Answer-level Calibration for Free-form Multiple Choice Question Answering (2022.acl-long)
Copied to clipboard
| Challenge: | Pre-trained language models exhibit few-shot and zero-shot learning capability on tasks such as commonsense reasoning. |
| Approach: | They propose to model context-independent biases in terms of the probability of a choice without the context and to remove it using an unsupervised estimate of similarity with the full context. |
| Outcome: | The proposed model improves over baselines on commonsense reasoning tasks. |
Similar Papers
Addressing Blind Guessing: Calibration of Selection Bias in Multiple-Choice Question Answering by Video Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Existing MCQA benchmarks fail to capture the full reasoning capabilities of video language models due to selection bias. |
| Approach: | They propose a method to reduce selection bias in video-to-text LLMs by suppressing "blind guessing" they propose 'bold' calibration technique to balance selection bias. |
| Outcome: | The proposed method reduces selection bias and improves model performance compared to existing methods. |
Calibrating Zero-shot Cross-lingual (Un-)structured Predictions (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing need for model calibration when natural language models are deployed in critical tasks. |
| Approach: | They compare model calibration methods in a context of zero-shot cross-lingual transfer with pre-trained language models. |
| Outcome: | The proposed method fails to calibrate more complex confidence estimations in structured predictions compared to expressive alternatives like Gaussian Process Calibration. |
Knowing More About Questions Can Help: Improving Calibration in Question Answering (2021.findings-acl)
Copied to clipboard
| Challenge: | Existing work on calibration focuses on model confidence, such as the max probability of the predicted class. |
| Approach: | They propose a calibration method which estimates whether model correctly predicts answer for each question. |
| Outcome: | The proposed calibration method achieves 5-10% gains on reading comprehension benchmarks. |
How Can We Know When Language Models Know? On the Calibration of Language Models for Question Answering (2021.tacl-1)
Copied to clipboard
| Challenge: | Recent studies have shown that language models capture different types of knowledge regarding facts or commonsense knowledge. |
| Approach: | They examine how language models can be calibrated to make their confidence scores correlate better with the likelihood of correctness. |
| Outcome: | The proposed calibration methods improve confidence scores on QA tasks and improve accuracy. |
Calibrating LLM-Based Evaluator (2024.lrec-main)
Copied to clipboard
Yuxuan Liu, Tianchi Yang, Shaohan Huang, Zihan Zhang, Haizhen Huang, Furu Wei, Weiwei Deng, Feng Sun, Qi Zhang
| Challenge: | Existing models for large language models lack the ability to calibrate their outputs towards human preference. |
| Approach: | They propose a multi-stage, gradient-free approach to calibrate an LLM-based evaluator toward human preference. |
| Outcome: | The proposed approach improves correlation with expert evaluation on multiple text quality evaluation datasets. |
Task Calibration: Calibrating Large Language Models on Inference Tasks (2025.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have shown impressive zero-shot performance on inference tasks, however, they may suffer from spurious correlations between input texts and output labels, which limits their ability to reason based purely on general language understanding. |
| Approach: | They propose a zero-shot and inference-only calibration method inspired by mutual information which recovers LLM performance through task reformulation. |
| Outcome: | The proposed calibration method improves on 13 benchmarks and prompt templates and can be integrated with other calibration methods. |
Language Model Probabilities are Not Calibrated in Numeric Contexts (2025.acl-long)
Copied to clipboard
Charles Lovering, Michael Krumdick, Viet Dac Lai, Varshini Reddy, Seth Ebner, Nilesh Kumar, Rik Koncel-Kedziorski, Chris Tanner
| Challenge: | Using language model outputs, we find that even in simple settings, the best LMs (1) are poorly calibrated and (2) have systematic biases. |
| Approach: | They argue that language model outputs should capture natural distributions over multiple options within their textual contexts. |
| Outcome: | The proposed model outputs are calibrated to the numeric content of their contexts. |
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. |
Calibrating Long-form Generations From Large Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Conventional calibration methods treat answer correctness as binary and do not work for long-form generation where an answer can be partially correct. |
| Approach: | They propose a framework where correctness of LLMs' responses and associated confidence levels are treated as distributions across a range of scores. |
| Outcome: | The proposed framework treats the correctness of the LLMs’ responses and their associated confidence levels as distributions across a range of scores. |
Rethinking Pragmatics in Large Language Models: Towards Open-Ended Evaluation and Preference Tuning (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to assess social-pragmatic inference in large language models are inadequacy, and preferential tuning is the best approach. |
| Approach: | They propose to use free-form models' responses as a measure to assess social-pragmatic reasoning and advocate for preference optimization over supervised finetuning (SFT). |
| Outcome: | The proposed model outperforms supervised finetuning (SFT) and offers a near-free launch in pragmatic abilities without compromising general capabilities. |