Papers by Chelsea Finn
Enhancing Self-Consistency and Performance of Pre-Trained Language Models through Natural Language Inference (2022.emnlp-main)
Copied to clipboard
Eric Mitchell, Joseph Noh, Siyan Li, Will Armstrong, Ananth Agarwal, Patrick Liu, Chelsea Finn, Christopher Manning
| Challenge: | Existing frameworks for boosting consistency and accuracy of pre-trained NLP models without fine-tuning or re-training are lacking. |
| Approach: | They propose a framework for boosting the consistency and accuracy of pre-trained NLP models using pre-trainer natural language inference models without fine-tuning or re-training. |
| Outcome: | The proposed framework boosts consistency and accuracy of pre-trained NLP models using pre-train natural language inference models without fine-tuning or re-training. |
Calibrating Language Models with Adaptive Temperature Scaling (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large language models' confidence scores are degraded after fine-tuning with reinforcement learning from human feedback. |
| Approach: | They propose a post-hoc calibration method that predicts a temperature scaling parameter for each token prediction. |
| Outcome: | Adaptive temperature scaling improves calibration by over 10% compared to prior methods . RLHF fine-tuning improves model accuracy, but degradation is not significant . |
Meta-Learning Online Adaptation of Language Models (2023.emnlp-main)
Copied to clipboard
| Challenge: | Large language models encode impressively broad world knowledge in their parameters, but the knowledge in static models becomes stale. |
| Approach: | They propose to use a small, autoregressive model to reweight the language modeling loss for each token during online fine-tuning to maximize the out-of-date model’s ability to answer questions about a document. |
| Outcome: | The proposed approach improves the model's ability to answer questions on streams of thousands of documents compared with fine-tuning and baseline heuristics. |
Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback (2023.emnlp-main)
Copied to clipboard
Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, Christopher Manning
| Challenge: | Recent studies have shown that unsupervised pre-training produces large language models whose conditional probabilities are remarkably well-calibrated. |
| Approach: | They propose to use verbalized confidences to extract confidence from large language models with reinforcement learning from human feedback to improve their accuracy. |
| Outcome: | The proposed methods reduce the expected calibration error by 50% for RLHF-LMs such as ChatGPT, GPT-4, and Claude. |
Disentangling Length from Quality in Direct Preference Optimization (2024.findings-acl)
Copied to clipboard
| Challenge: | Reinforcement Learning from Human Feedback (RLHF) exploits biases in human preferences, such as verbosity, and is under-explored for Direct Alignment Algorithms such as DPO. |
| Approach: | They propose to exploit verbosity biases in RLHF by using direct preference optimization to fine-tune models. |
| Outcome: | The proposed approach achieves 20% improvement in win rates when controlling for length on summarization and dialogue datasets, despite the GPT4 judge’s verbosity bias. |
PERSONA: A Reproducible Testbed for Pluralistic Alignment (2025.coling-main)
Copied to clipboard
| Challenge: | Currently, preference optimization approaches fail to capture the plurality of user opinions . Currently used methods do not account for the pluralities of users and difference of opinion . |
| Approach: | They propose a reproducible test bed to evaluate pluralistic alignment of language models . they generate user profiles from census data and use a large-scale evaluation dataset . |
| Outcome: | The proposed model improves pluralistic alignment of language models with diverse user values . it generates a large-scale evaluation dataset with 317,200 feedback pairs . |