PaperlayerView exact revision on arXiv

Source-linked AI summary

In-Context Learning Learns Label Relationships but Is Not Conventional Learning

Jannik Kossen, Yarin Gal, Tom Rainforth

arXiv:2307.12375v4cs.CLcs.AIcs.LG

TL;DR

Prior work disagrees about whether ICL learns label relationships from examples, motivating a closer study of its use of in-context information. The paper examines ICL across example counts with probabilistic measures and finds that it learns labels and novel tasks, yet retains pre-training preferences and weights context information unequally.

  • Problem

    Researchers disagree about whether ICL learns input–label relationships from demonstrations or behaves like a general-purpose learning algorithm.

    Source: §1 · p. 1

  • Method

    The paper studies ICL across increasing example counts and uses probabilistic metrics to examine its predictions and dynamics.

    Source: §1 · p. 2

  • Results

    ICL typically uses in-context label relationships and can learn novel tasks, but often fails to overcome pre-training preferences and favors information nearer the query.

    Source: §1 · p. 2, §11 · p. 9

  • Takeaways & Limitations

    ICL can learn from labels, but does so differently from an idealized conventional learner, informing understanding of its information processing.

    Source: §1 · p. 2

    The experiments focus on few-shot tasks evaluated with logits rather than free-form generation and do not cover RLHF-finetuned LLMs.

    Source: §9 · p. 9

Abstract

from arXiv · show

The predictions of Large Language Models (LLMs) on downstream tasks often improve significantly when including examples of the input--label relationship in the context. However, there is currently no consensus about how this in-context learning (ICL) ability of LLMs works. For example, while Xie et al. (2021) liken ICL to a general-purpose learning algorithm, Min et al. (2022) argue ICL does not even learn label relationships from in-context examples. In this paper, we provide novel insights into how ICL leverages label information, revealing both capabilities and limitations. To ensure we obtain a comprehensive picture of ICL behavior, we study probabilistic aspects of ICL predictions and thoroughly examine the dynamics of ICL as more examples are provided. Our experiments show that ICL predictions almost always depend on in-context labels and that ICL can learn truly novel tasks in-context. However, we also find that ICL struggles to fully overcome prediction preferences acquired from pre-training data and, further, that ICL does not consider all in-context information equally.

1 INTRODUCTION

Few-shot ICL improves downstream LLM predictions by placing input–label examples in context without updating model parameters, but its learning mechanism remains contested. This paper examines how ICL uses labels and finds both learning capabilities and limits relative to an idealized learner.

1 INTRODUCTION

  • Few-shot ICL prepends input–label examples to a query without updating model parameters, and is widely used across NLP tasks.The passage describes applications including sentiment and document classification, question answering, and natural language inference.Source: §1 · p. 1
  • Prior accounts disagree on whether ICL approximates general-purpose learning algorithms or fails to capture input–label correspondences.The paper highlights the incompatibility between the conventional-learning analogy and Min et al. (2022b)'s claim that LLMs do not learn new tasks at test time.Source: §1 · p. 1
  • The study tests whether ICL uses in-context labels, overcomes pre-training preferences, and treats all in-context information equally.These questions are framed against an idealized conventional learner.Source: §1 · pp. 1–2
  • The authors examine ICL across increasing example counts and use probabilistic metrics to reveal confidence changes that accuracy can miss.They study multiple LLMs and tasks, from no examples through the maximum context size.Source: §1 · p. 2
  • ICL usually uses in-context label relationships and can learn novel tasks, but pre-training preferences persist and nearer examples receive greater weight.The findings offer a middle ground between claims that ICL learns like a conventional algorithm and claims that it does not learn from labels.Source: §1 · p. 2

2 BACKGROUND

Prior work has offered competing accounts of ICL: some propose general-purpose learning, while others emphasize shortcomings and dependence on demonstration properties. The paper situates its investigation amid this unsettled evidence.

2 BACKGROUND

  • Few-shot ICL has become integral to LLM evaluation, including in HELM and other recent publications.The passage cites Liang et al. (2022), Chowdhery et al. (2022), and Hoffmann et al. (2022).Source: §2 · p. 2
  • Some work proposes that ICL implements general-purpose learning, such as gradient descent or Bayesian inference, but argues remain largely theoretical without solid experiments in actual LLMs.The cited proposals include Von Oswald et al. (2023) and Xie et al. (2022); Zhang et al. (2023), Huszár (2023), Wies et al. (2023), and Jiang (2023) are cited regarding limited experimental evidence.Source: §2 · p. 2
  • Other studies report sensitivity to example formatting or order, pre-training label preferences, prompt variation, and task formulations resembling pre-training data.The passage associates these observations with Min et al. (2022b), Lu et al. (2022), Zhao et al. (2021), Chen et al. (2022), and Wu et al. (2023).Source: §2 · p. 3
  • Min et al. (2022b) argue that ICL does not learn input–label relationships and can perform nearly as well with randomized demonstration labels.They instead attribute ICL to learning the label space, example formatting, and input distribution; this paper says it will revisit those claims.Source: §2 · p. 3

3 NULL HYPOTHESES ON HOW ICL INCORPORATES LABEL INFORMATION

The paper defines an idealized conventional learner through its use of labels, ability to learn supplied relationships, and equal treatment of information. It tests whether ICL exhibits these properties through three null hypotheses.

3 NULL HYPOTHESES ON HOW ICL INCORPORATES LABEL INFORMATION

  • A conventional learner is expected to use labels, learn the true label relationship with sufficient data, and consider all data information equally.The paper notes that actual learning approaches may not always meet these intuitions because of imperfect training schemes.Source: §3 · p. 3
  • NH1 states that ICL predictions are independent of in-context conditional label distributions; the paper tests this with randomized labels and a novel task.Probabilistic metrics are used to examine label randomization, and the novel task requires learning a relationship unavailable from pre-training.Source: §3 · p. 3
  • NH2 states that ICL can overcome zero-shot pre-training preferences, testing whether predictions eventually follow modified in-context label relationships.The motivation is that a conventional learner should follow training labels given sufficient data.Source: §3 · p. 3
  • NH3 states that ICL considers all in-context information equally, including when examples contain multiple sources of label relationships.For changing label relations, the hypothesis predicts that ICL predictions should not depend on the order in which relations are presented.Source: §3 · p. 3
  • On SST-2, accuracy and log likelihood improve as in-context dataset size grows, while entropy decreases.The figure caption reports averages over 500 random subsets and moving-average lines for clarity.Source: Figure 2 · p. 4

4 EXPERIMENTAL SETUP & ICL TRAINING DYNAMICS

The experiments evaluate few-shot ICL across models, tasks, and context sizes, using an efficient strategy to recover predictions at every size from each forward pass. Probabilistic metrics show continued improvement beyond the point where accuracy saturates.

4 EXPERIMENTAL SETUP & ICL TRAINING DYNAMICS

  • The study evaluates LLaMa-2, LLaMa, and Falcon models on nine classification and language-understanding tasks.Tasks include SST-2, Subjective, Financial Phrasebank, Hate Speech, AG News, MQP, MRPC, RTE, and WNLI.Source: §4 · p. 4
  • Performance is reported from zero-shot through the maximum number of demonstrations allowed by each model’s input token limit.This broad context-size sweep contrasts with prior work that often evaluates only a few set sizes.Source: §4 · p. 4
  • The evaluation strategy extracts label-token predictions from each forward pass to obtain few-shot ICL predictions at all context sizes without additional cost.Each pass supplies predicted probabilities for input tokens given preceding tokens, including labels of in-context examples.Source: §4 · p. 4
  • Metrics include accuracy, average label-prediction log likelihood, and entropy, averaged over randomly drawn training-set contexts and compared with a class-frequency guessing baseline.Entropy describes how spread predicted probabilities are over classes; it is not treated as a performance metric.Source: §4 · p. 4
  • Log likelihoods and entropies reveal continued ICL changes at larger context sizes after accuracy saturates, and distinguish model certainty on SST-2.Across Falcon models, accuracy and log likelihood increase and entropy decreases as examples accumulate; larger or instruction-tuned models show higher certainty. Similar patterns hold for LLaMa and LLaMA-2.Source: §4 · p. 4

5 DO ICL PREDICTIONS DEPEND ON IN-CONTEXT LABELS?

The authors test whether in-context predictions depend on demonstration labels by comparing standard and randomized-label ICL across models, tasks, and metrics. Randomization usually harms performance, with probabilistic metrics revealing differences more clearly than accuracy.

5 DO ICL PREDICTIONS DEPEND ON IN-CONTEXT LABELS?

  • When default ICL beats guessing, randomized labels almost always significantly reduce log likelihood across the evaluated models and tasks.Table 1 summarizes average differences at maximum context size; entries where default ICL does not beat guessing are excluded from this comparison.Source: Table 1 · p. 5, §5 · p. 5
  • Randomized labels significantly degrade likelihoods and increase entropy as context grows, while accuracy differences are often less clear.For LLaMa-2-70B, performance is always worse under randomization; some smaller Falcon models occasionally show improved likelihoods, but retain gaps from default ICL.Source: §5 · p. 5, Figure 3 · p. 5
  • The experiments reject NH1: ICL predictions depend on the conditional label distribution of in-context examples.The authors report this pattern across models, tasks, and metrics, especially when default ICL outperforms guessing.Source: §5 · p. 5
  • The authors suggest that earlier reports of small randomization effects may reflect fixed, short contexts, less-sensitive metrics, model size, or near-guessing performance.They note that random-label effects grow with context length and that probabilistic metrics are more sensitive.Source: §5 · p. 6

6 CAN ICL LEARN TRULY NOVEL LABEL RELATIONSHIPS?

To test whether ICL can learn label relations absent from pre-training, the authors construct a novel author-identification task from private messages. ICL performs better than chance on this task, supporting the conclusion that it can learn novel label relationships in-context.

6 CAN ICL LEARN TRULY NOVEL LABEL RELATIONSHIPS?

  • The authors designed the task to avoid familiar NLP tasks whose pre-training exposure or strong zero-shot performance could confound evidence of in-context learning.They use it to test whether LLMs can learn truly novel label relationships from context.Source: §6 · p. 6
  • The task uses private messages between two paper authors, ensuring the messages were absent from pre-training and their specific writing-style-to-author mapping was novel.The task asks the model to identify the author of a message from examples of the input-label relationship.Source: §6 · p. 6
  • LLaMa-2 models perform better than random on author identification, and performance improves with model size.All evaluated models except Falcon-7B-Instruct also beat random performance; accuracies and log likelihoods increase on the task.Source: §6 · p. 6

7 CAN ICL OVERCOME PRE-TRAINING PREFERENCE?

The authors test whether in-context examples can override pre-training preferences using flipped and arbitrary label mappings. ICL learns against those preferences to some extent, but flipped-label predictions plateau with a persistent gap from default-label predictions.

7 CAN ICL OVERCOME PRE-TRAINING PREFERENCE?

  • The authors test pre-training preference by mapping default labels to flipped labels or to arbitrary labels such as A and B.Flipped-label experiments evaluate NH2; arbitrary-label results provide a complementary comparison.Source: §7 · pp. 6–7
  • With flipped labels, accuracy rises above guessing, but predictions remain less certain than under default labels and the gap is unlikely to disappear with more observations.The authors note that input-size limits prevent adding examples indefinitely.Source: §7 · p. 7
  • Across the studied models and tasks, flipped-label predictions plateau with a significant entropy gap from default labels even at maximum context size, so NH2 is rejected.The reported pattern suggests that pre-training preferences cannot be fully overcome through in-context observations in these settings.Source: §7 · pp. 7–8, Table 2 · p. 7
  • Arbitrary label directions are similarly easy to learn, while arbitrary labels are learned more slowly than default labels but faster than flipped labels.The authors relate this ordering to inductive biases.Source: Figure 5 · p. 7, §7 · p. 8
  • Prompts can initially improve flipped-label predictions, but eventually they no longer improve them.Source: §7 · p. 8

8 HOW DOES ICL AGGREGATE IN-CONTEXT INFORMATION?

The section tests whether ICL weights in-context label relationships equally, then finds that predictions depend on where those relationships occur. After equal numbers of default and flipped examples, the three setups still produce significantly different predictions.

8 HOW DOES ICL AGGREGATE IN-CONTEXT INFORMATION?

  • The test changes labels from default to flipped, flipped to default, or alternates them after every observation.After 2N observations, each setup contains equal numbers of examples under both label relationships.Source: §8 · p. 8
  • After equal numbers of default and flipped examples, predictions differ significantly across the three setups, rejecting the hypothesis that ICL treats all in-context information equally.The comparison is made after 2N observations; Figure 6 shows few-shot accuracies across the changing-label setups.Source: §8 · p. 8, Figure 6 · p. 8
  • Predictions begin adjusting after the changepoint; at 2N observations, the flipped-to-default setup favors default labels, while the default-to-flipped setup favors flipped labels.This pattern indicates that LLMs preferentially use label information closer to the query rather than treating all observed examples equally.Source: §8 · pp. 8–9

9 DISCUSSION & LIMITATIONS

The discussion connects ICL’s interaction with pre-training preferences and later inputs to prompt-based alignment. It also limits the evaluation to few-shot tasks assessed with logits and excludes RLHF-finetuned LLMs.

9 DISCUSSION & LIMITATIONS

  • Prompt-based alignment may struggle to overwrite pre-training preferences, and future user input could itself overcome prompt-based alignment.The authors identify interactions among pre-training preferences, context strings, and user input as important to understand for alignment.Source: §9 · p. 9
  • For question answering, labels may matter less when in-context examples provide limited information about the query answer.In randomized-label experiments, capable LLMs might instead identify random answers and imitate them.Source: §9 · p. 9
  • The evaluation covers few-shot tasks assessed with logits, not free-form generation, and does not include RLHF-finetuned LLMs.The authors believe the results should transfer to free-form generation, but that setting is not covered by their experiments.Source: §9 · p. 9

10 RELATED WORK

Related work presents competing accounts of ICL, including label-independent explanations and general-purpose learning accounts. The paper distinguishes its approach through probabilistic analysis and full ICL training curves.

10 RELATED WORK

  • Prior studies examine label effects through randomization, arbitrary replacement tokens, or smaller and instruction-tuned models, but do not consider probabilistic metrics or full ICL training curves.The authors argue these omissions can underestimate prediction changes; they contrast their probabilistic findings with Pan et al. (2023) and Wei et al. (2023).Source: §10 · p. 9
  • The paper’s findings on modified labels contradict Pan et al. (2023)’s report of an insignificant gap for small models and do not support claims by Wei et al. (2023) about overriding pre-training knowledge.The passage attributes these contrasts to results discussed in §§5 and 7.Source: §10 · p. 9
  • Related work also studies example selection, uncertainty, in-context versus in-weights generalization, and the emergence of ICL; non-language deep non-parametric models can guarantee order invariance or approximate Bayesian predictions.The passage contrasts those guarantees with ICL.Source: §10 · p. 9

11 CONCLUSIONS

The paper concludes that ICL uses in-context labels and can learn novel tasks, but does not behave like an idealized conventional learner. In tested prompt variations, early gains on flipped labels do not persist at larger context sizes.

11 CONCLUSIONS

  • ICL incorporates in-context label information and can learn novel tasks, but pre-training label preferences persist, and ICL favors label information closer to the query.The conclusion describes ICL as both more capable and more limited than expected from analogies with conventional learning algorithms.Source: §11 · p. 9

A CAN PROMPTS HELP ICL LEARN FLIPPED LABEL RELATIONSHIPS?

  • The prompt experiments add instructions before the examples to inform the model about flipped labels, including instructions to ignore prior knowledge or invert answers.The study explores instruct, ignore, and invert prompts.Source: A CAN PROMPTS HELP ICL LEARN FLIPPED… · p. 15
  • In prompted few-shot ICL, instruct and invert prompts can improve flipped-label performance initially, but none of the tested prompts helps at large context sizes or removes the plateau below default-label performance.The reported results cover LLaMa-65B and Falcon-40B on SST-2, with broader figures for the largest models across tasks.Source: A CAN PROMPTS HELP ICL LEARN FLIPPED… · p. 15, Figure A.1 · p. 15
  • The authors leave open whether undiscovered prompts could change the result; for more capable models, they suspect zero-shot label flipping may be possible but are unsure it would improve ICL on flipped examples.This uncertainty concerns the models and prompting approaches considered in the discussion.Source: A CAN PROMPTS HELP ICL LEARN FLIPPED… · p. 15

B EVALUATION APPROACH FOR CHEAP IN-CONTEXT LEARNING DYNAMICS

The evaluation method extracts predictions for every in-context dataset size from one forward pass. The section also describes how the authors construct and limit their novel authorship task and identifies the evaluated datasets.

B EVALUATION APPROACH FOR CHEAP IN-CONTEXT LEARNING DYNAMICS

  • One forward pass yields few-shot predictions at every context size by extracting label-token probabilities and evaluating the joint probability of observed labels.The label probabilities are selected from the model’s vocabulary distribution; terms condition on observed, true labels rather than earlier predictions.Source: B EVALUATION APPROACH FOR CHEAP… · pp. 16–17
  • Examples are verbalized and concatenated with a verbalized test query, then tokenized so the positions corresponding to label tokens can be tracked.The example verbalizer formats sentence–label pairs, while the query verbalizer leaves the label position for the model to predict.Source: B EVALUATION APPROACH FOR CHEAP… · p. 16
  • The authors score multi-token labels from their first token, assuming it uniquely identifies each label; for LLaMa-7B on Subjectivity, subsequent [ive] probabilities were 0.9998 ± 0.0003.The estimate covers the following 12 instances after the first in-context “subjectivity” label.Source: B EVALUATION APPROACH FOR CHEAP… · p. 17

C AUTHORSHIP IDENTIFICATION TASK

  • The novel authorship task uses 109 Slack messages from two paper authors, filtered from the last 151 messages and truncated at 200 characters.The retained messages number 58 for one author and 51 for the other; they are presented in random order.Source: C AUTHORSHIP IDENTIFICATION TASK · p. 17
  • The authors withhold the authorship dataset because it contains private communication and public release could let future LLMs train on it.They provide six random examples instead.Source: C AUTHORSHIP IDENTIFICATION TASK · p. 17

D DATASET CITATIONS

  • The evaluated tasks include SST-2, Subjectivity, Financial Phrasebank, Hate Speech, AG News, MQP, MRPC, RTE, and WNLI.The passage gives citations for these datasets.Source: D DATASET CITATIONS · p. 17

E EXPERIMENT DETAILS

The experimental details specify label handling, baselines, model and input limits, and evaluation procedures, including tokenization, calibration, and significance testing.

E EXPERIMENT DETAILS

  • The class-frequency guessing baseline predicts the training-set class frequencies for every input, yielding majority-class accuracy and log likelihood based on those frequencies.For accuracy, its prediction is the most frequent class; its log likelihood is the sum of class frequencies times their log frequencies.Source: E EXPERIMENT DETAILS · p. 18
  • For multiclass label flipping, Financial Phrasebank and AG News labels are rotated cyclically rather than simply inverted.For Financial Phrasebank, this rotation changes the meaning of the “neutral” label.Source: E EXPERIMENT DETAILS · p. 18
  • The experiments use Hugging Face task datasets, selecting only Hate Speech examples with labels 0 and 1, and use distinct sentence-pair or question-pair templates where applicable.No custom processing is used for the other datasets.Source: E EXPERIMENT DETAILS · p. 18
  • For LLaMa-2-7B on Subjectivity, 8-bit quantization causes no significant performance degradation compared with 32-bit precision.The figure averages over 500 random in-context datasets and shows default and randomized labels.Source: E EXPERIMENT DETAILS · p. 18, Figure E.1 · p. 19
  • Incorrect whitespace handling can change label tokenization and adversely affect ICL; the test query must not end with trailing whitespace.The relevant positive-label token depends on the tokenizer and preceding whitespace, so the token matching the in-context label must be evaluated.Source: E EXPERIMENT DETAILS · pp. 18–19
  • Differences are marked significant at 95% when absolute average differences exceed 1.96 standard errors; comparisons against guessing use a 1.645-standard-error threshold.The latter test checks whether mean performance plus 1.645 standard errors exceeds the guessing baseline for accuracy and log likelihood.Source: E EXPERIMENT DETAILS · p. 19
  • The maximum reported context size is the smallest example count that exceeds the model’s 2048-token limit for Falcon and LLaMa or 4096-token limit for LLaMa-2.Longer task inputs therefore allow fewer in-context examples; Table E.1 lists the maximums by model and task.Source: E EXPERIMENT DETAILS · p. 19, Table E.1 · p. 19
  • Calibration rarely improves and can degrade predictions, particularly when labels use multiple tokens; the displayed LLaMa-65B comparison finds no performance improvement.The figure shows default and randomized labels with and without calibration.Source: E EXPERIMENT DETAILS · p. 19, Figure E.2 · p. 20

F EXTENDED RESULTS

The extended results examine how label setups, noise, calibration, demonstration counts, and task context relate to performance and probabilistic predictions, including whether examples are treated equally or copied exactly.

F EXTENDED RESULTS

  • On SST-2, accuracy and log likelihood improve as the in-context dataset grows, while entropy decreases.Figure F.1 averages over 500 random subsets of the training set.Source: F EXTENDED RESULTS · p. 21, Figure F.1 · p. 22
  • Across models and tasks, performance is usually significantly worse with randomized labels when default-label performance exceeds random guessing.Table F.1 reports mean differences and standard errors over 500 runs; positive accuracy or log-likelihood differences indicate worse performance with randomized labels.Source: F EXTENDED RESULTS · p. 21, Table F.1 · p. 34
  • All models achieve accuracy significantly above random guessing on the novel author-identification task, indicating that predictions depend on the in-context label relationship.Figure F.13 reports few-shot results for all models on the novel task.Source: F EXTENDED RESULTS · p. 21, Figure F.13 · p. 35
  • Across models and metrics, performance is usually significantly worse with flipped labels when default-label performance exceeds random guessing.Table F.2 summarizes differences at maximum context size over 100 runs and reports significant differences particularly for entropy.Source: F EXTENDED RESULTS · p. 21, Table F.2 · p. 54
  • At twice the changepoint, predictions are not equal despite equal numbers of examples for each label relationship; for LLaMa-2-70B on Hate Speech, differences are significant at a changepoint of 30.The additional changepoint comparison tests whether ICL treats all in-context information equally.Source: F EXTENDED RESULTS · p. 21, Figure F.35 · p. 58, Figure F.37 · p. 60
  • Some prompts improve few-shot ICL with flipped labels compared with using no prompt across the reported tasks.The prompted experiments cover SST-2, Subjectivity, Financial Phrasebank, Hate Speech, AG News, Medical Questions Pairs, MRPC, RTE, and WNLI.Source: F EXTENDED RESULTS · p. 21, Figure F.38 · p. 61, Figure F.39 · p. 62

G.1 PROGRESSIVE RANDOMIZATION

  • With LLaMa-2-70B on SST-2, log likelihood degrades at 0.2 random-label noise, while accuracy stays relatively constant until noise reaches 0.6.The reported accuracy stability lasts until more than half of the in-context points are noisy.Source: §G.1 · p. 70

G.2 CALIBRATED LABEL FLIPPING

  • Calibrated predictions produce generally small changes, but large performance differences between label setups remain at maximum context size, so NH2 remains rejected.The passage notes almost no difference for SST-2/Falcon-40B and small improvements for all labels on Hate Speech/LLaMa-65B and MQP/LLaMa-2-70B.Source: §G.2 · p. 70
  • At 16 demonstrations, calibration gains are insignificant, and the authors report that their conclusions remain unchanged when studying larger demonstration counts.They contrast their setting with Zhao et al. (2021), who focus on up to 16 in-context examples, where calibration is most important.Source: §G.2 · p. 70

G.3 LABEL FLIPPING FOR AUTHOR IDENTIFICATION

  • On novel author identification, accuracy is similar across label setups and probabilistic differences are smaller than in previous experiments, consistent with weaker pretraining preferences.The task is designed to be absent from training data; the authors argue flipped or replaced labels should matter more when pretraining favors a particular label setup.Source: §G.3 · pp. 70–71, Figure G.3 · p. 72
  • Initially, the model can favor author names over arbitrary symbols and larger LLaMa-2 models can slightly favor the default label direction, but either finding suffices to reject NH2.The authors suggest pretraining may explain the names-versus-symbols bias, while author names appearing in sentences may confuse the flipped direction.Source: §G.3 · p. 71

G.4 ANSWER IN CONTEXT

  • Four repetitions of a test query yield perfect accuracy across label setups, whereas one repetition improves arbitrary and flipped labels without producing perfect copying.The authors conclude that ICL is not a nearest-neighbor predictor and that a single observation has limited influence on its decision function.Source: §G.4 · p. 71, Figure G.4 · p. 72
  • One exact test-query repetition improves performance for arbitrary or flipped labels but not default labels, and does not produce perfect accuracy.Default-label performance is already high; the authors interpret the lack of perfect copying as evidence that ICL is not a nearest-neighbor predictor.Source: §G.4 · p. 71, Figure G.4 · p. 72

Related papers

Loading 2307.12375v4…