PaperlayerView exact revision on arXiv

Source-linked AI summary

A Reality Check of Inference-Time Context Augmentation Methods

Haeun Yu, Geert Heyman, Pepa Atanasova, Isabelle Augenstein

arXiv:2610.30983v1cs.CL

TL;DR

LLMs may fail to use provided context, while existing context-augmentation methods have been evaluated in isolation and by accuracy alone. This paper evaluates sentence-level interventions and finds that accuracy gains can reflect undirected prediction changes rather than reliable improvement.

  • Problem

    Context-augmentation methods have been evaluated in isolation and by performance gains alone, leaving their comparative reliability unclear.

    Source: §1 · p. 1

  • Method

    The framework evaluates sentence-level context interventions and uses sign-randomisation to test whether prediction changes consistently favor correct answers.

    Source: §3.2.2 · p. 4, §5.1 · p. 6

  • Results

    Across models and datasets, accuracy gains are inconsistent and often reflect undirected perturbations rather than movement toward correct answers.

    Source: §1 · p. 2, §5.1 · p. 6, §7 · p. 8

  • Takeaways & Limitations

    Researchers should verify that gains are directed and assess how often methods break correct predictions before applying context augmentation blindly.

    Source: §7 · p. 8

    The evaluation uses MCQA and binary classification with log-likelihood scoring and single-sentence interventions, so findings may not transfer to open-ended generation or whole-context and combined interventions.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Large language models (LLMs) often fail to utilise the context given in the input, resulting in hallucinations or a fallback to parametric knowledge. Context augmentation methods aim to improve model performance by modifying the input, from inserting filler tokens that add computational time to adding relevant content such as elaboration of a context during inference. However, these methods have been evaluated in isolation and only by their performance gains, leaving it unclear how they compare and whether they reliably improve performance. In this paper, we present an evaluation framework spanning five methods that differ in how informative the inserted content is and apply them at the sentence level across seven models spanning multiple sizes on three datasets. We find that accuracy gains are not always genuine improvements but undirected perturbations: the methods correct some predictions while breaking others, particularly when the inserted text allows only for more computational time. Our analysis reveals that methods that bring no additional content primarily change model predictions near the decision boundary. Our findings highlight the need to evaluate the reliability of the accuracy gains before adopting them in practice.

1 Introduction

LLMs can fail to use provided context, and prior context-augmentation methods have been evaluated in isolation and mainly by accuracy gains. This section introduces a shared evaluation framework and reports that gains can reflect undirected perturbations rather than consistent improvements.

1 Introduction

  • Prior methods insert content ranging from filler tokens to relevant knowledge, but their isolated evaluations leave the reliability of accuracy gains unclear.LLMs may still hallucinate or default to parametric knowledge despite context being provided.Source: §1 · p. 1
  • The framework compares five methods that vary in information content by applying each after every context sentence, one position at a time.This standardizes intervention positions across methods.Source: §1 · p. 2
  • Accuracy is paired with a sign-randomisation test to distinguish movement toward the context-supported correct answer from undirected prediction changes.The framework also examines which predictions change, how many positions matter, and whether successful positions can be identified in advance.Source: §1 · p. 2
  • Methods that add little information often perturb predictions without consistent direction, and their effects vary across models and datasets.Successful intervention positions remain an open problem for future work.Source: §1 · p. 2
  • Half of context positions can flip both incorrect and correct predictions, while successful positions cannot be predicted in advance.The analysis finds no advance prediction from sentence lexical characteristics or the last position.Source: §1 · p. 2

2 Related Work

Prior context augmentation ranges from adding computation through repetition or filler tokens to adding knowledge through retrieval or generated explanations. The paper situates its comparison against work evaluated in isolation and studies interventions with differing information content and positions.

2 Related Work

  • Filler tokens add computational space, while input repetition changes internal representations through bidirectional-like computation [Pfau et al., 2024; Goyal et al., 2024; Jang et al., 2025; Springer et al., 2025].Prior work reports downstream-task improvements for repetition [Xu et al., 2024; Leviathan et al., 2025].Source: §2 · p. 2
  • Knowledge-adding approaches include retrieval-augmented generation and explanations or elaborations generated by the model or a stronger model [Lewis et al., 2020; Zeng et al., 2025; Liu et al., 2022; Chen et al., 2023; Li et al., 2025].These approaches intervene over the whole input, between context and question, or at the input end [Xu et al., 2024; Leviathan et al., 2025; Liu et al., 2022; Li et al., 2025; Pfau et al., 2024; Greenblatt, 2025].Source: §2 · p. 2
  • Zhang et al. (2026) report that task-irrelevant context can leave aggregate accuracy nearly unchanged while individual predictions flip in both directions.Their study covers the least informative intervention; this paper compares interventions with increasing information content.Source: §2 · p. 2
  • The paper addresses isolated evaluations by comparing methods under common conditions; Table 1 illustrates intervention texts and marked insertion positions for an MMLU question.The correct option is marked in bold.Source: §2 · p. 2, Table 1 · p. 3

3 Benchmarking Context Augmentation Methods

The framework compares five sentence-level context augmentation methods that vary in the information they add, evaluating both accuracy and whether changes consistently favor correct answers. It also tests how intervention position affects predictions.

3.1 Context Augmentation Methods

  • The five methods range from filler tokens and repeated or paraphrased sentences to explanations generated by the evaluated model or a teacher model.Blank adds no information; Repeat and Paraphrase restate existing content, while Self-explanation and Teacher-explanation add generated information.Source: §3 · p. 3, §3.1 · p. 3
  • Each method is applied separately after every sentence, giving each context one intervention position at a time.The same question is paired with each intervened input, enabling comparison across positions and methods.Source: §3 · p. 3, §3.2 · p. 4

3.2.1 Sentence-Averaged Accuracy

  • Sentence-averaged accuracy measures correctness across intervened inputs, and its change decomposes into corrections minus breaks relative to original accuracy.Corr counts incorrect-to-correct flips; Brk counts correct-to-incorrect flips.Source: §3.2.1 · p. 4

3.2.2 Sign-Randomisation Test

  • The sign-randomisation test checks whether interventions move predictions toward correct answers more consistently than equally sized changes with random directions.Random signs preserve each margin-change magnitude while removing its direction; the test compares observed Corr and Brk with the resulting null distribution.Source: §3.2.2 · p. 4
  • A method rejects the null only when observed corrections exceed the null distribution’s 95% interval and breaks do not.This distinguishes directed improvement from undirected perturbation, which can raise both Corr and Brk.Source: §3.2.2 · p. 4

4 Experimental Setup

The experiments apply the intervention framework to seven open-weight models and three datasets selected to support sentence-level interventions. Intervention texts are generated or constructed to vary in information content.

4 Experimental Setup

  • The evaluation spans seven open-weight models from four model families and three datasets.The models range from OLMo-2-0425-1B-Instruct to Qwen2.5-72B-Instruct-AWQ; datasets are MedQA, MMLU, and DRUID.Source: §4 · p. 4, §4.3 · p. 5, §4.1 · p. 5
  • Table 2 reports MedQA sentence-averaged accuracy and marks correction and break fractions alongside sign-test results.Higher-than-Original accuracies are bolded, and colored cells indicate methods passing the sign-randomisation test.Source: Table 2 · p. 5

4.1 Datasets

  • MedQA, MMLU, and DRUID provide background contexts necessary for questions without explicitly stating the gold answers.MMLU is restricted to five subjects with multi-sentence passages, and DRUID claims with insufficient evidence are excluded.Source: §4.1 · p. 5

4.2 Implementation of Interventions

  • Blank uses whitespace, Repeat copies the sentence, and Paraphrase restates it; self- and teacher-generated explanations elaborate the sentence.The explanation-generation prompt excludes the question and answer options to prevent an answer shortcut.Source: §4.2 · p. 5

5 Results

Across models and datasets, sentence-level accuracy gains are inconsistent and often reflect undirected prediction changes rather than reliable movement toward correct answers. Methods adding information more often pass the sign-randomisation test, though results vary across models.

5.1 Sentence-Averaged Accuracy Results

  • Teacher-explanation improves MedQA sentence-averaged accuracy over Original for six of seven models by 2.0 to 9.3 points; Self-explanation gains 0.3 to 3.7 points for five.The largest reported gains, +9.3 for Gemma-4b and +7.0 for OLMo-7b, do not reject the null hypothesis.Source: §5.1 · p. 6
  • Blank, Repeat, and Paraphrase barely change sentence-averaged accuracy for most models because similar correction and break rates cancel.For Blank with OLMo-7b, Corr is +3.2 and Brk is −3.0.Source: §5.1 · p. 6
  • Across models and datasets, many accuracy gains arise from undirected perturbations rather than consistent movement toward correct answers.The authors conclude that accuracy-based evaluation should be complemented by tests of direction, such as their sign-randomisation test.Source: §5.1 · p. 6

5.2 Sign-Randomisation Test Results

  • Teacher-explanation rejects the sign-test null in 8 of 21 model–dataset pairs and Self-explanation in 6, versus 4 each for Repeat and Blank.The test identifies cases where observed corrections exceed the 95% null interval while breaks do not.Source: §5.2 · p. 6, Table 3 · p. 6
  • Results vary across models: undirected perturbation prevails for OLMo models, while Qwen-72b and Llama models reject the null most often.The sign-test outcome is not uniform across model families.Source: §5.2 · p. 6

6 Analysis

The analysis examines which sentence-level interventions change predictions, how many positions can rescue or break them, and whether successful positions can be identified in advance.

6.1 Which Instances are Affected?

  • Rescue and Break interventions mainly affect instances near the decision boundary, shifting margins by similar amounts in opposite directions; Unchanged instances lie farther away.The boundary pattern holds across datasets, while knowledge-oriented methods shift margins more strongly than computation-oriented methods.Source: §6.1 · pp. 6–7

6.2 How Many Positions are Affected?

  • Once a context can be affected, interventions change predictions at about half its sentence positions, with similar Rescue and Break fractions.The chosen position determines whether the intervention helps or harms.Source: §6.2 · p. 7

6.3 Can Positions be Identified?

  • Lexical characteristics alone cannot identify successful positions: their AUROCs are 0.48–0.54, near chance.Last-position performance is mostly on par with randomly selected non-last positions; successful positions are plentiful but cannot be identified in advance from sentence content or location.Source: §6.3 · p. 8

7 Discussion & Conclusion

Accuracy gains from context augmentation are inconsistent across models and datasets, and can reflect undirected prediction changes rather than movement toward correct answers. The discussion argues for testing whether gains are directed and for validating information-adding methods across settings.

7 Discussion & Conclusion

  • Intervening after the last sentence is mostly on par with other positions, while identifying successful positions in advance remains open.Figure 3 compares last-position accuracy with accuracy at randomly selected non-last positions.Source: Figure 3 · p. 8, §7 · p. 9
  • Context augmentation effects vary by model and dataset, and gains can arise from undirected perturbations even for methods that add content.Rescues and breaks occur at similar fractions of sentence positions once a prediction can be changed.Source: §7 · p. 8
  • Accuracy gains alone cannot distinguish guidance toward the correct answer from prediction perturbation.The authors recommend checking whether gains are directed and how often correct predictions are broken, and favor validating methods that add new information for each model family and dataset.Source: §7 · p. 8

Limitations

The evaluation is limited to classification-style tasks and single-sentence interventions, and its model and language coverage is restricted.

Limitations

  • The evaluation uses multiple-choice and binary classification tasks scored by answer-option log-likelihood, not free-form generation.The authors say findings may not transfer to open-ended generation; they also intervene at one sentence at a time, so results may differ from whole-context or combined interventions.Source: Limitations · p. 9
  • The study covers seven instruction-tuned open-weight models from 1B to 72B parameters, all on English datasets, and excludes reasoning models.The authors excluded reasoning models because their chain of thought may interact with inserted text.Source: Limitations · p. 9

A.1 Dataset Details

The datasets cover medical and world-knowledge multiple-choice question answering and evidence-based fact-checking. The appendix also describes dataset statistics and intervention implementation.

A.2 Implementation of Interventions

  • Table 5 reports sentences per context as mean ± standard deviation, with total sentences equaling intervention positions per model and method.The implementation segments sentences with spaCy; runs took one to two hours on an AMD Instinct MI250X GPU, or 0.5 hours per Qwen-72b run on an NVIDIA B200 GPU.Source: Table 5 · p. 11, §A.2 · p. 11

A.1 Dataset Details

  • MedQA and MMLU are multiple-choice question-answering datasets, while DRUID is fact-checking with support and refute labels.MedQA uses patient descriptions as context; DRUID provides claims and real-world retrieved evidence.Source: §A.1 · pp. 10–11
  • MMLU is restricted to five subjects with questions averaging 654–1,391 characters because most other questions are short and lack intervention positions.The selected subjects are high school European, US and world history, professional law, and professional medicine.Source: §A.1 · p. 11

B Results

Across MMLU and DRUID, context augmentation methods show different rates of statistically supported accuracy gains, with Teacher-explanation gaining most often. Detailed sign-randomisation results are provided for all three datasets.

B Results

  • Teacher-explanation shows the most frequent accuracy gains on both MMLU and DRUID.On MMLU, methods adding only computation do not consistently improve performance.Source: B Results · p. 11
  • On MMLU, context augmentation methods reject the null hypothesis 5 out of 35 times.DRUID has the most rejections, at 12 out of 35.Source: B Results · p. 11
  • Detailed sign-randomisation test results are reported for MedQA, MMLU, and DRUID.The tests report correct and broken predictions and compare observed values with sign-randomised null distributions.Source: B Results · p. 11

C.1 Margin Analysis Results

Across the supplementary analyses, methods adding computation and new content produce larger margin changes than computation-only methods. Multiplicity rates are similar across models and methods, while lexical characteristics generally show little relation to intervention outcomes.

C.1 Margin Analysis Results

C.2 Multiplicity Analysis

C.3 Predictability with Lexical Characteristics

  • Most lexical-characteristic AUROC values are near 0.5, indicating no relation to intervention outcomes.The table reports mean AUROC across seven models and three datasets for identifying positions leading to Rescue or Break.Source: §C.3 · p. 12, Table 16 · p. 18
  • The AUROC analysis labels each intervention position by whether it leads to Rescue or Break, then tests how well each characteristic separates those positions.Positions are pooled across contexts, and confidence intervals are bootstrapped over contexts.Source: §C.3 · p. 12

Related papers

Loading 2610.30983v1…