for each question, and categorizing them into positive and negative ones based on whether they achieve the correct answer; 3) calibrating the scores of positive and negative responses given by LLMs with a novel constraint alignment loss. Specifically, the constraint alignment loss has two objectives: a) Alignment, which guarantees that positive scores surpass negative scores to encourage answers with high-quality COTs; b) Constraint, which keeps the negative scores confined to a reasonable range to prevent the model degradation. Beyond just the binary positive and negative feedback, the constraint alignment loss can be seamlessly adapted to the ranking situations when ranking feedback is accessible. Furthermore, we also delve deeply into recent ranking-based alignment methods, such as DPO, RRHF, and PRO, and discover that the constraint, which has been overlooked by these approaches, is also crucial for their performance. Extensive experiments on four reasoning benchmarks with both binary and ranking feedback demonstrate the effectiveness of AFT.
Abstract:Recent years have witnessed the wide adoption of large language models (LLM) in different fields, especially natural language processing and computer vision. Such a trend can also be observed in recommender systems (RS). However, most of related work treat LLM as a component of the conventional recommendation pipeline (e.g., as a feature extractor) which may not be able to fully leverage the generative power of LLM. Instead of separating the recommendation process into multiple stages such as score computation and re-ranking, this process can be simplified to one stage with LLM: directly generating recommendations from the complete pool of items. This survey reviews the progress, methods and future directions of LLM-based generative recommendation by examining three questions: 1) What generative recommendation is, 2) Why RS should advance to generative recommendation, and 3) How to implement LLM-based generative recommendation for various RS tasks. We hope that the survey can provide the context and guidance needed to explore this interesting and emerging topic.
Abstract:Due to the unbalanced training data distribution, the language ability of large language models (LLMs) is often biased towards English. In this paper, we propose to empower pre-trained LLMs on non-English languages by building semantic alignment across languages. We perform instruction-tuning on LLaMA with both translation task data and cross-lingual general task data to obtain cross-lingual models (x-LLaMA). Experiment results on cross-lingual benchmark XQUAD and MLQA show that x-LLaMA models outperform the English instruction-tuned counterpart (Alpaca) by 42.50% on average on six non-English languages. Further experiments on Chinese benchmark C-Eval show that x-LLaMA achieves significant improvement on Chinese humanities tasks, outperforming Alpaca by 8.2%. We also discover that incorporating non-English text on the target side of translation data is particularly effective for boosting non-English ability. Besides, we find that semantic alignment within LLM can be further strengthened as translation task data scales up and we present the formulation of the underlying scaling law. Evaluation results on translation dataset Flores-101 show that \method outperforms previous LLaMA-based models in all evaluated directions. Code and data will be available at: https://github.com/OwenNJU/x-LLM.
Abstract:Myocardial infarction (MI) is one of the most prevalent cardiovascular diseases with associated clinical decision-making typically based on single-valued imaging biomarkers. However, such metrics only approximate the complex 3D structure and physiology of the heart and hence hinder a better understanding and prediction of MI outcomes. In this work, we investigate the utility of complete 3D cardiac shapes in the form of point clouds for an improved detection of MI events. To this end, we propose a fully automatic multi-step pipeline consisting of a 3D cardiac surface reconstruction step followed by a point cloud classification network. Our method utilizes recent advances in geometric deep learning on point clouds to enable direct and efficient multi-scale learning on high-resolution surface models of the cardiac anatomy. We evaluate our approach on 1068 UK Biobank subjects for the tasks of prevalent MI detection and incident MI prediction and find improvements of ~13% and ~5% respectively over clinical benchmarks. Furthermore, we analyze the role of each ventricle and cardiac phase for 3D shape-based MI detection and conduct a visual analysis of the morphological and physiological patterns typically associated with MI outcomes.