Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Learning to Select In-Context Demonstration Preferred by Large Language Model

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read GenICL claims that demonstration selection for in-context learning can be directly optimized as generative preference learning from LLM feedback, outperforming retrieval-based baselines across 19 datasets.

desk verdict GenICL is a solid empirical contribution with a shaky theoretical derivation and an acknowledged set-level gap; worth a serious referee but needs honest reframing and reproducibility artifacts. read the letter →

arxiv 2505.19966 v1 pith:RZ4T6YQW submitted 2025-05-26 cs.LG cs.AI

Zheng Zhang , Shaocheng Lan , Lei Song , Jiang Bian , Yexin Li , Kan Ren This is my paper
classification cs.LGcs.AI
keywords in-contextlearningdemonstrationselectionpreferenceLLMfeedbacklatentvariablegenerativeBayesianoptimizationKTOfew-shotprompting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the usual way of choosing in-context demonstrations for large language models—retrieving examples by embedding similarity, or training a retriever with contrastive losses—optimizes a surrogate objective that is misaligned with what actually helps the model. It proposes GenICL, a generative preference-learning method that treats demonstration selection as a Bayesian inference problem with a latent variable representing the LLM's preference, and optimizes that variable directly against LLM feedback. The training signal is a pair of losses, one at the demonstration level and one at the answer level, both formulated with preference learning on preferred and non-preferred examples. At inference, candidates are scored by the probability they generate the learned preference variable, and the top K are placed in the prompt. If the claim holds, a modestly sized trainable module plus a frozen LLM can consistently outperform retrieval-based selectors, including on tasks where most candidate demonstrations are useless.

What carries the argument

The load-bearing object is the latent demonstration-preference variable $z$, realized in implementation as a task-specific description with low-rank adapter (LoRA) parameters. It is introduced through the Bayesian identity $P_{\mathcal{M}}(Y\mid \{(X_k,Y_k)\}_{k=1}^K, X) = \int_z P_{\mathcal{M}}(Y\mid z, X) P_{\mathcal{M}}(z\mid \{(X_k,Y_k)\}_{k=1}^K, X)\,dz$, and an ELBO argument turns optimization into two terms: an answer-level loss on $P_{\mathcal{M}}(Y\mid z, X)$ and a demonstration-level loss on $P_{\mathcal{M}}(z\mid (X_k,Y_k), X)$. Both losses are KTO-style preference objectives—utility-based updates that push up the utility of preferred demonstrations and answers while pushing down non-preferred ones relative to a reference model—and they are optimized alternately. At inference, each candidate $(x_k,y_k)$ is scored independently by $P_{\mathcal{M}}(z\mid (x_k,y_k), x_{\text{test}})$ and the top $K$ are selected, so the latent variable is what converts the combinatorial set-selection problem into a per-example ranking problem.

What would settle it

On a single dataset, take a pool of about 20 candidate demonstrations, compute GenICL's per-example score for each, then exhaustively evaluate all K-combinations, with a fixed order, under the frozen LLM. If the set formed by the top-K individual scores does not match the highest-accuracy combination, the independence assumption fails; the paper's Limitations predict this will happen for some datasets. A simpler observation-level check is to compare accuracy when the top-K set is replaced by a lower-scoring set that is more diverse in gold answers.

Watch

Extended reading notes

Core claim

The paper's central claim is that the demonstration-selection problem for in-context learning, $\arg\max_{(x_k,y_k)\in\mathcal{P}} P_{\mathcal{M}}(y \mid \{(x_k,y_k)\}_{k=1}^K, x)$, can be reorganized into a generative Bayesian optimization problem by introducing a latent variable $z$ representing the LLM's demonstration preference. Using an ELBO argument, the objective separates into a term $-\log P_{\mathcal{M}}(Y\mid z,X)$ and a term $-\log P_{\mathcal{M}}(z\mid (X_k,Y_k), X)$, which GenICL optimizes alternately as an answer-level preference loss and a demonstration-level preference loss, both built with KTO-style updates on preferred and non-preferred demonstrations. The result is a trained latent variable, realized as a task-specific description with LoRA parameters, such that at inference, ranking demonstrations by $P_{\mathcal{M}}(z\mid (x_k,y_k), x_{\text{test}})$ and taking the top $K$ yields demonstrations the LLM actually benefits from. The paper reports that this direct optimization outperforms retrieval and retriever-training baselines on 19 datasets across classification, multiple-choice, and generation tasks, and that demonstrations selected by a small model transfer to larger models.

Load-bearing premise

The load-bearing premise is that the best prompt can be assembled by ranking each candidate demonstration on its own and taking the top few; the paper's Limitations explicitly admit that individually optimal demonstrations need not form the overall best combination.

Editorial extensions

If this is right

  • A moderately sized trainable module, about 79M parameters via LoRA, plus a frozen LLM can replace per-task retrieval pipelines and still improve few-shot accuracy.
  • Demonstrations selected on a small LLM transfer to larger LLMs without retraining, so one trained selector can serve multiple models.
  • High-quality demonstrations chosen by preference learning are less sensitive to ordering, weakening the common failure mode where random few-shot prompts depend heavily on example order.
  • More demonstrations is not uniformly better: the selected top-K can degrade accuracy when K grows, so K should be treated as a tunable per-task choice.
  • Directly optimizing the ICL objective gives gains over trained retrievers on generation metrics like ROUGE and exact match, where retrieval baselines plateau.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's own Limitations imply that a set-level reranking stage, one that scores combinations of demonstrations rather than individual ones, is the natural next step; GenICL does not test this, and such a stage could close the gap between Eq. (1) and the top-K approximation.
  • Because the preference signal is just the LLM's log-likelihood of the gold output conditioned on a candidate, the same training recipe could be applied to other prompt components such as instructions or output templates; the paper restricts the claim to demonstration examples.
  • The transfer result across model sizes suggests a cheap deployment path, namely training one small selector and reusing its ranked demonstrations as a frozen front-end for larger models, but the paper reports this on only four tasks, so broader transfer is an open extension.
  • The observed non-monotonic effect of K implies the number of demonstrations could be optimized per task as part of the same preference framework; the paper reports the effect but does not exploit it.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes GenICL, a generative preference-learning framework for selecting in-context demonstrations. A latent variable z is introduced to represent the LLM's demonstration preference; the method derives an ELBO-style objective from the ICL marginal likelihood, trains z with KTO-style preference losses at both the demonstration level and the answer level, and at inference scores each candidate by the probability of generating z given the candidate and the test query, selecting the top-K demonstrations. Experiments on 19 datasets across classification, multi-choice, and generation tasks report consistent gains over retrieval and learned-retriever baselines, with additional results on multiple backbone LLMs and cross-model transfer of selected demonstrations.

Significance. If the central claim were fully established, GenICL would be a useful contribution: a lightweight trainable selector that, together with a frozen LLM, improves few-shot ICL across many task types. The paper deserves credit for its broad evaluation, ablations of the two loss terms, generalization across several LLMs, and the cross-model transfer experiment in Table 4. However, the theoretical derivation and the relationship between the set-level objective and the item-level scoring procedure need substantial clarification before the 'directly optimizes demonstration selection' claim can be accepted.

major comments (4)
  1. [Section 3.2 and Appendix D] Equation (2) is presented as a transformation, but it assumes PM(Y | z, {(Xk,Yk)}_K, X) = PM(Y | z, X), i.e., that the target output is conditionally independent of the demonstrations once z is known. This is an unstated modeling assumption, not an identity. The ELBO derivation in Appendix D then uses q(z)=delta(z-z*) and discards delta(0) as a constant; a Dirac measure is not a density with respect to Lebesgue measure, the ratio PM(...)/delta(0) is not well-defined, and optimizing the location of a point-mass variational distribution does not produce the stated objective. These issues are load-bearing for the claim that Eq. (4) is derived from Eq. (1).
  2. [Section 3.1-3.4 and Limitations] The formal objective in Eq. (1) is over sets of K demonstrations, but the method scores and selects demonstrations independently: Eq. (4) replaces the set posterior with a single-demonstration term, Eq. (6) trains on each candidate in isolation, and Eq. (11) selects the top-K by individual scores. The paper's own Limitations state that 'the combination of individually optimal demonstrations does not necessarily result in the overall best combination.' Table 5 only tests the order of a fixed selected set and does not vary set composition, while Figure 4 shows non-monotonic performance as K changes, which is at least consistent with set-level interactions. The experiments therefore do not establish the set-level 'direct optimization' claim; at present the method should be described as an item-level scoring heuristic.
  3. [Section 4.1 and Appendix A] Appendix A states that EPR results are 'directly adopt[ed]' from the LLM-R paper, but Section 4.1 claims that 'our method and all baselines use the same template.' Borrowed numbers may come from different prompt templates, evaluation subsets, or decoding settings, so the head-to-head comparisons with EPR in Table 1 are not controlled. EPR should be rerun under the same protocol as the other methods, or the EPR column should be removed from the comparison.
  4. [Tables 1-4] All experimental tables report single point estimates without standard errors, number of seeds, or significance tests. Some reported gains are small (e.g., Table 3, Qwen2.5-3B on E2E NLG: 37.9 vs. 37.0; Table 1, AGNews: 92.6 vs. 92.4), so the claim of consistent, significant superiority is not statistically supported. The authors should report mean and variance over multiple seeds or at least provide significance tests for the main comparisons.
minor comments (4)
  1. [Section 3.3, Eq. (5)] The text calls s_k the 'log-likelihood' of the LLM generation, but the displayed formula contains no logarithm.
  2. [Section 4.4] There is a typo in the first sentence: 'th generalizability' should be 'the generalizability'.
  3. [Section 3.2 and 3.4] The latent variable z is called a 'task-specific description' with trainable parameters, but the paper does not specify how probabilities such as PM(z | (xk,yk), x) are computed; please clarify whether z is a token sequence and whether these are token-level log-probabilities.
  4. [Figure 3] The figure caption says 'predicted probability of the ground truth' but does not explain how the distribution is computed or over what set of queries; please add this information.

Circularity Check

2 steps flagged · score 3.0 of 10

GenICL's set-level ICL objective is optimized through an item-level surrogate trained on the same LLM's likelihood scores; the claimed direct optimization is therefore partially self-referential, though held-out evaluation and baselines provide independent content.

  1. fitted input called prediction [Section 3.3 Eq. (5) and Section 3.4 Eq. (11)]
    "The preference score sk is the log-likelihood of LLM generation w.r.t. the ground-truth output y given the candidate (xk, yk) and the input x, reflecting the LLM's preference for the demonstration. sk = P(y|(xk, yk), x) ... we select the top-K demonstrations based on the probability of generating the latent variable, as shown below: PM(z|(xk, yk), xtest)"

    The training labels for the latent variable z are the frozen LLM's own per-demonstration likelihoods of the ground-truth output (Eq. 5): preferred demonstrations are those with highest P(y|demo,x), non-preferred those with lowest. The demonstration-level KTO loss (Eq. 6) trains z to reproduce that ranking, and inference ranks candidates by P(z|demo,x_test) (Eq. 11), taking the top-K. Evaluating the selected demonstrations on the same frozen LLM's ICL accuracy therefore measures how well z recapitulates the LLM's own likelihood scores. The claimed 'prediction' of effective demonstrations is a distilled version of the LLM's self-preference rather than an independent derivation, and the set-level target of Eq. (1) is not what is actually optimized.

  2. self definitional [Section 3.2 Eqs. (3)-(4)]
    "Due to the combinatory search space of (X1,Y1),(X2,Y2),...,(XK,YK) being immense, we simplify the optimization objective to the following formula. L(z) = −log PM(Y|z,X) − log PM(z|(Xk,Yk),X)"

    Eq. (1) is an arg max over sets of K demonstrations. Eq. (3) still conditions z on the full set, but Eq. (4) replaces the set with a single demonstration (Xk,Yk), so the optimization target is no longer the ICL objective it claims to optimize. The paper's own Limitations concede: 'we treat each demonstration independently, ignoring the interactions between demonstrations (such as order and combination). In the inference phase, we select a set of demonstrations based on their scores, but the combination of individually optimal demonstrations does not necessarily result in the overall best combination.' Thus the claimed equivalence between the method's objective and Eq. (1) is assumed rather than derived; the 'direct optimization' claim is built on an item-level proxy by construction.

full rationale

The paper does not rely on load-bearing self-citation: the Bayesian latent-variable framing is attributed to Xie et al. (2021) and Wang et al. (2024b), and the preference-learning machinery to KTO; none of these references overlap with the present authors. The main circularity concern is internal. Eq. (1) defines a set-level objective, but Eq. (4) simplifies the conditioning to a single demonstration, and the Limitations explicitly admit that item-wise optimal demonstrations need not form the best set. Table 5 only tests order sensitivity of a fixed selected set; it does not vary set composition. Additionally, the training labels in Eq. (5) are the frozen LLM's own per-demonstration likelihoods of the gold answer, and inference in Eq. (11) ranks by the trained z's approximation of that same likelihood, with evaluation performed on the same LLM. This makes the procedure substantially a distillation of the LLM's self-preference rather than an externally anchored first-principles derivation. Nevertheless, the selector is trained on training splits and tested on held-out queries across 19 datasets, and comparisons against EPR, LLM-R, BM25, SBERT, and CBDS give the empirical results independent content beyond the training signal. The circularity is therefore partial—an item-level proxy and a self-referential feedback loop—not a wholesale reduction of the prediction to its inputs, so a moderate score is appropriate.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central method rests on a latent variable z whose behavior is defined by a trained model, plus several hand-chosen hyperparameters. The most consequential assumptions are the sufficiency of z for the demonstrations, the concentration of the posterior, and the independence of demonstration scores. The E5 prefilter and the single-demo scoring rule are additional assumptions that the paper itself acknowledges as limitations.

free parameters (5)
  • beta (KTO KL regularization coefficient) = 0.1
    Chosen by hand; controls the strength of the KL constraint in both demonstration-level and answer-level losses. The paper does not report a sensitivity study for this value.
  • lambda_w and lambda_l (preferred and non-preferred loss weights) = 1.0 each
    Chosen by hand; weights for preferred and non-preferred demonstrations in the KTO losses.
  • K (number of in-context demonstrations) = 8
    Fixed at 8 for all experiments; Figure 4 shows that performance is sensitive to K, so this is a consequential hand-chosen value.
  • E5 candidate subset size (xi) = not reported
    The paper says E5 base is used to reduce the pool to a subset xi, but the size of that subset is never reported. This parameter controls the tradeoff between search cost and coverage.
  • Training steps and warmup steps = 20,000 steps, 3,000 warmup
    Chosen by hand; the paper reports them in Table 6 but does not justify the values or study their sensitivity.
assumptions (5)
  • ad hoc to paper There exists a latent variable z such that PM(Y|z, X) does not depend on the demonstrations D, so Eq. (2) holds.
    Section 3.2 introduces Eq. (2) as a transformation without stating this conditional independence. The equality is not true for an arbitrary z; it is a modeling assumption.
  • ad hoc to paper The posterior PM(z|D, X) is concentrated around a point z*, allowing the Dirac-delta substitution in the ELBO derivation.
    Appendix D replaces q(z) by delta(z-z*) and discards log delta(0) as an irrelevant infinite constant. This is a heuristic concentration assumption, not a proven limit.
  • domain assumption Scoring each demonstration independently and selecting the top-K yields a good demonstration set; interactions among demonstrations are negligible.
    Required to go from the set-level objective Eq. (1) to the single-demonstration loss Eq. (4) and the inference rule Eq. (11). The paper's Limitations section explicitly contradicts this assumption.
  • domain assumption The E5 base retriever preserves the demonstrations the LLM would prefer, so the prefilter does not remove the best candidates.
    Section 3.4 uses E5 base to shrink the pool. The Limitations section acknowledges the prefilter 'may filter out some valuable demonstrations.'
  • domain assumption The LLM feedback score s_k = P(y|(x_k,y_k), x) is a valid preference signal for demonstration usefulness.
    Used in Eq. (5) to label preferred and non-preferred demonstrations. It measures individual benefit of one demonstration to one query, not the benefit of a set.
invented entities (1)
  • Latent variable z representing LLM demonstration preference
    purpose: Bridges demonstration selection and LLM inference; z is a task-specific description with trainable parameters theta. In inference, demonstrations are ranked by P(z|(x_k,y_k), x).
    The paper provides no external falsifiable handle for z outside its own training objective; its meaning is defined by the trained LoRA parameters and is not independently measurable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to Select In-Context Demonstration Preferred by Large Language Model." pith.science (2026). https://pith.science/paper/RZ4T6YQW

@misc{pith2026250519966,
  author       = {Pith},
  title        = {Pith review of: Learning to Select In-Context Demonstration Preferred by Large Language Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RZ4T6YQW}},
  note         = {Machine review of arXiv:2505.19966}
}
read the original abstract

In-context learning (ICL) enables large language models (LLMs) to adapt to new tasks during inference using only a few demonstrations. However, ICL performance is highly dependent on the selection of these demonstrations. Recent work explores retrieval-based methods for selecting query-specific demonstrations, but these approaches often rely on surrogate objectives such as metric learning, failing to directly optimize ICL performance. Consequently, they struggle to identify truly beneficial demonstrations. Moreover, their discriminative retrieval paradigm is ineffective when the candidate pool lacks sufficient high-quality demonstrations. To address these challenges, we propose GenICL, a novel generative preference learning framework that leverages LLM feedback to directly optimize demonstration selection for ICL. Experiments on 19 datasets across 11 task categories demonstrate that GenICL achieves superior performance than existing methods in selecting the most effective demonstrations, leading to better ICL performance.

Figures

Figures reproduced from arXiv: 2505.19966 by the authors.

Figure 1
Figure 1. The distribution of the useful example ratio [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Training and inference pipeline of GENICL. conditioned on either the effective demonstration (xk, yk)w or the relatively ineffective one (xk, yk)l . This ensures that (xk, yk)w ≻ (xk, yk)l given the input x. To achieve this, we reformulate the op￾timization objectives, − log PM(Y | z, X) and − log PM(z | (Xk, Yk), X) in Eq. (4), as a two￾stage alternating optimization process. The First Stage. In this stage, we opti… view at source ↗
Figure 3
Figure 3. The predicted probability of the ground truth [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The effect of demonstration number K. To investigate the impact of the order of demon￾strations on downstream task performance, we com￾pare three different order settings: • Shuffle: the top-K selected demonstrations are randomly shuffled. • Descending: the top-K selec…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Learning to Select Visual In-Context Demonstrations

    cs.LG 2026-03 reject novelty 5.0 of 10

    A Dueling-DQN agent selects visual in-context demonstrations and outperforms kNN retrieval on objective regression benchmarks but not on subjective preference tasks, per the paper's main table.

Reference graph

Works this paper leans on

58 extracted references · 8 canonical work pages · cited by 1 Pith paper

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Sid Black, Leo Gao, Phil Wang, Connor Leahy, and Stella Biderman. 2021. Gpt-neo: Large scale autoregressive language modeling with mesh-tensorflow. If you use this software, please cite it using these metadata, 58(2)

    work page 2021

  4. [4]

    Samuel R Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. 2015. A large annotated corpus for learning natural language inference. arXiv preprint arXiv:1508.05326

    arXiv 2015

  5. [5]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877--1901

    2020

  6. [6]

    Yanda Chen, Chen Zhao, Zhou Yu, Kathleen McKeown, and He He. 2022. On the relation between sensitivity and accuracy in in-context learning. arXiv preprint arXiv:2209.07661

    work page Pith review arXiv 2022

  7. [7]

    Yunmo Chen, Tongfei Chen, Harsh Jhamtani, Patrick Xia, Richard Shin, Jason Eisner, and Benjamin Van Durme. 2024. Learning to retrieve iteratively for in-context learning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 7156--7168

    work page 2024

  8. [8]

    Daixuan Cheng, Shaohan Huang, Junyu Bi, Yuefeng Zhan, Jianfeng Liu, Yujing Wang, Hao Sun, Furu Wei, Denvy Deng, and Qi Zhang. 2023. Uprise: Universal prompt retrieval for improving zero-shot evaluation. arXiv preprint arXiv:2303.08518

    arXiv 2023

Show all 58 references
  1. [9]

    Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2017. Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30

    2017

  2. [10]

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019. Boolq: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044

    2019 arXiv

  3. [11]

    Ond r ej Du s ek, David M Howcroft, and Verena Rieser. 2019. Semantic noise matters for neural natural language generation. arXiv preprint arXiv:1911.03905

    2019 arXiv

  4. [12]

    Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306

    2024 arXiv

  5. [13]

    Alec Go, Richa Bhayani, and Lei Huang. 2009. Twitter sentiment classification using distant supervision. CS224N project report, Stanford, 1(12):2009

    2009

  6. [14]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  7. [15]

    Alex Havrilla, Yuqing Du, Sharath Chandra Raparthy, Christoforos Nalmpantis, Jane Dwivedi-Yu, Maksym Zhuravinskyi, Eric Hambro, Sainbayar Sukhbaatar, and Roberta Raileanu. 2024. Teaching large language models to reason with reinforcement learning. arXiv preprint arXiv:2403.04642

    2024 arXiv

  8. [16]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685

    2021 arXiv

  9. [17]

    Daniel Khashabi, Snigdha Chaturvedi, Michael Roth, Shyam Upadhyay, and Dan Roth. 2018. Looking beyond the surface: A challenge set for reading comprehension over multiple sentences. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Comp...

    2018

  10. [18]

    Diederik P Kingma. 2013. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114

    2013 arXiv

  11. [19]

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for C...

    2019

  12. [20]

    Jialian Li, Yipin Zhang, Wei Shen, Yuzi Yan, Jian Xie, and Dong Yan. 2024. Boosting deductive reasoning with step signals in rlhf. arXiv preprint arXiv:2410.09528

    2024 arXiv

  13. [21]

    Xiaonan Li, Kai Lv, Hang Yan, Tianyang Lin, Wei Zhu, Yuan Ni, Guotong Xie, Xiaoling Wang, and Xipeng Qiu. 2023. Unified demonstration retriever for in-context learning. arXiv preprint arXiv:2305.04320

    2023 arXiv

  14. [22]

    Xiaonan Li and Xipeng Qiu. 2023. Mot: Pre-thinking and recalling enable chatgpt to self-improve with memory-of-thoughts. arXiv preprint arXiv:2305.05181, 16

    2023 arXiv

  15. [23]

    Bill Yuchen Lin, Wangchunshu Zhou, Ming Shen, Pei Zhou, Chandra Bhagavatula, Yejin Choi, and Xiang Ren. 2019. Commongen: A constrained text generation challenge for generative commonsense reasoning. arXiv preprint arXiv:1911.03705

    2019 arXiv

  16. [24]

    Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. 2021. What makes good in-context examples for gpt- 3 ? arXiv preprint arXiv:2101.06804

    2021 arXiv

  17. [25]

    Wenhan Liu, Yutao Zhu, and Zhicheng Dou. 2024. Demorank: Selecting effective demonstrations for large language models in ranking task. arXiv preprint arXiv:2406.16332

    2024 arXiv

  18. [26]

    Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. 2021. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. arXiv preprint arXiv:2104.08786

    2021 arXiv

  19. [27]

    Man Luo, Xin Xu, Zhuyun Dai, Panupong Pasupat, Mehran Kazemi, Chitta Baral, Vaiva Imbrasaite, and Vincent Y Zhao. 2023. Dr. icl: Demonstration-retrieved in-context learning. arXiv preprint arXiv:2305.14128

    2023 arXiv

  20. [28]

    Man Luo, Xin Xu, Yue Liu, Panupong Pasupat, and Mehran Kazemi. 2024. In-context learning with retrieved demonstrations for language models: A survey. arXiv preprint arXiv:2401.11624

    2024 arXiv

  21. [29]

    Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. 2018. Can a suit of armor conduct electricity? a new dataset for open book question answering. arXiv preprint arXiv:1809.02789

    2018 arXiv

  22. [30]

    Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2022 a . https://doi.org/10.18653/v1/2022.naacl-main.201 M eta ICL : Learning to learn in context . In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Lingui...

    2022 doi

  23. [31]

    Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022 b . Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837

    2022 arXiv

  24. [32]

    Linyong Nan, Dragomir Radev, Rui Zhang, Amrit Rau, Abhinand Sivaprasad, Chiachun Hsieh, Xiangru Tang, Aadit Vyas, Neha Verma, Pranav Krishna, et al. 2020. Dart: Open-domain structured data record to text generation. arXiv preprint arXiv:2007.02871

    2020 arXiv

  25. [33]

    Courtney Napoles, Matthew R Gormley, and Benjamin Van Durme. 2012. Annotated gigaword. In Proceedings of the joint workshop on automatic knowledge base construction and web-scale knowledge extraction (AKBC-WEKEX), pages 95--100

    2012

  26. [34]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 3...

    2022

  27. [35]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36

    2024

  28. [36]

    P Rajpurkar. 2016. Squad: 100,000+ questions for machine comprehension of text. arXiv preprint arXiv:1606.05250

    2016 arXiv

  29. [37]

    N Reimers. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084

    2019 arXiv

  30. [38]

    Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval , 3(4):333--389

    2009

  31. [39]

    Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S Gordon. 2011. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In 2011 AAAI spring symposium series

    2011

  32. [40]

    Ohad Rubin, Jonathan Herzig, and Jonathan Berant. 2021. Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633

    2021 arXiv

  33. [41]

    Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2021. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64(9):99--106

    2021

  34. [42]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347

    2017 arXiv

  35. [43]

    Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language...

    2013

  36. [44]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

    2023 arXiv

  37. [45]

    Alex Wang. 2018. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461

    2018 arXiv

  38. [46]

    Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text embeddings by weakly-supervised contrastive pre-training. arXiv preprint arXiv:2212.03533

    2022 arXiv

  39. [47]

    Liang Wang, Nan Yang, and Furu Wei. 2023. Learning to retrieve in-context examples for large language models. arXiv preprint arXiv:2307.07164

    2023 arXiv

  40. [48]

    Peng Wang, Xiaobin Wang, Chao Lou, Shengyu Mao, Pengjun Xie, and Yong Jiang. 2024 a . Effective demonstration annotation for in-context learning via language model-based determinantal point process. arXiv preprint arXiv:2408.02103

    2024 arXiv

  41. [49]

    Xinyi Wang, Wanrong Zhu, Michael Saxon, Mark Steyvers, and William Yang Wang. 2024 b . Large language models are latent variable models: Explaining and finding good demonstrations for in-context learning. Advances in Neural Information Processing Systems, 36

    2024

  42. [50]

    Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. 2021. An explanation of in-context learning as implicit bayesian inference. arXiv preprint arXiv:2111.02080

    2021 arXiv

  43. [51]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. 2024. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115

    2024 arXiv

  44. [52]

    Jiacheng Ye, Zhiyong Wu, Jiangtao Feng, Tao Yu, and Lingpeng Kong. 2023. Compositional exemplars for in-context learning. In International Conference on Machine Learning, pages 39818--39833. PMLR

    2023

  45. [53]

    Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830

    2019 arXiv

  46. [54]

    Rui Zhang and Joel Tetreault. 2019. This email could save your life: Introducing the task of email subject line generation. arXiv preprint arXiv:1906.03497

    2019 arXiv

  47. [55]

    Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. Advances in neural information processing systems, 28

    2015

  48. [56]

    Yuan Zhang, Jason Baldridge, and Luheng He. 2019. Paws: Paraphrase adversaries from word scrambling. arXiv preprint arXiv:1904.01130

    2019 arXiv

  49. [57]

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595--46623

    2023

  50. [58]

    Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. 2019. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593

    2019 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.