Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

DICE: Dynamic In-Context Example Selection in LLM Agents via Efficient Knowledge Transfer

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read DICE argues that dynamically re-selecting in-context demonstrations at every agent step, by maximizing estimated transferable knowledge, makes LLM agents more accurate and robust while tightening generalization bounds.

desk verdict Solid empirical plug-in with consistent gains; the formal guarantee is overclaimed and the theory currently decouples from the implementation. read the letter →

arxiv 2507.23554 v1 pith:C6647QZN submitted 2025-07-31 cs.AI

Ruoyu Wang , Junda Wu , Yu Xia , Tong Yu , Ryan A. Rossi , Julian McAuley , Lina Yao This is my paper
classification cs.AI
keywords in-contextlearningdemonstrationselectionLLMagentstransferableknowledgeinformationbottleneckgeneralizationbounddynamicpromptingagenticreasoning
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

LLM agents often fail or degrade when the examples prepended to their prompt are irrelevant to the current step, and static demonstration sets are especially fragile in long-horizon tasks. DICE is a training-free, plug-in module that re-selects a fresh set of demonstrations at each timestep, scoring candidates by how much 'transferable knowledge' they share with the agent's evolving context. The scoring rule is derived from a causal decomposition of demonstration knowledge plus an information-bottleneck criterion, and the paper proves that following it yields a strictly tighter generalization-gap bound than using unreduced demonstrations. On HotpotQA, ALFWorld, and Webshop, adding DICE to ReAct, Reflexion, or LATS improves exact match, success rate, and score over static baselines, and stepwise selection beats taskwise selection. If the paper is right, dynamic, context-aware demonstration retrieval is a general and cheap way to make LLM agents more accurate and robust.

What carries the argument

The load-bearing objects are the causal graph $D \leftarrow TK \rightarrow A_t$ (with $\epsilon_D \rightarrow D$ and $H_t \rightarrow A_t$) and the information-bottleneck selection criterion $J_i = I(d_i; TK_{d_i}) - \beta I(TK_{d_i}; A_t)$. The graph identifies why non-transferable noise in a demonstration leaks into the agent's next action through a collider, and the criterion turns the causal story into a ranking: pick the demo whose TK is most predictive of $A_t$ while carrying least extra detail. To make this computable, the paper instantiates TK with summaries from a small open-source language model, assumes the retriever is a fixed-capacity channel so $I(d; TK_d)$ is constant across candidates, and uses an InfoNCE/cosine-similarity objective between $TK_d$ and a $TK_t$ proxy of $A_t$. The theory rides on two generalization-gap bounds: Theorem 2.2 bounds the gap by $\tilde O(\sqrt{(I(X;T)+1)/n})$, and Theorem 2.3 bounds it by $\tilde O(\sqrt{I(X;T|Y)/n})$, which is what ties the selection criterion to a strictly smaller gap.

What would settle it

Run DICE on the same benchmarks but replace the cosine-similarity retrieval with a brute-force oracle that evaluates every candidate demonstration subset by actually rolling out the agent; if the oracle pick consistently beats DICE's pick, the InfoNCE and cosine approximation is not capturing the mutual information the theory requires.

Watch

Extended reading notes

Core claim

The paper's central claim is that a demonstration helps an agent only through its transferable knowledge (TK), the information shared between the demonstration and the current decision, while the rest is task-specific noise that can hurt via a collider path $D \leftarrow TK \rightarrow A_t$ with $\epsilon_D \rightarrow D$. Conditioning on $D$ opens the backdoor $\epsilon_D \rightarrow D \leftarrow TK \rightarrow A_t$, so spurious associations can drive the next action. DICE therefore selects, at each step $t$, the demonstrations minimizing $I(d; TK_d) - \beta I(TK_d; A_t)$, which for $\beta = 1$ equals $I(d; TK_d | A_t)$; by Theorem 2.3 this makes the generalization-gap bound strictly smaller than for unreduced examples. Since exact mutual information is intractable, the paper uses a fixed-capacity knowledge retriever, a small open-source language model, to summarize demonstrations and the current context, assumes $I(d; TK_d)$ is constant across candidates, and replaces $I(TK_d; A_t)$ with an InfoNCE estimate of similarity between $TK_d$ and $TK_t$. The paper reports that this practical rule improves HotpotQA exact match by 7.3–9.3 points, ALFWorld success rate by 7.5–10.4 points, and Webshop score and success rate by 1.2–4.4 and 1.5–7.0 points respectively over the corresponding ReAct, Reflexion, and LATS baselines.

Load-bearing premise

The whole argument hinges on the retrieval module faithfully implementing the theoretical criterion: the fixed summary channel must treat every candidate demonstration equally, and the summary of the current context must predict the agent's next action well enough.

Editorial extensions

If this is right

  • Adding DICE to ReAct, Reflexion, and LATS improves HotpotQA exact match by 7.3–9.3 points, ALFWorld success rate by 7.5–10.4 points, and Webshop score and success rate by 1.2–4.4 and 1.5–7.0 points respectively.
  • Stepwise selection consistently beats taskwise selection, so the benefit comes from adapting demonstrations to the evolving context, not just from better static examples.
  • DICE maintains or widens its edge over standard ICL even when the pool contains only low-quality demonstrations, so it reduces the need for large or carefully curated demo pools.
  • Because DICE is training-free and framework-agnostic, any existing agent that prepends demonstrations can adopt it without re-training or changing the base model.

Reading between the lines

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

  • A likely reason DICE works is that the summary of the current trajectory acts as a dynamic query into a library of successful plans; if so, the same recipe should transfer to other stateful in-context learners, not just tool-use agents.
  • The theory and the implementation are only loosely coupled: the fixed-capacity and TK-proxy assumptions are untested, so a natural next step is a trainable encoder that estimates $I(d; TK_d)$ directly instead of assuming it constant.
  • DICE's per-step retrieval could be combined with memory or reflection mechanisms: instead of a fixed pool, successful trajectories generated online by the agent itself could replenish the demonstration pool, making the system self-improving without training.
  • The reported gains may come partly from prompt diversity or recency effects rather than from the information-bottleneck criterion; an ablation that compares DICE's scores against random, recency-based, or kNN retrieval with the same per-step refresh would isolate the mechanism.
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 / 5 minor

Summary. The paper proposes DICE, a training-free method that dynamically selects in-context demonstrations at each agent step by maximizing a transferable-knowledge criterion. The method is motivated by a causal graph in which demonstrations contain transferable and non-transferable components, and is implemented by using gemma-2-2b-it to summarize candidate demonstrations and the current context, then ranking by cosine similarity via an InfoNCE-style objective. The authors claim a formal guarantee of improved agent performance from tighter generalization-gap bounds and support the method with experiments on HotpotQA, ALFWorld, and Webshop using ReAct, Reflexion, and LATS baselines, reporting consistent gains (e.g., +7.3 to +9.3 exact match on HotpotQA, +7.5 to +10.4 success rate on ALFWorld, and +1.2 to +7.0 on Webshop metrics).

Significance. If the reported gains are stable, DICE is a practically valuable, framework-agnostic plug-in that requires no additional training; the stepwise-versus-taskwise ablation in Table 4 supports the value of dynamic selection, and the causal/IB framing is a useful organizing perspective. The paper also deserves credit for running experiments across three benchmarks and three different agent frameworks. However, the central formal claim is currently overclaimed: the presented bounds are upper bounds on a generalization gap, not performance guarantees, and the implemented retrieval objective is not rigorously shown to optimize the selection criterion used in the theory. The empirical claims would be substantially stronger with repeated-seed statistics, significance tests, and released code and data.

major comments (4)
  1. [Section 2.4, Eq. (1) and InfoNCE objective]
  2. [Section 2.3 and Abstract] The paper repeatedly states that the method provides a "formal guarantee of improved agent performance" (Abstract and Section 2.3). The actual results are upper bounds on a generalization gap: Theorem 2.2 gives Δ ≤ Õ(sqrt((I(X;T)+1)/n)), and Theorem 2.3 gives a similar conditional-mutual-information bound. A tighter upper bound does not imply a smaller generalization gap, and no argument connects these bounds to Exact Match or success rate. Moreover, the reduction Ji = I(di; TKi) − I(TKi; At) = I(di; TKi | At) when β = 1 is not generally true; it relies on conditional independence d ⊥ At | TK from the causal graph in Figure 2, which is assumed rather than derived or tested. The formal claim should either be substantiated with a direct performance guarantee or withdrawn in favor of a more modest statement about the selection criterion.
  3. [Definition 2.1 and Section 2.2] The generalization-gap framework assumes that the encoder T is fixed independently of the training data and that (X, Y) are i.i.d. samples. In DICE, demonstrations are selected online using the current context H_t, which includes observations from the test instance, so the selected d is not independent of the test instance and the i.i.d. assumption in Definition 2.1 is violated. The paper does not explain how the bound applies to instance-dependent, dynamically selected demonstrations; this gap affects the validity of the theoretical argument as stated.
  4. [Section 3, Tables 1–3] All experimental results are reported as single numbers with no standard deviations, number of seeds, or significance tests. Given that the reported gains vary from +1.2 to +10.4 points across settings, and that some gains are small (e.g., Webshop LATS +1.2 score and +1.5 success rate), it is not possible to assess whether the improvements are statistically reliable or whether the "consistent and robust" claim is supportable. The manuscript also does not provide code or data links, which limits reproducibility. Please report variance across multiple runs and, where feasible, significance tests, and release the implementation and demo pools.
minor comments (5)
  1. [Section 2.3] The text says "generalization gap of a classifier built on top of Z" but Z is not defined; it should refer to the representation T used in the theorem.
  2. [Section 2.3] The notation Õ is used without definition, and all bounds suppress constant factors; this is acceptable but should be stated explicitly.
  3. [Throughout] The benchmark name is spelled inconsistently as "AlfWorld" in several places and "ALFWorld" in the table and references; please standardize.
  4. [Section 3.1] The paper references an Appendix for the prompting strategy, proof details, and additional results, but no appendix is included in the submitted manuscript, making these details unavailable for review.
  5. [Section 3.4, Table 4] The taskwise/stepwise comparison is only shown for ReAct; adding the same ablation for Reflexion and LATS would more strongly support the claim that stepwise selection is generally beneficial.

Circularity Check

1 steps flagged · score 2.0 of 10

The Section 2.3 'formal guarantee' is partly a restatement of the selection criterion: Eq. 1 at beta=1 minimizes the same mutual-information quantity that Theorem 2.3's bound depends on. The empirical results are not circular.

  1. self definitional [Section 2.3, Theoretical Guarantee, paragraph after Theorem 2.3]
    "when β = 1, the criterion in Equation 1 can be reduced to: Ji = I(di; T Ki) − I(T Ki; At) = I(di; T Ki | At) ... According to the generalization bound from Theorem 2.3, which scales with I(d; T K| At), we obtain ... Thus, selecting demonstrations based on our proposed criterion provably yields a tighter generalization bound."

    The selection rule is defined (Equation 1 with β=1) to minimize I(d;TK|At), and Theorem 2.3's bound is a monotone function of I(X;T|Y). Therefore 'selecting by the criterion tightens the bound' follows by direct substitution of the criterion into the theorem; it is not an independent discovery. The advertised 'formal guarantee of improved agent performance' also overstates the step: a tighter upper bound on the generalization gap does not by itself imply lower error, so the guarantee is a restatement of the objective rather than a demonstrated performance consequence.

full rationale

The paper's empirical claim is not circular: DICE is evaluated on external benchmarks (HotpotQA, ALFWorld, Webshop) with a non-overlapping demonstration-pool construction, no trainable selection parameters, and a fixed pretrained retriever (gemma-2-2b-it), so the reported gains are not fitted inputs renamed as predictions. The theoretical sections apply external information-theoretic results (Xu and Raginsky; Kawaguchi et al.), which are independent support. The one circular flavor is that the proposed criterion is deliberately chosen to minimize the mutual-information quantity appearing in Theorem 2.3's bound, making the 'tighter bound' conclusion a near-tautological application of the theorem rather than an empirically or theoretically independent result. Section 2.4's implementation relies on untested assumptions, notably a fixed-capacity retriever with constant I(d;TKd) and TK_t as a proxy for the unobserved At; if these fail, the deployed selector is not shown to optimize the criterion that the bound covers, but that is a correctness risk, not circularity. Self-citations in the related-work section ([12], [25]) are not load-bearing. Overall, the central derivation is largely self-contained, with only a mild self-definitional overlap between the selection objective and the bound it claims to improve.

Assumptions & free parameters 3 free parameters · 5 assumptions · 2 invented entities

No parameters are fitted to optimize the benchmarks, but several design choices such as β, M, pool size, and the TK prompt are set by hand and not reported. The 'transferable knowledge' construct is the main invented quantity and lacks independent evidence. The causal graph and the TK_t proxy are the weakest links between theory and implementation.

free parameters (3)
  • β (regularization coefficient in Eq. 1) = not reported; analysis sets β=1
    Controls the trade-off between compressing TK and preserving action-relevant information in the selection objective, but the implemented retriever never explicitly tunes it.
  • M (number of demonstrations selected per step) = not reported
    Number of trajectories retrieved at each timestep; core to the method, value absent from the main text.
  • Demonstration pool size N = not reported
    Pool is constructed from successful baseline trajectories; its size and the training/evaluation split are not specified.
assumptions (5)
  • domain assumption Causal graph D ← TK → At, εD → D, Ht → At is the correct account of how demonstrations influence agent actions.
    Section 2.1 states the graph and uses it to justify selecting against spurious dependencies; no identification, counterfactual, or test is offered.
  • domain assumption Conditioning on D opens a collider path εD → D ← TK → At that impairs generalization.
    Section 2.1 interprets the causal graph as explaining ICL instability; this is an interpretive claim, not proven.
  • ad hoc to paper A fixed-capacity pretrained LLM channel makes I(d;TK_d) constant across all candidate demonstrations.
    Section 2.4 asserts this to simplify Eq. (1); no capacity measurement or MI estimation supports it.
  • ad hoc to paper TK_t extracted from current context is an adequate proxy for the unobserved next action At.
    Section 2.4 substitutes TK_t for At and uses InfoNCE; the proxy is unvalidated.
  • standard math Published information-bottleneck generalization bounds (Xu and Raginsky; Kawaguchi et al.) apply to this agentic ICL setting with the stated substitutions.
    Section 2.3 imports Theorem 2.2 from [52] and Theorem 2.3 from [10]; the paper does not derive or adapt them to the causal graph.
invented entities (2)
  • Transferable Knowledge (TK)
    purpose: Latent semantic component of demonstrations that should influence the agent's next action; used as the retrieval key.
    TK is operationalized only through LLM prompts and cosine similarity; there is no external or falsifiable handle to confirm it captures a real causal quantity.
  • Task-specific noise εD
    purpose: Explains spurious correlations opened by conditioning on demonstrations in the causal graph.
    Purely conceptual noise variable in Section 2.1; never measured or falsified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DICE: Dynamic In-Context Example Selection in LLM Agents via Efficient Knowledge Transfer." pith.science (2026). https://pith.science/paper/C6647QZN

@misc{pith2026250723554,
  author       = {Pith},
  title        = {Pith review of: DICE: Dynamic In-Context Example Selection in LLM Agents via Efficient Knowledge Transfer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C6647QZN}},
  note         = {Machine review of arXiv:2507.23554}
}
read the original abstract

Large language model-based agents, empowered by in-context learning (ICL), have demonstrated strong capabilities in complex reasoning and tool-use tasks. However, existing works have shown that the effectiveness of ICL is highly sensitive to the choice of demonstrations, with suboptimal examples often leading to unstable or degraded performance. While prior work has explored example selection, including in some agentic or multi-step settings, existing approaches typically rely on heuristics or task-specific designs and lack a general, theoretically grounded criterion for what constitutes an effective demonstration across reasoning steps. Therefore, it is non-trivial to develop a principled, general-purpose method for selecting demonstrations that consistently benefit agent performance. In this paper, we address this challenge with DICE, Dynamic In-Context Example Selection for LLM Agents, a theoretically grounded ICL framework for agentic tasks that selects the most relevant demonstrations at each step of reasoning. Our approach decomposes demonstration knowledge into transferable and non-transferable components through a causal lens, showing how the latter can introduce spurious dependencies that impair generalization. We further propose a stepwise selection criterion with a formal guarantee of improved agent performance. Importantly, DICE is a general, framework-agnostic solution that can be integrated as a plug-in module into existing agentic frameworks without any additional training cost. Extensive experiments across diverse domains demonstrate our method's effectiveness and generality, highlighting the importance of principled, context-aware demo selection for robust and efficient LLM agents.

Figures

Figures reproduced from arXiv: 2507.23554 by the authors.

Figure 1
Figure 1. Overview of our stepwise dynamic demonstration selection framework. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Causal graph show￾ing how demo-specific noise ϵD affects the next action At in ICL, introducing spurious correlations that can hinder agent performance. Intuitively, when using in-context learning (ICL) to guide an agent, the provided examples inevitably contain a mixture of knowl￾edge—some of which is relevant and transferable to the new task, and some of which is not. While the transferable knowledge can be transf… view at source ↗
Figure 3
Figure 3. (a) Using demonstrations with higher average score in DICE results in better performance. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Case studies. Top: Without DICE, the agent overlooks the correct answer already present in Observation 2 and unnecessarily attempts a Lookup[middle name], resulting in failure. With DICE, the agent correctly extracts the middle name directly from the observation. Botto…

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. OLIVIA: Online Learning via Inference-time Action Adaptation for Decision Making in LLM ReAct Agents

    cs.AI 2026-05 unverdicted novelty 7.0 of 10

    OLIVIA treats LLM agent action selection as a contextual linear bandit over frozen hidden states and applies UCB exploration to adapt online, yielding consistent gains over static ReAct and prompt-based baselines on f...

  2. Skill-CMIB: Multimodal Agent Skill for Consistent Action via Conditional Multimodal Information Bottleneck

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    CMIB uses a conditional multimodal information bottleneck to create reusable agent skills that separate verbalizable text content from predictive perceptual residuals, improving execution stability.

  3. Generate, Filter, Control, Replay: A Comprehensive Survey of Rollout Strategies for LLM Reinforcement Learning

    cs.LG 2026-04 unverdicted novelty 7.0 of 10

    This survey introduces the Generate-Filter-Control-Replay (GFCR) taxonomy to structure rollout pipelines for RL-based post-training of reasoning LLMs.

Reference graph

Works this paper leans on

62 extracted references · 16 canonical work pages · cited by 3 Pith papers

  1. [1]

    What learn- ing algorithm is in-context learning? investigations with linear models

    Ekin Akyürek, Dale Schuurmans, Jacob Andreas, Tengyu Ma, and Denny Zhou. What learn- ing algorithm is in-context learning? investigations with linear models. In The Eleventh International Conference on Learning Representations

    work page

  2. [2]

    Language models are few-shot learners

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

    1901

  3. [3]

    Scaling instruction-finetuned language models

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. Journal of Machine Learning Research, 25(70):1–53, 2024

    2024

  4. [4]

    Why can gpt learn in-context? language models secretly perform gradient descent as meta-optimizers

    Damai Dai, Yutao Sun, Li Dong, Yaru Hao, Shuming Ma, Zhifang Sui, and Furu Wei. Why can gpt learn in-context? language models secretly perform gradient descent as meta-optimizers. In Findings of the Association for Computational Linguistics: ACL 2023, pages 4005–4019, 2023

    2023

  5. [5]

    A survey on in-context learning

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022

    arXiv 2022

  6. [6]

    Demystifying prompts in language models via perplexity estimation

    Hila Gonen, Srini Iyer, Terra Blevins, Noah A Smith, and Luke Zettlemoyer. Demystifying prompts in language models via perplexity estimation. arXiv preprint arXiv:2212.04037, 2022

    arXiv 2022

  7. [7]

    Pre-Training to Learn in Context

    Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Pre-training to learn in context. arXiv preprint arXiv:2305.09137, 2023

    work page Pith review arXiv 2023

  8. [8]

    Structured prompting: Scaling in-context learning to 1,000 examples

    Yaru Hao, Yutao Sun, Li Dong, Zhixiong Han, Yuxian Gu, and Furu Wei. Structured prompting: Scaling in-context learning to 1,000 examples. arXiv preprint arXiv:2212.06713, 2022

    arXiv 2022

Show all 62 references
  1. [9]

    Opt-iml: Scaling language model instruction meta learning through the lens of generalization

    Srinivasan Iyer, Xi Victoria Lin, Ramakanth Pasunuru, Todor Mihaylov, Daniel Simig, Ping Yu, Kurt Shuster, Tianlu Wang, Qing Liu, Punit Singh Koura, et al. Opt-iml: Scaling language model instruction meta learning through the lens of generalization. arXiv preprint arXiv:2212.1...

    2022 arXiv

  2. [10]

    How does information bottleneck help deep learning? In International Conference on Machine Learning, pages 16049–16096

    Kenji Kawaguchi, Zhun Deng, Xu Ji, and Jiaoyang Huang. How does information bottleneck help deep learning? In International Conference on Machine Learning, pages 16049–16096. PMLR, 2023

    2023

  3. [11]

    Self-generated in-context learning: Leveraging auto-regressive language models as a demonstration generator

    Hyuhng Joon Kim, Hyunsoo Cho, Junyeob Kim, Taeuk Kim, Kang Min Yoo, and Sang-goo Lee. Self-generated in-context learning: Leveraging auto-regressive language models as a demonstration generator. arXiv preprint arXiv:2206.08082, 2022

    2022 arXiv

  4. [12]

    The closeness of in-context learning and weight shifting for softmax regression

    Shuai Li, Zhao Song, Yu Xia, Tong Yu, and Tianyi Zhou. The closeness of in-context learning and weight shifting for softmax regression. Advances in Neural Information Processing Systems, 37:62584–62616, 2024

    2024

  5. [13]

    Finding supporting examples for in-context learning

    Xiaonan Li and Xipeng Qiu. Finding supporting examples for in-context learning. CoRR, 2023

    2023

  6. [14]

    Mend: Meta demonstration distillation for efficient and effective in-context learning

    Yichuan Li, Xiyao Ma, Sixing Lu, Kyumin Lee, Xiaohu Liu, and Chenlei Guo. Mend: Meta demonstration distillation for efficient and effective in-context learning. arXiv preprint arXiv:2403.06914, 2024. 10

    2024 arXiv

  7. [15]

    What makes good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804, 2021

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

    2021 arXiv

  8. [16]

    In-context vectors: Making in context learning more effective and controllable through latent space steering

    Sheng Liu, Haotian Ye, Lei Xing, and James Zou. In-context vectors: Making in context learning more effective and controllable through latent space steering. arXiv preprint arXiv:2311.06668, 2023

    2023 arXiv

  9. [17]

    Let’s learn step by step: Enhancing in-context learning ability with curriculum learning

    Yinpeng Liu, Jiawei Liu, Xiang Shi, Qikai Cheng, Yong Huang, and Wei Lu. Let’s learn step by step: Enhancing in-context learning ability with curriculum learning. arXiv preprint arXiv:2402.10738, 2024

    2024 arXiv

  10. [18]

    Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity

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

    2021 arXiv

  11. [19]

    Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity

    Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1...

    2022

  12. [20]

    Wilbur: Adaptive in-context learning for robust and accurate web agents

    Michael Lutz, Arth Bohra, Manvel Saroyan, Artem Harutyunyan, and Giovanni Campagna. Wilbur: Adaptive in-context learning for robust and accurate web agents. arXiv preprint arXiv:2404.05902, 2024

    2024 arXiv

  13. [21]

    Laser: Llm agent with state-space exploration for web navigation

    Kaixin Ma, Hongming Zhang, Hongwei Wang, Xiaoman Pan, Wenhao Yu, and Dong Yu. Laser: Llm agent with state-space exploration for web navigation. arXiv preprint arXiv:2309.08172, 2023

    2023 arXiv

  14. [22]

    Active learning principles for in-context learning with large language models

    Katerina Margatina, Timo Schick, Nikolaos Aletras, and Jane Dwivedi-Yu. Active learning principles for in-context learning with large language models. arXiv preprint arXiv:2305.14264, 2023

    2023 arXiv

  15. [23]

    Metaicl: Learning to learn in context

    Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. Metaicl: Learning to learn in context. arXiv preprint arXiv:2110.15943, 2021

    2021 arXiv

  16. [24]

    Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. Rethinking the role of demonstrations: What makes in-context learning work? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, page...

    2022

  17. [25]

    Gui agents: A survey

    Dang Nguyen, Jian Chen, Yu Wang, Gang Wu, Namyong Park, Zhengmian Hu, Hanjia Lyu, Junda Wu, Ryan Aponte, Yu Xia, et al. Gui agents: A survey. arXiv preprint arXiv:2412.13501, 2024

    2024

  18. [26]

    In-context example selection with influences

    Tai Nguyen and Eric Wong. In-context example selection with influences. arXiv preprint arXiv:2302.11042, 2023

    2023 arXiv

  19. [27]

    Representation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

    2018 arXiv

  20. [28]

    In-context learning with iterative demonstration selection

    Chengwei Qin, Aston Zhang, Chen Chen, Anirudh Dagar, and Wenming Ye. In-context learning with iterative demonstration selection. arXiv preprint arXiv:2310.09881, 2023

    2023 arXiv

  21. [29]

    In-context learning with iterative demonstration selection

    Chengwei Qin, Aston Zhang, Chen Chen, Anirudh Dagar, and Wenming Ye. In-context learning with iterative demonstration selection. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 7441–7455, 2024

    2024

  22. [30]

    Learning to retrieve prompts for in-context learning

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

    2021 arXiv

  23. [31]

    Toolformer: Language models can teach themselves to use tools

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36: 68539–68551, 2023

    2023

  24. [32]

    In-context pretraining: Language modeling beyond document boundaries

    Weijia Shi, Sewon Min, Maria Lomeli, Chunting Zhou, Margaret Li, Gergely Szilvasy, Rich James, Xi Victoria Lin, Noah A Smith, Luke Zettlemoyer, et al. In-context pretraining: Language modeling beyond document boundaries. arXiv preprint arXiv:2310.10638, 2023

    arXiv

  25. [33]

    Reflexion: Language agents with verbal reinforcement learning.Advances in Neural Information Processing Systems, 36:8634–8652, 2023

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning.Advances in Neural Information Processing Systems, 36:8634–8652, 2023. 11

    2023

  26. [34]

    Alfred: A benchmark for interpreting grounded instructions for everyday tasks

    Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mot- taghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In Proceedings of the IEEE/CVF conference on computer vision and patter...

    2020

  27. [35]

    Alfworld: Aligning text and embodied environments for interactive learning

    Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. Alfworld: Aligning text and embodied environments for interactive learning. arXiv preprint arXiv:2010.03768, 2020

    2010 arXiv

  28. [36]

    An information-theoretic approach to prompt engineering without ground truth labels

    Taylor Sorensen, Joshua Robinson, Christopher Michael Rytting, Alexander Glenn Shaw, Kyle Jeffrey Rogers, Alexia Pauline Delorey, Mahmoud Khalil, Nancy Fulda, and David Wingate. An information-theoretic approach to prompt engineering without ground truth labels. arXiv preprint...

    2022 arXiv

  29. [37]

    Selective annotation makes language models better few-shot learners

    Hongjin Su, Jungo Kasai, Chen Henry Wu, Weijia Shi, Tianlu Wang, Jiayi Xin, Rui Zhang, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, et al. Selective annotation makes language models better few-shot learners. arXiv preprint arXiv:2209.01975, 2022

    2022 arXiv

  30. [38]

    Multi- lingual llms are better cross-lingual in-context learners with alignment

    Eshaan Tanwar, Subhabrata Dutta, Manish Borthakur, and Tanmoy Chakraborty. Multi- lingual llms are better cross-lingual in-context learners with alignment. arXiv preprint arXiv:2305.05940, 2023

    2023 arXiv

  31. [39]

    Gemma 2: Improving open language models at a practical size

    Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024

    2024 arXiv

  32. [40]

    In-context learning demonstration selection via influence analysis

    Minh-Hao Van, Xintao Wu, et al. In-context learning demonstration selection via influence analysis. arXiv preprint arXiv:2402.11750, 2024

    2024 arXiv

  33. [41]

    Curriculum demonstration selection for in-context learning

    Duc Anh Vu, Nguyen Tran Cong Duy, Xiaobao Wu, Hoang Minh Nhat, Du Mingzhe, Nguyen Thanh Thong, and Anh Tuan Luu. Curriculum demonstration selection for in-context learning. arXiv preprint arXiv:2411.18126, 2024

    2024 arXiv

  34. [42]

    A survey on large language model based autonomous agents

    Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6):186345, 2024

    2024

  35. [43]

    Learning to retrieve in-context examples for large language models

    Liang Wang, Nan Yang, and Furu Wei. Learning to retrieve in-context examples for large language models. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1752–1767, 2024

    2024

  36. [44]

    Large language models are implicitly topic models: Explaining and finding good demonstrations for in-context learning

    Xinyi Wang, Wanrong Zhu, and William Yang Wang. Large language models are implicitly topic models: Explaining and finding good demonstrations for in-context learning. arXiv preprint arXiv:2301.11916, 1:15, 2023

    2023 arXiv

  37. [45]

    Demonstration selection for in-context learning via reinforcement learning

    Xubin Wang, Jianfei Wu, Yichen Yuan, Mingzhe Li, Deyu Cai, and Weijia Jia. Demonstration selection for in-context learning via reinforcement learning. arXiv preprint arXiv:2412.03966, 2024

    2024 arXiv

  38. [46]

    Super-naturalinstructions: Generalization via declarative instructions on 1600+ nlp tasks

    Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, Anjana Arunkumar, Arjun Ashok, Arut Selvan Dhanasekaran, Atharva Naik, David Stap, et al. Super-naturalinstructions: Generalization via declarative instructions on 1600+ nlp tasks. arXiv pre...

    2022 arXiv

  39. [47]

    Finetuned language models are zero-shot learners

    Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021

    2021 arXiv

  40. [49]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  41. [50]

    Self-adaptive in-context learning: An information compression perspective for in-context example selection and ordering

    Zhiyong Wu, Yaoxiang Wang, Jiacheng Ye, and Lingpeng Kong. Self-adaptive in-context learning: An information compression perspective for in-context example selection and ordering. arXiv preprint arXiv:2212.10375, 2022. 12

    2022 arXiv

  42. [51]

    Self-adaptive in-context learning: An information compression perspective for in-context example selection and ordering

    Zhiyong Wu, Yaoxiang Wang, Jiacheng Ye, and Lingpeng Kong. Self-adaptive in-context learning: An information compression perspective for in-context example selection and ordering. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volum...

    2023

  43. [52]

    Information-theoretic analysis of generalization capability of learning algorithms

    Aolin Xu and Maxim Raginsky. Information-theoretic analysis of generalization capability of learning algorithms. Advances in neural information processing systems, 30, 2017

    2017

  44. [53]

    Misconfidence-based demonstration selection for llm in-context learning

    Shangqing Xu and Chao Zhang. Misconfidence-based demonstration selection for llm in-context learning. arXiv preprint arXiv:2401.06301, 2024

    2024 arXiv

  45. [54]

    Cops: Empowering llm agents with provable cross-task experience sharing

    Chen Yang, Chenyang Zhao, Quanquan Gu, and Dongruo Zhou. Cops: Empowering llm agents with provable cross-task experience sharing. arXiv preprint arXiv:2410.16670, 2024

    2024 arXiv

  46. [55]

    Hotpotqa: A dataset for diverse, explainable multi-hop question answering

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhut- dinov, and Christopher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. arXiv preprint arXiv:1809.09600, 2018

    2018 arXiv

  47. [56]

    Webshop: Towards scalable real-world web interaction with grounded language agents

    Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems, 35:20744–20757, 2022

    2022

  48. [57]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR), 2023

    2023

  49. [58]

    Compositional ex- emplars for in-context learning

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

    2023

  50. [59]

    Active example selection for in-context learning

    Yiming Zhang, Shi Feng, and Chenhao Tan. Active example selection for in-context learning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9134–9148, 2022

    2022

  51. [60]

    Expel: Llm agents are experiential learners

    Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. Expel: Llm agents are experiential learners. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19632–19642, 2024

    2024

  52. [61]

    Dynamic demonstrations controller for in-context learning

    Fei Zhao, Taotian Pang, Zhen Wu, Zheng Ma, Shujian Huang, and Xinyu Dai. Dynamic demonstrations controller for in-context learning. arXiv preprint arXiv:2310.00385, 2023

    2023 arXiv

  53. [62]

    Calibrate before use: Improving few-shot performance of language models

    Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use: Improving few-shot performance of language models. In International conference on machine learning, pages 12697–12706. PMLR, 2021

    2021

  54. [63]

    Language agent tree search unifies reasoning acting and planning in language models

    Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu-Xiong Wang. Language agent tree search unifies reasoning acting and planning in language models. arXiv preprint arXiv:2310.04406, 2023. 13

    2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.