An official website of the United States government
Language models (LMs) are pretrained to imitate text from large and diverse datasets that contain content that would violate human preferences if generated by an LM: falsehoods, offensive comments, personally identifiable information, low-quality or buggy code, among others. Here, we explore alternative objectives for pretraining LMs in a way that also guides them to generate text aligned with human preferences. We benchmark five objectives for pretraining with human feedback across three tasks and study how they affect the alignment and capabilities of pretrained LMs. We find a Pareto-optimal and simple approach among those we explored: conditional training, or learning distribution over tokens conditional on their human preference scores. Conditional training reduces the rate of undesirable content by up to an order of magnitude, both when generating without a prompt and with an adversarially-chosen prompt. Moreover, conditional training maintains the downstream task performance of standard LM pretraining, both before and after task-specific finetuning. Pretraining with human feedback results in much better preference satisfaction than standard LM pretraining followed by finetuning with feedback, i.e., learning and then unlearning undesirable behavior. Our results suggest that we should move beyond imitation learning when pretraining LMs and incorporate human preferences from the start of training.
more »
« less
This content will become publicly available on October 1, 2026
Exploring Human and Language Model Alignment in Perceived Design Similarity Using Ordinal Embeddings
Abstract In recent years, large language models (LLMs) and vision language models (VLMs) have excelled at tasks requiring human-like reasoning, inspiring researchers in engineering design to use language models (LMs) as surrogate evaluators of design concepts. But do these models actually evaluate designs like humans? While recent work has shown that LM evaluations sometimes fall within human variance on Likert-scale grading tasks, those tasks often obscure the reasoning and biases behind the scores. To address this limitation, we compare LM word embeddings (trained to capture semantic similarity) with human-rated similarity embeddings derived from triplet comparisons (“is A closer to B than C?”) on a dataset of design sketches and descriptions. We assess alignment via local tripletwise similarity and embedding distances, allowing for deeper insights than raw Likert-scale scores provide. We also explore whether describing the designs to LMs through text or images improves alignment with human judgments. Our findings suggest that text alone may not fully capture the nuances humans key into, yet text-based embeddings outperform their multimodal counterparts on satisfying local triplets. On the basis of these insights, we offer recommendations for effectively integrating LMs into design evaluation tasks.
more »
« less
- Award ID(s):
- 2155072
- PAR ID:
- 10651764
- Publisher / Repository:
- The American Society of Mechanical Engineers
- Date Published:
- Journal Name:
- Journal of Mechanical Design
- Volume:
- 147
- Issue:
- 10
- ISSN:
- 1050-0472
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Language models (LMs) are pretrained to imitate internet text, including content that would violate human preferences if generated by an LM: falsehoods, offensive comments, personally identifiable information, low-quality or buggy code, and more. Here, we explore alternative objectives for pretraining LMs in a way that also guides them to generate text aligned with human preferences. We benchmark five objectives for pretraining with human feedback across three tasks and study how they affect the trade-off between alignment and capabilities of pretrained LMs. We find a Pareto optimal and simple approach among those we explored: conditional training, or learning distribution over tokens conditional on their human preference scores given by a reward model. Conditional training reduces the rate of undesirable content by up to an order of magnitude, both when generating without a prompt and with an adversarially chosen prompt. Moreover, conditional training maintains the downstream task performance of standard LM pretraining, both before and after task-specific finetuning. Pretraining with human feedback results in much better preference satisfaction than standard LM pretraining followed by finetuning with feedback, i.e., learning and then unlearning undesirable behavior. Our results suggest that we should move beyond imitation learning when pretraining LMs and incorporate human preferences from the start of training.more » « less
-
Abstract Well-studied techniques that enhance diversity in early design concept generation require effective metrics for evaluating human-perceived similarity between ideas. Recent work suggests collecting triplet comparisons between designs directly from human raters and using those triplets to form an embedding where similarity is expressed as a Euclidean distance. While effective at modeling human-perceived similarity judgments, these methods are expensive and require a large number of triplets to be hand-labeled. However, what if there were a way to use AI to replicate the human similarity judgments captured in triplet embedding methods? In this paper, we explore the potential for pretrained Large Language Models (LLMs) to be used in this context. Using a dataset of crowdsourced text descriptions written about engineering design sketches, we generate LLM embeddings and compare them to an embedding created from human-provided triplets of those same sketches. From these embeddings, we can use Euclidean distances to describe areas where human perception and LLM perception disagree regarding design similarity. We then implement this same procedure but with descriptions written from a template that attempts to isolate a particular modality of a design (i.e., functions, behaviors, structures). By comparing the templated description embeddings to both the triplet-generated and pre-template LLM embeddings, we identify ways of describing designs such that LLM and human similarity perception better agree. We use these results to better understand how humans and LLMs interpret similarity in engineering designs.more » « less
-
Prompting has shown impressive success in enabling large pre-trained language models (LMs) to perform diverse NLP tasks, especially with only few downstream data. Automatically finding the optimal prompt for each task, however, is challenging. Most existing work resorts to tuning *soft* prompts (e.g., embeddings) which fall short of interpretability, reusability across LMs, and applicability when gradients are not accessible. *Discrete* prompts, on the other hand, are difficult to optimize, and are often created by “enumeration (e.g., paraphrasing)-then-selection” heuristics that do not explore the prompt space systematically. This paper proposes RLPrompt, an efficient discrete prompt optimization approach with reinforcement learning (RL). RLPrompt formulates a parameter-efficient policy network that generates the optimized discrete prompt after training with reward. To harness the complex and stochastic reward signals from the large LM environment, we incorporate effective reward stabilization that substantially enhances training efficiency. RLPrompt is flexibly applicable to different types of LMs, such as masked (e.g., BERT) and left-to-right models (e.g., GPTs), for both classification and generation tasks. Experiments on few-shot classification and unsupervised text style transfer show superior performance over a wide range of existing fine-tuning or prompting methods. Interestingly, the resulting optimized prompts are often ungrammatical gibberish text; and surprisingly, those gibberish prompts are transferrable between different LMs to retain significant performance, indicating that LM prompting may not follow human language patterns.more » « less
-
Moens, Marie-Francine; Huang, Xuanjing; Specia, Lucia; Yih, Scott Wen-tau (Ed.)Pre-trained Transformer language models (LM) have become go-to text representation encoders. Prior research fine-tunes deep LMs to encode text sequences such as sentences and passages into single dense vector representations for efficient text comparison and retrieval. However, dense encoders require a lot of data and sophisticated techniques to effectively train and suffer in low data situations. This paper finds a key reason is that standard LMs’ internal attention structure is not ready-to-use for dense encoders, which needs to aggregate text information into the dense representation. We propose to pre-train towards dense encoder with a novel Transformer architecture, Condenser, where LM prediction CONditions on DENSE Representation. Our experiments show Condenser improves over standard LM by large margins on various text retrieval and similarity tasks.more » « less
