Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Batch Calibration: Rethinking Calibration for In-Context Learning and Prompt Engineering (2309.17249v3)

Published 29 Sep 2023 in cs.CL, cs.AI, and cs.LG

Abstract: Prompting and in-context learning (ICL) have become efficient learning paradigms for LLMs. However, LLMs suffer from prompt brittleness and various bias factors in the prompt, including but not limited to the formatting, the choice verbalizers, and the ICL examples. To address this problem that results in unexpected performance degradation, calibration methods have been developed to mitigate the effects of these biases while recovering LLM performance. In this work, we first conduct a systematic analysis of the existing calibration methods, where we both provide a unified view and reveal the failure cases. Inspired by these analyses, we propose Batch Calibration (BC), a simple yet intuitive method that controls the contextual bias from the batched input, unifies various prior approaches, and effectively addresses the aforementioned issues. BC is zero-shot, inference-only, and incurs negligible additional costs. In the few-shot setup, we further extend BC to allow it to learn the contextual bias from labeled data. We validate the effectiveness of BC with PaLM 2-(S, M, L) and CLIP models and demonstrate state-of-the-art performance over previous calibration baselines across more than 10 natural language understanding and image classification tasks.

Definition Search Book Streamline Icon: https://streamlinehq.com
References (73)
  1. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022.
  2. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023.
  3. anonymous. Survival of the most influential prompts: Efficient black-box prompt search via clustering and pruning. 2023.
  4. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020.
  5. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022.
  6. Describing textures in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014, pp.  3606–3613, 2014.
  7. BoolQ: Exploring the surprising difficulty of natural yes/no questions. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pp.  2924–2936, Minneapolis, Minnesota, June 2019. Association for Computational Linguistics.
  8. Decoder tuning: Efficient language understanding as decoding. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.  15072–15087, Toronto, Canada, July 2023. Association for Computational Linguistics.
  9. The commitmentbank: Investigating projection in naturally occurring discourse. In proceedings of Sinn und Bedeutung, volume 23, pp.  107–124, 2019.
  10. RLPrompt: Optimizing discrete text prompts with reinforcement learning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pp.  3369–3391, Abu Dhabi, United Arab Emirates, December 2022. Association for Computational Linguistics.
  11. Black-box prompt learning for pre-trained language models. Transactions on Machine Learning Research, 2023. ISSN 2835-8856.
  12. Automatically constructing a corpus of sentential paraphrases. In Proceedings of the Third International Workshop on Paraphrasing (IWP2005), 2005.
  13. Mitigating label biases for in-context learning. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.  14014–14031, Toronto, Canada, July 2023. Association for Computational Linguistics.
  14. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, pp.  1321–1330, 2017.
  15. Prototypical calibration for few-shot learning of language models. In The Eleventh International Conference on Learning Representations, 2023.
  16. Towards a unified view of parameter-efficient transfer learning. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022.
  17. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217–2226, 2019.
  18. Surface form competition: Why the highest probability answer isn’t always right. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pp.  7038–7051, Online and Punta Cana, Dominican Republic, November 2021. Association for Computational Linguistics.
  19. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, pp.  2790–2799, 2019.
  20. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022.
  21. Mixnorm: Test-time adaptation through online normalization estimation. arXiv preprint arXiv:2110.11478, 2021.
  22. CLEVR: A diagnostic dataset for compositional language and elementary visual reasoning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pp.  1988–1997, 2017.
  23. Sita: Single image test-time adaptation. arXiv preprint arXiv:2112.02355, 2021.
  24. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pp.  3045–3059, Online and Punta Cana, Dominican Republic, November 2021. Association for Computational Linguistics.
  25. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pp.  4582–4597, Online, August 2021. Association for Computational Linguistics.
  26. TTN: A domain-shift aware batch normalization in test-time adaptation. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023, 2023.
  27. What makes good in-context examples for GPT-3? In Proceedings of Deep Learning Inside Out (DeeLIO 2022): The 3rd Workshop on Knowledge Extraction and Integration for Deep Learning Architectures, pp.  100–114, Dublin, Ireland and Online, May 2022a. Association for Computational Linguistics.
  28. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1–35, 2023.
  29. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pp.  61–68, Dublin, Ireland, May 2022b. Association for Computational Linguistics.
  30. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.  8086–8098, Dublin, Ireland, May 2022. Association for Computational Linguistics.
  31. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013.
  32. Noisy channel language model prompting for few-shot text classification. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.  5316–5330, Dublin, Ireland, May 2022a. Association for Computational Linguistics.
  33. Rethinking the role of demonstrations: What makes in-context learning work? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pp.  11048–11064, Abu Dhabi, United Arab Emirates, December 2022b. Association for Computational Linguistics.
  34. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pp.  14745–14755, 2022.
  35. Todd K Moon. The expectation-maximization algorithm. IEEE Signal processing magazine, 13(6):47–60, 1996.
  36. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020.
  37. Adversarial NLI: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp.  4885–4901, Online, July 2020. Association for Computational Linguistics.
  38. Blackvip: Black-box visual prompting for robust transfer learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pp.  24224–24235, 2023.
  39. What in-context learning “learns” in-context: Disentangling task recognition and task learning. In Findings of the Association for Computational Linguistics: ACL 2023, pp.  8298–8319, Toronto, Canada, July 2023. Association for Computational Linguistics.
  40. Large language models sensitivity to the order of options in multiple-choice questions. arXiv preprint arXiv:2308.11483, 2023.
  41. WiC: the word-in-context dataset for evaluating context-sensitive meaning representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pp.  1267–1273, Minneapolis, Minnesota, June 2019. Association for Computational Linguistics.
  42. John Platt et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61–74, 1999.
  43. GrIPS: Gradient-free, edit-based instruction search for prompting large language models. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pp.  3845–3864, Dubrovnik, Croatia, May 2023. Association for Computational Linguistics.
  44. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pp.  8748–8763, 2021.
  45. Impact of pretraining term frequencies on few-shot numerical reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2022, pp.  840–854, Abu Dhabi, United Arab Emirates, December 2022. Association for Computational Linguistics.
  46. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In Logical Formalizations of Commonsense Reasoning, Papers from the 2011 AAAI Spring Symposium, Technical Report SS-11-06, Stanford, California, USA, March 21-23, 2011, 2011.
  47. Learning to retrieve prompts for in-context learning. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pp.  2655–2671, Seattle, United States, July 2022. Association for Computational Linguistics.
  48. Improving robustness against common corruptions by covariate shift adaptation. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020.
  49. On the effect of pretraining corpora on in-context learning by a large-scale language model. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pp.  5168–5186, Seattle, United States, July 2022. Association for Computational Linguistics.
  50. AutoPrompt: Eliciting Knowledge from Language Models with Automatically Generated Prompts. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp.  4222–4235, Online, November 2020. Association for Computational Linguistics.
  51. Measuring inductive biases of in-context learning with underspecified demonstrations. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.  11289–11310, Toronto, Canada, July 2023. Association for Computational Linguistics.
  52. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pp.  1631–1642, Seattle, Washington, USA, October 2013. Association for Computational Linguistics.
  53. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.
  54. An information-theoretic approach to prompt engineering without ground truth labels. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.  819–862, Dublin, Ireland, May 2022. Association for Computational Linguistics.
  55. Better zero-shot reasoning with self-adaptive prompting. In Findings of the Association for Computational Linguistics: ACL 2023, pp.  3493–3514, Toronto, Canada, July 2023a. Association for Computational Linguistics.
  56. Universal self-adaptive prompting. arXiv preprint arXiv:2305.14926, 2023b.
  57. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pp.  353–355, Brussels, Belgium, November 2018. Association for Computational Linguistics.
  58. Superglue: A stickier benchmark for general-purpose language understanding systems. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pp.  3261–3275, 2019.
  59. Tent: Fully test-time adaptation by entropy minimization. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021.
  60. Large language models are not fair evaluators. arXiv preprint arXiv:2305.17926, 2023.
  61. Larger language models do in-context learning differently. arXiv preprint arXiv:2303.03846, 2023.
  62. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pp.  1112–1122, New Orleans, Louisiana, June 2018. Association for Computational Linguistics.
  63. An explanation of in-context learning as implicit bayesian inference. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022.
  64. $k$nn prompting: Beyond-context learning with calibration-free nearest neighbor inference. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023, 2023.
  65. Guess the instruction! flipped learning makes language models stronger zero-shot learners. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023, 2023.
  66. Did you read the instructions? rethinking the effectiveness of task definitions in instruction learning. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp.  3063–3079, Toronto, Canada, July 2023. Association for Computational Linguistics.
  67. Test-time batch statistics calibration for covariate shift. arXiv preprint arXiv:2110.04065, 2021.
  68. Netzer Yuval. Reading digits in natural images with unsupervised feature learning. In Proceedings of the NIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011.
  69. TEMPERA: Test-time prompt editing via reinforcement learning. In The Eleventh International Conference on Learning Representations, 2023.
  70. Calibrate before use: Improving few-shot performance of language models. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pp.  12697–12706, 2021.
  71. On large language models’ selection bias in multi-choice questions. arXiv preprint arXiv:2309.03882, 2023.
  72. Autopeft: Automatic configuration search for parameter-efficient fine-tuning. arXiv preprint arXiv:2301.12132, 2023.
  73. Learning instance-specific adaptation for cross-domain segmentation. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pp.  459–476, 2022.
User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (7)
  1. Han Zhou (72 papers)
  2. Xingchen Wan (31 papers)
  3. Lev Proleev (6 papers)
  4. Diana Mincu (11 papers)
  5. Jilin Chen (32 papers)
  6. Katherine Heller (46 papers)
  7. Subhrajit Roy (24 papers)
Citations (36)
View on Semantic Scholar

Summary

Batch Calibration: Advancements in In-Context Learning and Prompt Engineering

The paper "Batch Calibration: Rethinking Calibration for In-Context Learning and Prompt Engineering" authored by Han Zhou et al., addresses a critical challenge in leveraging LLMs for natural language processing tasks: the inherent fragility and biases present in prompt-based learning. This work meticulously dissects existing calibration methodologies and introduces a novel approach named Batch Calibration (BC), which aims to alleviate biases in prompt engineering and in-context learning (ICL) with minimal computational overhead.

Context and Motivation

Prompting and in-context learning has emerged as efficient approaches to adapt LLMs for specific tasks by conditioning them with human-designed instructions. However, despite their utility, they are susceptible to various biases related to the format of prompts, choice of verbalizers, and selection of ICL examples. These biases result in significant performance variations, highlighting the necessity for robust calibration techniques. Conventional methods like Contextual Calibration (CC), Domain-Context Calibration (DC), and Prototypical Calibration (PC) have sought to address these biases but tend to fall short in providing consistent and holistic solutions across varied tasks.

Methodological Innovations

The primary proposition of this paper, Batch Calibration, operates on the principle of contextual bias reduction, achieved by marginalizing LLM scores over a batched input. This allows BC to function in a zero-shot inference regime, thus bypassing the need for additional labeled data and computational cost. Furthermore, BC extends its potential in few-shot learning scenarios by adapting a learnable parameter, henceforth named Black-box Few-shot Learning (BCL), to refine calibration through available labeled samples. This modularity and adaptability make BC a versatile addition to the arsenal of tools available for prompt engineering.

Empirical Evaluation

The authors conducted extensive experiments validating BC’s effectiveness against state-of-the-art calibration methods using datasets spanning over ten natural language understanding and image classification tasks. Utilizing PaLM 2 and CLIP models, BC demonstrated superior performance across configurations, underscoring its efficacy in mitigating prompt brittle biases. The results show statistically significant improvements in classification accuracy, consolidating BC as a robust methodology for enhancing LLM performance.

Implications and Future Directions

The introduction of BC has substantial implications for both practical applications and theoretical exploration in AI. It not only paves the way for more reliable usage of LLMs in industry settings but also provides a foundational framework for future studies focused on reducing bias in machine learning models. Additionally, extending BC to multi-modal learning contexts, such as in vision-LLMs like CLIP, reveals its potential applicability across modalities. This broad applicability suggests exciting avenues for research, particularly in exploring BC's benefits in generative tasks, potentially revolutionizing aspects of machine learning that rely heavily on context understanding.

Conclusion

Batch Calibration offers a streamlined, computationally efficient solution to prompt-induced biases, reinforcing the reliability and adaptability of LLMs and VLMs in diverse applications. Its introduction marks a meaningful stride toward contextually robust language and vision models, setting the stage for enhanced model generalization and more user-friendly prompt engineering practices. As researchers continue to explore the boundaries of LLM capabilities, methodologies like BC will undoubtedly play a critical role in defining the future landscape of AI-driven decision-making systems.

File Document Download Save Streamline Icon: https://streamlinehq.com PDF File Document Download Save Streamline Icon: https://streamlinehq.com Markdown
Youtube Logo Streamline Icon: https://streamlinehq.com

YouTube