Calibrate Before Use: Improving Few-Shot Performance of Language Models

1. Introduction

• GPT-3 performs numerous tasks from a few examples in natural-language prompts (Brown et al., 2020), without updating model parameters.This is the paper’s in-context few-shot learning setting.Source: §1 · p. 1
• In-context learning enables rapid prototyping and natural-language system creation by nontechnical users, while model reuse reduces memory requirements and serving complexity across tasks.Changing the prompt immediately produces a new model behavior.Source: §1 · p. 1
• 54% to 93% sentiment-analysis accuracy results from merely reordering training examples, illustrating GPT-3’s sensitivity to prompt format, example selection, and ordering.This instability means manually designed prompts cannot reliably deliver good accuracy.Source: §1 · p. 1
• Majority-label, recency, and common-token biases favor frequent training answers, answers near the prompt’s end, and answers common in pretraining, respectively.These biases typically shift the model’s output distribution.Source: §1 · p. 1
• Up to 30.0% absolute accuracy improvement accompanies consistently better GPT-3 and GPT-2 accuracy and reduced variability across prompt formats and examples.Greater stability mitigates the need for prompt engineering.Source: Figure 1 · p. 2, §1 · p. 2
• Contextual calibration estimates answer biases using content-free inputs, then fits parameters to make those inputs’ answer scores uniform without additional training data.A dummy test input such as “N/A” reveals the model’s preference among answers.Source: §1 · p. 2

2. Background and Experimental Setup

• In-context learning prompts combine a format, training examples, and their ordering, which matters because language models update hidden states left to right.Training examples demonstrate how to solve the task; zero-shot prompts contain none.Source: §2 · p. 2

2.1. Datasets and Prompt Formats

• Six text-classification datasets cover sentiment analysis, question classification, textual entailment, and topic classification; each dataset uses a fixed prompt format unless otherwise specified.Classification uses SST-2, 6-way TREC, 3-way CB, and binary RTE, with CB and RTE drawn from SuperGLUE (Wang et al., 2019). Topic classification uses 4-way AGNews (Zhang et al., 2015) and 14-way DBPedia (Zhang et al., 2015).Source: §2.1 · pp. 2–3

2. Background and Experimental Setup

• GPT-3 2.7B shows high accuracy variance on SST-2 when training-example sets, example order, or prompt formats change.Figure 2 varies permutations within ten sets of four examples; Figure 3 compares ten formats using different sets of four examples.Source: Figure 2 · p. 3, Figure 3 · p. 3

2.1. Datasets and Prompt Formats

• Fact retrieval uses LAMA, retaining templates whose missing answers occur at the end and reporting average accuracy across all triples.Knowledge-base triples become prompts with missing objects; answers are always single tokens.Source: §2.1 · p. 3
• ATIS and MIT Movies trivia10k13 evaluate slot filling using Exact Match between generated outputs and ground-truth spans.The selected slots are airline and departure date for ATIS, and director name and movie genre for MIT Movies.Source: §2.1 · p. 3

2.2. Model Details

• Experiments use three GPT-3 sizes and GPT-2, scoring classification through label-name probabilities and generating greedily until a newline for generation tasks.GPT-3 has 2.7B, 13B, or 175B parameters; GPT-2 has 1.5B. Test inputs are inserted into the prompt’s test placeholder.Source: §2 · p. 2, §2.2 · p. 3

3. Accuracy Varies Highly Across Prompts

• 54.3% to 93.4% accuracy on 4-shot SST-2 with GPT-3 2.7B can result solely from changing the order of training examples.With prompt format fixed, ordering can matter as much as or more than example selection. Majority-label and recency biases favor labels prevalent in the prompt or repeated near its end.Source: §3 · p. 3, Figure 4 · p. 4
• Accuracy variance remains high even with 16 training examples and can persist with larger models; adding training examples to prompts does not necessarily reduce variance and can sometimes hurt accuracy.Training-example counts were swept for three datasets; DBPedia mean accuracy drops from 36.0% at 0-shot to 25.9% at 1-shot.Source: §3 · p. 3
• Prompt formats differ in average SST-2 accuracy, but all still exhibit high variance across training sets.Holding examples and ordering fixed, the authors test 14 additional formats, including question-answer, conversation-style, Web-page-like templates, and variations in label names.Source: §3 · p. 4

Majority Label Bias

• 50.2% of GPT-3 2.7B’s 4-shot LAMA validation predictions repeat a training answer, versus a correct repeat rate of 24.7%, illustrating majority-label bias.Favoring labels frequent in the prompt degrades accuracy on balanced SST-2. Repeating the lone training example’s class also explains frequent accuracy drops from 0-shot to 1-shot.Source: §4 · p. 4

Recency Bias

• Nearly 90% of predictions are Negative for a P P P N prompt, showing that recency bias can outweigh a three-to-one Positive majority.For 4-shot LAMA, overprediction of answers from training positions 1–4 is 8.5%, 8.3%, 14.3%, and 16.1%, respectively, helping explain sensitivity to example order. Overprediction means model repeat rate minus ground-truth repeat rate.Source: §4 · p. 4

Common Token Bias

• 11× more frequent predictions for “book” than “artist” on balanced 14-way DBPedia illustrate GPT-3’s common-token bias.Label-name frequency correlates with class prediction rate (r = 0.67), helping explain sensitivity to label names. On LAMA, the model often predicts common entities instead of rare ground-truth answers.Source: §4 · pp. 4–5

The Impact of Biases on Model Predictions

• 94% validation accuracy would be achievable for the illustrated SST-2 classifier with an optimal p(Positive) = 0.68 threshold, correcting frequent false positives.The three biases typically shift the model’s output distribution. For this prompt, high Positive confidence causes frequent false positives under the default 50% threshold.Source: Figure 5 · p. 5, §4 · p. 5

5. Contextual Calibration

• Contextual calibration estimates answer bias from content-free inputs and sets transformation parameters to make their class scores uniform, without calibration training data.For content-free probabilities p_cf, the method sets W = diag(p_cf)^-1 and b = 0, then predicts using argmax(Wp + b). Content-free predictions depend on the training examples, their order, and prompt format.Source: §5 · p. 5
• Implementation adds trivial computational overhead by saving content-free probabilities and adjusting outputs; experiments average predictions from “N/A”, “[MASK]”, and the empty string.This ensemble performed best on AGNews and was reused across the other datasets.Source: §5 · p. 5
• Inverse-probability scaling performs better for classification, while subtracting content-free probabilities with W set to the identity performs better for generation.Classification uses renormalized label probabilities; generation calibrates only first-token probabilities because later predictions were highly deterministic on the studied tasks. Restricting W to diagonal avoids quadratic parameter growth.Source: §5 · p. 5

5.1. Results for Contextual Calibration

• Up to 30.0% absolute gains improve GPT-3’s mean and worst-case accuracy across classification and generation tasks.Calibrated GPT-3 2.7B sometimes outperforms the GPT-3 175B baseline by up to 19.3%, despite being over 50x smaller.Source: Table 1 · p. 6, §5.1 · p. 6
• Across the evaluated datasets and language models, contextual calibration considerably reduces accuracy variance across training-example sets in a majority of cases, with only small increases in the remaining cases.Evaluation used five random training-example sets, a fixed prompt format, and labeling budgets of 0–8 examples; contextual calibration and the baseline (standard decoding without calibration) used the same sets.Source: §5.1 · p. 6, Figure 6 · p. 7
• GPT-2 1.5B also exhibits substantial prompt sensitivity, and contextual calibration works out-of-the-box to improve mean accuracy and reduce variance for most tasks.This extends the observed prompt instability and calibration benefits beyond GPT-3.Source: §5.1 · p. 7
• Across prompt formats, contextual calibration improves average and worst-case accuracy for SST-2 and LAMA, while reducing variance for SST-2.Experiments held training examples fixed and tested 15 formats for SST-2 and each of three LAMA relations.Source: §5.1 · p. 7, Figure 7 · p. 7

5.2. Ablations on Contextual Calibration

• Contextual calibration approaches validation-tuned oracle accuracy on AGNews, while SST-2 and AGNews ablations show that content-free input choice matters but many choices work well.The oracle uses the validation set to find the best diagonal W.Source: §5.2 · p. 7

Does Calibration Eliminate the Need to Engineer Prompts?

• Contextual calibration makes best-, average-, and worst-case prompt accuracies higher and more similar, reducing rather than eliminating the need for prompt engineering.Prompt format and examples still warrant tuning for the best performance (Brown et al., 2020; Gao et al., 2020).Source: §6 · p. 7

Should You Finetune in the Few-shot Setting?

• Using a fixed language model without finetuning enables rapid prototyping, a fully natural language interface, and lower memory requirements and serving complexity across tasks.These practical advantages motivate the study’s choice not to finetune.Source: §6 · p. 7
• Finetuning can improve accuracy over in-context learning when its disadvantages are acceptable or avoidable (Schick & Sch¨utze, 2020; Gao et al., 2020).Finetuning can also be unstable in few-shot settings (Schick & Sch¨utze, 2021). Future work could test whether contextual calibration reduces the need for finetuning, or vice versa.Source: §6 · p. 7

Few-shot Learning with Language Models

• Contextual calibration improves the few-shot setting of Radford et al. (2019) and Brown et al. (2020), while potentially correcting underestimation in prompt-based probing.The paper recommends calibration for probes of factual knowledge (Shin et al., 2020) and commonsense knowledge (Bosselut et al., 2019).Source: §7 · pp. 7–8

Volatility of Few-shot Learning in NLP

• The paper extends instability findings to left-to-right in-context learning without finetuning, identifies example-order sensitivity, and analyzes causes to develop mitigation.Prior work found prompt-format sensitivity in zero-shot masked models (Shin et al., 2020). Independent concurrent work found training-example sensitivity during few-shot finetuning (Schick & Sch¨utze, 2020; Gao et al., 2020).Source: §7 · p. 8

Failures of Language Models

• The paper connects in-context learning failures, including recency bias, to previously documented failures in language-model text generation.These include repetition (Holtzman et al., 2020), overconfidence (Braverman et al., 2020; Jiang et al., 2020a), recency bias (Khandelwal et al., 2018; Ravfogel et al., 2019), and preference for generic responses over rare text.Source: §7 · p. 8

• Small decisions such as calibration can greatly influence few-shot results, complicating the development and comparison of pretraining schemes and model architectures.Source: §8 · p. 8
• Future methods work aims to make few-shot learning more robust and extend calibration to a wider range of tasks, including open-ended generation.Source: §8 · p. 8
• GPT-3 improves with more training examples but also learns superficial prompt patterns, including repetition of common answers, motivating closer analysis of in-context learning.Source: §8 · p. 8

• 88.5% accuracy falls to 51.3%, near random chance, for GPT-3 2.7B when the order of two training examples is reversed.Source: Table 2 · p. 11, A. Additional Results on Variance and… · p. 11
• The choice of content-free input affects accuracy, but many good choices exist.Table 3 examines 1-shot SST-2 and 0-shot AGNews. The experiments ensemble over N/A, [MASK], and the empty string.Source: A. Additional Results on Variance and… · p. 11, Table 3 · p. 11
• Figure 9 compares GPT-3 accuracy across LAMA prompt formats with and without calibration; Table 4 examines calibration’s effect on GPT-2.Source: A. Additional Results on Variance and… · p. 11

B. Prompt Formats Used

• 15 prompt formats are used to study the effect of prompt format on SST-2.Table 7 illustrates the formats with one training example.Source: B. Prompt Formats Used · p. 11, Table 7 · p. 15
• Text-classification predictions use the language model’s probabilities for the specified label-name tokens.Table 5 illustrates each task’s prompt with one training example.Source: Table 5 · p. 13
• Default generation-task prompts are documented in Table 6, with one training example per task for illustration.Together, Tables 5 and 6 provide the default prompt formats used across all tasks.Source: B. Prompt Formats Used · p. 11, Table 6 · p. 14