Prometheus: Inducing Fine-grained Evaluation Capability in Language Models

Recently, using a powerful proprietary Large Language Model (LLM) (e.g., GPT-4) as an evaluator for long-form responses has become the de facto standard. However, for practitioners with large-scale evaluation tasks and custom criteria in consideration (e.g., child-readability), using proprietary LLMs as an evaluator is unreliable due to the closed-source nature, uncontrolled versioning, and prohibitive costs. In this work, we propose Prometheus, a fully open-source LLM that is on par with GPT-4's evaluation capabilities when the appropriate reference materials (reference answer, score rubric) are accompanied. We first construct the Feedback Collection, a new dataset that consists of 1K fine-grained score rubrics, 20K instructions, and 100K responses and language feedback generated by GPT-4. Using the Feedback Collection, we train Prometheus, a 13B evaluator LLM that can assess any given long-form text based on customized score rubric provided by the user. Experimental results show that Prometheus scores a Pearson correlation of 0.897 with human evaluators when evaluating with 45 customized score rubrics, which is on par with GPT-4 (0.882), and greatly outperforms ChatGPT (0.392). Furthermore, measuring correlation with GPT-4 with 1222 customized score rubrics across four benchmarks (MT Bench, Vicuna Bench, Feedback Bench, Flask Eval) shows similar trends, bolstering Prometheus's capability as an evaluator LLM. Lastly, Prometheus achieves the highest accuracy on two human preference benchmarks (HHH Alignment & MT Bench Human Judgment) compared to open-sourced reward models explicitly trained on human preference datasets, highlighting its potential as an universal reward model. We open-source our code, dataset, and model at https://github.com/kaistAI/Prometheus.

EvalLM: Interactive Evaluation of Large Language Model Prompts on User-Defined Criteria

By simply composing prompts, developers can prototype novel generative applications with Large Language Models (LLMs). To refine prototypes into products, however, developers must iteratively revise prompts by evaluating outputs to diagnose weaknesses. Formative interviews (N=8) revealed that developers invest significant effort in manually evaluating outputs as they assess context-specific and subjective criteria. We present EvalLM, an interactive system for iteratively refining prompts by evaluating multiple outputs on user-defined criteria. By describing criteria in natural language, users can employ the system's LLM-based evaluator to get an overview of where prompts excel or fail, and improve these based on the evaluator's feedback. A comparative study (N=12) showed that EvalLM, when compared to manual evaluation, helped participants compose more diverse criteria, examine twice as many outputs, and reach satisfactory prompts with 59% fewer revisions. Beyond prompts, our work can be extended to augment model evaluation and alignment in specific application contexts.

Who Wrote this Code? Watermarking for Code Generation

Large language models for code have recently shown remarkable performance in generating executable code. However, this rapid advancement has been accompanied by many legal and ethical concerns, such as code licensing issues, code plagiarism, and malware generation, making watermarking machine-generated code a very timely problem. Despite such imminent needs, we discover that existing watermarking and machine-generated text detection methods for LLMs fail to function with code generation tasks properly. Hence, in this work, we propose a new watermarking method, SWEET, that significantly improves upon previous approaches when watermarking machine-generated code. Our proposed method selectively applies watermarking to the tokens with high enough entropy, surpassing a defined threshold. The experiments on code generation benchmarks show that our watermarked code has superior quality compared to code produced by the previous state-of-the-art LLM watermarking method. Furthermore, our watermark method also outperforms DetectGPT for the task of machine-generated code detection.

The CoT Collection: Improving Zero-shot and Few-shot Learning of Language Models via Chain-of-Thought Fine-Tuning

Large Language Models (LLMs) have shown enhanced capabilities of solving novel tasks by reasoning step-by-step known as Chain-of-Thought (CoT) reasoning; how can we instill the same capability of reasoning step-by-step on unseen tasks into LMs that possess less than <100B parameters? To address this question, we first introduce the CoT Collection, a new instruction-tuning dataset that augments 1.88 million CoT rationales across 1,060 tasks. We show that continually fine-tuning Flan-T5 (3B & 11B) with the CoT Collection enables the 3B & 11B LMs to perform CoT better on unseen tasks, leading to an improvement in the average zero-shot accuracy on 27 datasets of the BIG-Bench-Hard benchmark by +4.34% and +2.44%, respectively. Furthermore, we show that instruction tuning with CoT allows LMs to possess stronger few-shot learning capabilities, resulting in an improvement of +2.97% and +2.37% on 4 domain-specific tasks over Flan-T5 (3B & 11B), respectively. We make our CoT Collection data and our trained models publicly available at https://github.com/kaist-lklab/CoT-Collection.

Revealing User Familiarity Bias in Task-Oriented Dialogue via Interactive Evaluation

Most task-oriented dialogue (TOD) benchmarks assume users that know exactly how to use the system by constraining the user behaviors within the system's capabilities via strict user goals, namely "user familiarity" bias. This data bias deepens when it combines with data-driven TOD systems, as it is impossible to fathom the effect of it with existing static evaluations. Hence, we conduct an interactive user study to unveil how vulnerable TOD systems are against realistic scenarios. In particular, we compare users with 1) detailed goal instructions that conform to the system boundaries (closed-goal) and 2) vague goal instructions that are often unsupported but realistic (open-goal). Our study reveals that conversations in open-goal settings lead to catastrophic failures of the system, in which 92% of the dialogues had significant issues. Moreover, we conduct a thorough analysis to identify distinctive features between the two settings through error annotation. From this, we discover a novel "pretending" behavior, in which the system pretends to handle the user requests even though they are beyond the system's capabilities. We discuss its characteristics and toxicity while emphasizing transparency and a fallback strategy for robust TOD systems.

Aligning Large Language Models through Synthetic Feedback

Aligning large language models (LLMs) to human values has become increasingly important as it enables sophisticated steering of LLMs, e.g., making them follow given instructions while keeping them less toxic. However, it requires a significant amount of human demonstrations and feedback. Recently, open-sourced models have attempted to replicate the alignment learning process by distilling data from already aligned LLMs like InstructGPT or ChatGPT. While this process reduces human efforts, constructing these datasets has a heavy dependency on the teacher models. In this work, we propose a novel framework for alignment learning with almost no human labor and no dependency on pre-aligned LLMs. First, we perform reward modeling (RM) with synthetic feedback by contrasting responses from vanilla LLMs with various sizes and prompts. Then, we use the RM for simulating high-quality demonstrations to train a supervised policy and for further optimizing the model with reinforcement learning. Our resulting model, Aligned Language Model with Synthetic Training dataset (ALMoST), outperforms open-sourced models, including Alpaca, Dolly, and OpenAssistant, which are trained on the outputs of InstructGPT or human-annotated instructions. Our 7B-sized model outperforms the 12-13B models in the A/B tests using GPT-4 as the judge with about 75% winning rate on average.

Towards Zero-Shot Functional Compositionality of Language Models

Large Pre-trained Language Models (PLM) have become the most desirable starting point in the field of NLP, as they have become remarkably good at solving many individual tasks. Despite such success, in this paper, we argue that current paradigms of working with PLMs are neglecting a critical aspect of modeling human intelligence: functional compositionality. Functional compositionality - the ability to compose learned tasks - has been a long-standing challenge in the field of AI (and many other fields) as it is considered one of the hallmarks of human intelligence. An illustrative example of such is cross-lingual summarization, where a bilingual person (English-French) could directly summarize an English document into French sentences without having to translate the English document or summary into French explicitly. We discuss why this matter is an important open problem that requires further attention from the field. Then, we show that current PLMs (e.g., GPT-2 and T5) don't have functional compositionality yet and it is far from human-level generalizability. Finally, we suggest several research directions that could push the field towards zero-shot functional compositionality of language models.

Evaluating the Knowledge Dependency of Questions

The automatic generation of Multiple Choice Questions (MCQ) has the potential to reduce the time educators spend on student assessment significantly. However, existing evaluation metrics for MCQ generation, such as BLEU, ROUGE, and METEOR, focus on the n-gram based similarity of the generated MCQ to the gold sample in the dataset and disregard their educational value. They fail to evaluate the MCQ's ability to assess the student's knowledge of the corresponding target fact. To tackle this issue, we propose a novel automatic evaluation metric, coined Knowledge Dependent Answerability (KDA), which measures the MCQ's answerability given knowledge of the target fact. Specifically, we first show how to measure KDA based on student responses from a human survey. Then, we propose two automatic evaluation metrics, KDA_disc and KDA_cont, that approximate KDA by leveraging pre-trained language models to imitate students' problem-solving behavior. Through our human studies, we show that KDA_disc and KDA_soft have strong correlations with both (1) KDA and (2) usability in an actual classroom setting, labeled by experts. Furthermore, when combined with n-gram based similarity metrics, KDA_disc and KDA_cont are shown to have a strong predictive power for various expert-labeled MCQ quality measures.

Dialogue Summaries as Dialogue States (DS2), Template-Guided Summarization for Few-shot Dialogue State Tracking

Annotating task-oriented dialogues is notorious for the expensive and difficult data collection process. Few-shot dialogue state tracking (DST) is a realistic solution to this problem. In this paper, we hypothesize that dialogue summaries are essentially unstructured dialogue states; hence, we propose to reformulate dialogue state tracking as a dialogue summarization problem. To elaborate, we train a text-to-text language model with synthetic template-based dialogue summaries, generated by a set of rules from the dialogue states. Then, the dialogue states can be recovered by inversely applying the summary generation rules. We empirically show that our method DS2 outperforms previous works on few-shot DST in MultiWoZ 2.0 and 2.1, in both cross-domain and multi-domain settings. Our method also exhibits vast speedup during both training and inference as it can generate all states at once. Finally, based on our analysis, we discover that the naturalness of the summary templates plays a key role for successful training.

Pedagogical Word Recommendation: A novel task and dataset on personalized vocabulary acquisition for L2 learners

When learning a second language (L2), one of the most important but tedious components that often demoralizes students with its ineffectiveness and inefficiency is vocabulary acquisition, or more simply put, memorizing words. In light of such, a personalized and educational vocabulary recommendation system that traces a learner's vocabulary knowledge state would have an immense learning impact as it could resolve both issues. Therefore, in this paper, we propose and release data for a novel task called Pedagogical Word Recommendation (PWR). The main goal of PWR is to predict whether a given learner knows a given word based on other words the learner has already seen. To elaborate, we collect this data via an Intelligent Tutoring System (ITS) that is serviced to ~1M L2 learners who study for the standardized English exam, TOEIC. As a feature of this ITS, students can directly indicate words they do not know from the questions they solved to create wordbooks. Finally, we report the evaluation results of a Neural Collaborative Filtering approach along with an exploratory data analysis and discuss the impact and efficacy of this dataset as a baseline for future studies on this task.