ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools

We introduce ChatGLM, an evolving family of large language models that we have been developing over time. This report primarily focuses on the GLM-4 language series, which includes GLM-4, GLM-4-Air, and GLM-4-9B. They represent our most capable models that are trained with all the insights and lessons gained from the preceding three generations of ChatGLM. To date, the GLM-4 models are pre-trained on ten trillions of tokens mostly in Chinese and English, along with a small set of corpus from 24 languages, and aligned primarily for Chinese and English usage. The high-quality alignment is achieved via a multi-stage post-training process, which involves supervised fine-tuning and learning from human feedback. Evaluations show that GLM-4 1) closely rivals or outperforms GPT-4 in terms of general metrics such as MMLU, GSM8K, MATH, BBH, GPQA, and HumanEval, 2) gets close to GPT-4-Turbo in instruction following as measured by IFEval, 3) matches GPT-4 Turbo (128K) and Claude 3 for long context tasks, and 4) outperforms GPT-4 in Chinese alignments as measured by AlignBench. The GLM-4 All Tools model is further aligned to understand user intent and autonomously decide when and which tool(s) touse -- including web browser, Python interpreter, text-to-image model, and user-defined functions -- to effectively complete complex tasks. In practical applications, it matches and even surpasses GPT-4 All Tools in tasks like accessing online information via web browsing and solving math problems using Python interpreter. Over the course, we have open-sourced a series of models, including ChatGLM-6B (three generations), GLM-4-9B (128K, 1M), GLM-4V-9B, WebGLM, and CodeGeeX, attracting over 10 million downloads on Hugging face in the year 2023 alone. The open models can be accessed through https://github.com/THUDM and https://huggingface.co/THUDM.

Trial and Error: Exploration-Based Trajectory Optimization for LLM Agents

Large Language Models (LLMs) have become integral components in various autonomous agent systems. In this study, we present an exploration-based trajectory optimization approach, referred to as ETO. This learning method is designed to enhance the performance of open LLM agents. Contrary to previous studies that exclusively train on successful expert trajectories, our method allows agents to learn from their exploration failures. This leads to improved performance through an iterative optimization framework. During the exploration phase, the agent interacts with the environment while completing given tasks, gathering failure trajectories to create contrastive trajectory pairs. In the subsequent training phase, the agent utilizes these trajectory preference pairs to update its policy using contrastive learning methods like DPO. This iterative cycle of exploration and training fosters continued improvement in the agents. Our experiments on three complex tasks demonstrate that ETO consistently surpasses baseline performance by a large margin. Furthermore, an examination of task-solving efficiency and potential in scenarios lacking expert trajectory underscores the effectiveness of our approach.

PST-Bench: Tracing and Benchmarking the Source of Publications

Tracing the source of research papers is a fundamental yet challenging task for researchers. The billion-scale citation relations between papers hinder researchers from understanding the evolution of science efficiently. To date, there is still a lack of an accurate and scalable dataset constructed by professional researchers to identify the direct source of their studied papers, based on which automatic algorithms can be developed to expand the evolutionary knowledge of science. In this paper, we study the problem of paper source tracing (PST) and construct a high-quality and ever-increasing dataset PST-Bench in computer science. Based on PST-Bench, we reveal several intriguing discoveries, such as the differing evolution patterns across various topics. An exploration of various methods underscores the hardness of PST-Bench, pinpointing potential directions on this topic. The dataset and codes have been available at https://github.com/THUDM/paper-source-trace.

Lumos: Learning Agents with Unified Data, Modular Design, and Open-Source LLMs

We introduce Lumos, a novel framework for training language agents that employs a unified data format and a modular architecture based on open-source large language models (LLMs). Lumos consists of three distinct modules: planning, grounding, and execution. The planning module breaks down a task into a series of high-level, tool-agnostic subgoals, which are then made specific by the grounding module through a set of low-level actions. These actions are subsequently executed by the execution module, utilizing a range of off-the-shelf tools and APIs. In order to train these modules effectively, high-quality annotations of subgoals and actions were collected and are made available for fine-tuning open-source LLMs for various tasks such as complex question answering, web tasks, and math problems. Leveraging this unified data and modular design, Lumos not only achieves comparable or superior performance to current, state-of-the-art agents, but also exhibits several key advantages: (1) Lumos surpasses GPT-4/3.5-based agents in complex question answering and web tasks, while equalling the performance of significantly larger LLM agents on math tasks; (2) Lumos outperforms open-source agents created through conventional training methods and those using chain-of-thoughts training; and (3) Lumos is capable of effectively generalizing to unseen interactive tasks, outperforming larger LLM-based agents and even exceeding performance of specialized agents.

Will the Prince Get True Love's Kiss? On the Model Sensitivity to Gender Perturbation over Fairytale Texts

Recent studies show that traditional fairytales are rife with harmful gender biases. To help mitigate these gender biases in fairytales, this work aims to assess learned biases of language models by evaluating their robustness against gender perturbations. Specifically, we focus on Question Answering (QA) tasks in fairytales. Using counterfactual data augmentation to the FairytaleQA dataset, we evaluate model robustness against swapped gender character information, and then mitigate learned biases by introducing counterfactual gender stereotypes during training time. We additionally introduce a novel approach that utilizes the massive vocabulary of language models to support text genres beyond fairytales. Our experimental results suggest that models are sensitive to gender perturbations, with significant performance drops compared to the original testing set. However, when first fine-tuned on a counterfactual training dataset, models are less sensitive to the later introduced anti-gender stereotyped text.

The Magic of IF: Investigating Causal Reasoning Abilities in Large Language Models of Code

Causal reasoning, the ability to identify cause-and-effect relationship, is crucial in human thinking. Although large language models (LLMs) succeed in many NLP tasks, it is still challenging for them to conduct complex causal reasoning like abductive reasoning and counterfactual reasoning. Given the fact that programming code may express causal relations more often and explicitly with conditional statements like ``if``, we want to explore whether Code-LLMs acquire better causal reasoning abilities. Our experiments show that compared to text-only LLMs, Code-LLMs with code prompts are significantly better in causal reasoning. We further intervene on the prompts from different aspects, and discover that the programming structure is crucial in code prompt design, while Code-LLMs are robust towards format perturbations.

Dynosaur: A Dynamic Growth Paradigm for Instruction-Tuning Data Curation

Instruction tuning has emerged to enhance the capabilities of large language models (LLMs) in providing appropriate outputs based on input instructions. However, existing methods for collecting instruction-tuning data suffer from limitations in scalability and affordability. In this paper, we propose Dynosaur, a dynamic growth paradigm for instruction-tuning data curation. Built upon the metadata of existing NLP datasets, we generate multiple task instructions applicable to various NLP datasets and determine the relevant data fields for constructing instruction-tuning data with LLMs. Dynosaur offers several advantages: 1) lower generation costs (less than $12 for generating 800K instruction-tuning data), 2) good quality of instruction-tuning data (better performance than Alpaca and Instruction GPT-4 on Super-NI with comparable data sizes), and 3) the ability to grow dynamically by incorporating new datasets from Huggingface Datasets Platform. We further investigate continual learning as an approach to learning with the ever-growing instruction-tuning dataset. We demonstrate that replay methods not only help mitigate forgetting issues but help generalize to unseen tasks better. As a novel continual learning scenario for instruction tuning, selecting tasks based on instruction representations can be an effective replaying strategy. Code and data are released at \url{https://github.com/WadeYin9712/Dynosaur}.

KPEval: Towards Fine-grained Semantic-based Evaluation of Keyphrase Extraction and Generation Systems

Despite the significant advancements in keyphrase extraction and keyphrase generation methods, the predominant approach for evaluation only relies on exact matching with human references and disregards reference-free attributes. This scheme fails to recognize systems that generate keyphrases that are semantically equivalent to the references or keyphrases that have practical utility. To better understand the strengths and weaknesses of different keyphrase systems, we propose a comprehensive evaluation framework consisting of six critical dimensions: naturalness, faithfulness, saliency, coverage, diversity, and utility. For each dimension, we discuss the desiderata and design semantic-based metrics that align with the evaluation objectives. Rigorous meta-evaluation studies demonstrate that our evaluation strategy correlates better with human preferences compared to a range of previously used metrics. Using this framework, we re-evaluate 18 keyphrase systems and further discover that (1) the best model differs in different dimensions, with pre-trained language models achieving the best in most dimensions; (2) the utility in downstream tasks does not always correlate well with reference-based metrics; and (3) large language models exhibit a strong performance in reference-free evaluation.

GIVL: Improving Geographical Inclusivity of Vision-Language Models with Pre-Training Methods

A key goal for the advancement of AI is to develop technologies that serve the needs not just of one group but of all communities regardless of their geographical region. In fact, a significant proportion of knowledge is locally shared by people from certain regions but may not apply equally in other regions because of cultural differences. If a model is unaware of regional characteristics, it may lead to performance disparity across regions and result in bias against underrepresented groups. We propose GIVL, a Geographically Inclusive Vision-and-Language Pre-trained model. There are two attributes of geo-diverse visual concepts which can help to learn geo-diverse knowledge: 1) concepts under similar categories have unique knowledge and visual characteristics, 2) concepts with similar visual features may fall in completely different categories. Motivated by the attributes, we design new pre-training objectives Image Knowledge Matching (IKM) and Image Edit Checking (IEC) to pre-train GIVL. Compared with similar-size models pre-trained with similar scale of data, GIVL achieves state-of-the-art (SOTA) and more balanced performance on geo-diverse V&L tasks.

How well can Text-to-Image Generative Models understand Ethical Natural Language Interventions?

Text-to-image generative models have achieved unprecedented success in generating high-quality images based on natural language descriptions. However, it is shown that these models tend to favor specific social groups when prompted with neutral text descriptions (e.g., 'a photo of a lawyer'). Following Zhao et al. (2021), we study the effect on the diversity of the generated images when adding ethical intervention that supports equitable judgment (e.g., 'if all individuals can be a lawyer irrespective of their gender') in the input prompts. To this end, we introduce an Ethical NaTural Language Interventions in Text-to-Image GENeration (ENTIGEN) benchmark dataset to evaluate the change in image generations conditional on ethical interventions across three social axes -- gender, skin color, and culture. Through ENTIGEN framework, we find that the generations from minDALL.E, DALL.E-mini and Stable Diffusion cover diverse social groups while preserving the image quality. Preliminary studies indicate that a large change in the model predictions is triggered by certain phrases such as 'irrespective of gender' in the context of gender bias in the ethical interventions. We release code and annotated data at https://github.com/Hritikbansal/entigen_emnlp.