Pith. sign in

REVIEW 5 major objections 6 minor 71 references

Selecting in-context examples for both relevance and diversity, via submodular mutual information, improves few-shot performance.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A submodular mutual information framework for selecting and training in-context learning exemplars improves average accuracy on nine benchmarks by about five points over the IDEAL baseline.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection InSQuaD is a plausible SMI-based approach to ICL selection with a real training contribution, but the headline gains are not convincingly attributed to the learned retriever; the paper deserves peer review but needs substantial revision. the 5 major comments →

arxiv 2508.21003 v1 pith:DTLARONO submitted 2025-08-28 cs.LG

InSQuAD: In-Context Learning for Efficient Retrieval via Submodular Mutual Information to Enforce Quality and Diversity

Souradeep Nanda , Anay Majee , Rishabh Iyer This is my paper
classification cs.LG
keywords in-context learningexemplar selectionsubmodular mutual informationretrieval model trainingdiversitylikelihood-based lossparaphrase augmentationfew-shot learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

In-context learning hinges on which demonstrations are put into the prompt, and this paper claims that those demonstrations should be chosen as a set, not one-by-one: relevant to the query but also diverse among themselves. The authors build both halves of the pipeline from submodular mutual information, a measure of how much information a candidate set shares with a target while rewarding coverage. The same SMI objective is then converted into a trainable loss that teaches a sentence-embedding retriever to prefer relevant documents and suppress paraphrased distractors, trained on a multi-hop question-answering corpus augmented with synthetic paraphrases. Across nine classification, multi-choice, dialogue, and generation benchmarks, the combined method improves over the baselines it compares against, up to 21.6% on classification, 16.4% on multi-choice, and 7% on generation. If the claim holds, ICL systems can get quality, diversity, and exemplar ordering from one combinatorial formulation.

Core claim

This paper sets out to show that in-context learning improves when exemplar selection is treated as a set-selection problem rather than a nearest-neighbor search. It defines InSQuaD-RETRIEVE, which uses submodular mutual information to shortlist a diverse annotation pool and then, for each test query, greedily picks a small set of exemplars that share high mutual information with the query while remaining diverse among themselves. It then defines InSQuaD-LEARN, which treats retrieval as a submodular point process and trains a sentence-embedding retriever by minimizing the log-ratio of the SMI of distractors to the SMI of relevant documents, split into a quality term and a paraphrase-driven d

What carries the argument

The load-bearing object is submodular mutual information (SMI): for a submodular set function f, SMI If(A;Q) measures how much information a candidate set A shares with a query set Q, so maximizing it pulls in items that are relevant to Q while discouraging redundancy within A. InSQuaD uses three instantiations—Graph-Cut, Facility-Location, and Log-Determinant—both in a greedy targeted-selection algorithm that yields an exemplar ordering by decreasing information gain, and in a Submodular Point Process likelihood. The training objective is the negative log-ratio between the SMI of distractors and the SMI of relevant documents, split into a quality term and a paraphrase-driven diversity term;

Load-bearing premise

The diversity signal learned by the retriever comes entirely from synthetic paraphrases of one question-answering dataset; if those paraphrases are noisy or unlike the language of the downstream tasks, the diversity objective may not transfer.

What would settle it

Train the retriever with paraphrases replaced by random, unrelated documents while keeping everything else fixed; if performance on the nine benchmarks stays the same, the diversity loss is not the source of the gains. A second check: measure the lexical and semantic overlap between generated paraphrases and their originals—near-duplicate paraphrases would leave the diversity objective with almost no signal.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • ICL pipelines can use a single SMI-based selection step instead of separate relevance and diversity stages, with lower inference cost.
  • Retrievers for ICL benefit from training on diversity, not just query relevance, because paraphrase suppression is learnable.
  • The Graph-Cut variant offers a practical default: best average performance and the fastest inference among the three instances.
  • Annotation budget size is not the main driver; targeted selection maintains performance even when the candidate pool grows.
  • Greedy SMI maximization supplies a principled ordering of exemplars by decreasing information gain, addressing prompt-order sensitivity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the paraphrase-driven diversity loss is the active ingredient, then the quality of paraphrases should be measurable in downstream results; replacing them with human-written or adversarial paraphrases would be a direct stress test.
  • The same likelihood-based SMI training could transfer to retrieval-augmented generation and other settings where the selected context must balance relevance and coverage.
  • Since the optimal trade-off parameter λ differs by task, practical use would need an automatic calibration rule; the paper reports the sensitivity but leaves that open.
  • The budget ablation suggests diminishing returns from annotating more examples once diversity is accounted for, a testable prediction for active-learning pipelines.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes InSQuaD, a two-component approach to in-context learning (ICL) example selection. InSQuaD-RETRIEVE models exemplar annotation and retrieval as a targeted selection problem using submodular mutual information (SMI), aiming to balance quality and diversity. InSQuaD-LEARN trains a sentence embedding retriever (SBERT/MPNET) using a likelihood-based loss derived from submodular point processes, with a quality term L_q and a diversity term L_d supervised by GPT-3.5-generated paraphrases of HotpotQA documents. The trained retriever is then used in the selection pipeline. Experiments on nine ICL benchmarks with Gemma 2B report improvements over several baselines, with InSQuaD-GC the best variant; ablations examine retrieval method, annotation budget, the λ trade-off, and model scale.

Significance. If the results are reproducible, the paper offers a useful unified submodular framework for ICL, combining a principled selection objective with a retriever trained to align with that objective. The authors provide code and report three trials, and the NT vs full comparison is a reasonable first step toward isolating the training contribution. However, the lack of retrieval-level evaluation, unvalidated paraphrase supervision, a fixed λ that is suboptimal for most tasks, and the omission of a key end-to-end submodular baseline mean the central claims are not yet fully supported. The empirical gains, while positive, are modest in average terms and often within one standard deviation of baselines.

major comments (5)
  1. [III-C2, Table II] The paper attributes downstream ICL gains to InSQuaD-LEARN, but the only evidence is end-to-end accuracy. No retrieval-level metrics (e.g., recall, diversity of selected sets) are reported on the nine downstream tasks. The NT (no-training) variants already beat or tie IDEAL on several datasets (e.g., InSQuaD-GC NT: MRPC 0.57 vs 0.47, GeoQ 0.84 vs 0.82), so the marginal benefit of the learned retriever is not quantified. Please report retrieval quality/diversity metrics or an explicit delta analysis between NT and full models to support the LEARN contribution.
  2. [III-C2, Eq. (7)] L_d is defined entirely over GPT-3.5 paraphrases of HotpotQA documents. The paper provides no validation of paraphrase quality or evidence that this diversity signal transfers to the nine ICL benchmarks. The λ ablation (Table IV) shows λ>0 helps on 7/9 tasks, but this does not isolate the paraphrase signal from a generic auxiliary loss. Please include paraphrase quality checks (e.g., human/embedding-based) or an ablation with an alternative diversity supervision.
  3. [V, Table IV] Table II uses a fixed λ=0.5, yet Table IV shows that for InSQuaD-GC the optimal λ is 0.5 only for HellaSwag; for most tasks λ=0 or λ=1 is better. The choice of λ=0.5 is therefore not justified and may understate or distort the comparison. Either perform per-task λ selection on a validation split and report the selected values, or provide a principled reason for a single λ across tasks.
  4. [IV-C, Table II] The experimental comparison omits the end-to-end submodular ICL framework of Kumari et al. (2024, ref. [4]), which is cited in the Introduction as related work. MFL and GC baselines may be components of that framework, but not the full end-to-end trained method. A direct comparison is necessary to support the claim that InSQuaD improves on the state of the art in submodular ICL.
  5. [IV-B, Table II] The paper reports three trials with standard deviations but no significance tests. Several headline differences are within one standard deviation (e.g., InSQuaD-GC vs IDEAL on SST5: 0.43±0.05 vs 0.42±0.01; GeoQ: 0.85±0.03 vs 0.82±0.07). The abstract and contributions use the word 'significant'; please add paired bootstrap or permutation tests to support this claim.
minor comments (6)
  1. [Title/Abstract] Inconsistent spelling: title uses 'InSQuAD', body uses 'InSQuaD'. Also typo 'exisitng' in the abstract.
  2. [IV-A] Typo 'benachmarks' in the datasets paragraph.
  3. [III-C2] Duplicate phrase 'modeling the modeling' in the first paragraph of InSQuaD-LEARN.
  4. [Fig. 3 caption] The caption enumerates (a), (b), and (d) but no (c); also λ=0.5 is referenced but not defined in the main text.
  5. [IV-C] The MFL baseline acronym is not defined; please define it where it first appears.
  6. [Table I] The InSQuaD-LD expressions are dense; defining the submatrix notation (e.g., S_{S−, q}) would improve readability.

Circularity Check

0 steps flagged

No significant circularity: InSQuaD's central claim is an external empirical comparison, not a quantity defined by its fitted parameters.

full rationale

The derivation chain is self-contained in the relevant sense. InSQuaD-LEARN's objective L_InSQuaD = exp((1-lambda)L_q + lambda L_d) is a contrastive loss over SMI scores (Eqs. 5-7), trained on HotpotQA augmented with GPT-3.5 paraphrases. The test-time selection (Eq. 3) uses the same SMI family with the trained embeddings. The final claim is downstream ICL accuracy on nine external benchmarks (Table II), which is not equal to any training objective or selection score by construction. The self-citations to SPP [46] and SCORE [47] supply the mathematical form of the SMI/SPP likelihood and kernel similarity, but they do not assert the empirical result, and the paper does not invoke a uniqueness theorem to rule out alternatives. The weak spot—unvalidated paraphrase transfer from HotpotQA—is a data-distribution/experimental-support issue, not a definitional equivalence: even if the transfer were poor, the reported numbers would not be forced by the equations. NT comparisons indicate that some gains may come from RETRIEVE alone, but that is an attribution question, not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The empirical claim is evaluated on external benchmarks, so the parameter count is modest. The main free parameters are lambda, the submodular function choice, and the unreported eta; the approach relies on standard submodular approximation bounds and the assumption that embedding similarity captures quality and diversity.

free parameters (4)
  • lambda = 0.5
    Trade-off between quality L_q and diversity L_d; Table IV shows the optimal value varies by task, but 0.5 is used for the main results.
  • submodular function f = Graph-Cut (GC) chosen as best
    Three instantiations FL, GC, LD; the paper highlights GC as the best and uses it for the headline results.
  • eta = Not reported
    Appears in the FL and LD loss formulas in Table I; the paper does not state its value or tuning procedure.
  • annotation budget B = 18 (main), 100 (ablation)
    Set to match baseline papers [5,6]; ablation shows weak effect on performance.
axioms (4)
  • standard math Greedy submodular maximization yields a (1-1/e) approximation
    Invoked in Section III-B and used in Algorithm 1 for all SMI maximizations.
  • domain assumption Submodular Point Process likelihood from [46] is a valid probabilistic model of set relevance
    Used to define the likelihood ratio in InSQuaD-LEARN, Section III-C2.
  • domain assumption Cosine similarity on MPNet embeddings reflects semantic relevance and diversity
    The similarity kernel S_ab is the core input to all SMI computations; the method assumes this kernel encodes the needed notions.
  • domain assumption Quality and diversity of exemplars improve in-context learning
    Borrowed from prior works [2,6], this motivates the entire selection and training framework.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of InSQuAD: In-Context Learning for Efficient Retrieval via Submodular Mutual Information to Enforce Quality and Diversity." pith.science (2026). https://pith.science/paper/DTLARONO

@misc{pith2026250821003,
  author       = {Pith},
  title        = {Pith review of: InSQuAD: In-Context Learning for Efficient Retrieval via Submodular Mutual Information to Enforce Quality and Diversity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DTLARONO}},
  note         = {Machine review of arXiv:2508.21003}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In this paper, we introduce InSQuAD, designed to enhance the performance of In-Context Learning (ICL) models through Submodular Mutual Information} (SMI) enforcing Quality and Diversity among in-context exemplars. InSQuAD achieves this through two principal strategies: First, we model the ICL task as a targeted selection problem and introduce a unified selection strategy based on SMIs which mines relevant yet diverse in-context examples encapsulating the notions of quality and diversity. Secondly, we address a common pitfall in existing retrieval models which model query relevance, often overlooking diversity, critical for ICL. InSQuAD introduces a combinatorial training paradigm which learns the parameters of an SMI function to enforce both quality and diversity in the retrieval model through a novel likelihood-based loss. To further aid the learning process we augment an existing multi-hop question answering dataset with synthetically generated paraphrases. Adopting the retrieval model trained using this strategy alongside the novel targeted selection formulation for ICL on nine benchmark datasets shows significant improvements validating the efficacy of our approach.

Figures

Figures reproduced from arXiv: 2508.21003 by Anay Majee, Rishabh Iyer, Souradeep Nanda.

Figure 1
Figure 1. Figure 1: Overview of InSQuaD comprising two principle components - (a) InSQuaD-RETRIEVE which models exemplar annotation and retrieval as a targeted selection problem injecting the notions of quality, diversity and order. (b) InSQuaD-LEARN which introduces a novel learning strategy to enforce the notions of quality and diversity into the retrieval model R. the goal of the downstream ICL tasks which is represented a… view at source ↗
Figure 2
Figure 2. Figure 2: Training Workflow of InSQuaD-LEARN which learns the parameters of an SMI function to update the parameters of a retrieval model R through a novel likelihood based objective LInSQuaD enforcing quality and diversity in learnt embeddings. C ← argmax C⊆Vlabeled , |C|≤k If (C; qtest) (3) The greedy optimization [13] strategy detailed in Alg.1, adopted during the selection process orders the exemplars in C based… view at source ↗
Figure 3
Figure 3. Figure 3: Ablations in InSQuaD-RETRIEVE which studies the effect of (a) Model size of the underlying LLM M, (b) the choice of retrieval method between {Random, Similar, and Combinatorial} with M fixed to gemma2b and (d) the effect of annotation budget B with M as gemma2b and λ = 0.5. hop question answering datasets like MuSiQue [56] (207 citations), ConcurrentQA [57] (22 citations). We will release the augmented dat… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of inference time (in Log scale) across different methods. Adopting a combinatorial approach significantly reduces inference times making them a suitable choice for ICL tasks. The model R is trained for 7 epochs. The learning rate is set at 3 × 10−5 , a value chosen to strike an optimal balance between rapid convergence and stability in training. We adopt a weight decay of 0.01 to mitigate overf… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

71 extracted references · 48 canonical work pages · 6 internal anchors

  1. [1]

    Active example selection for in-context learning,

    Y Zhang, S Feng, and C Tan, “Active example selection for in-context learning,” arXiv, vol. abs/2211.04486, 2022

    Pith/arXiv arXiv 2022

  2. [2]

    Self-adaptive in- context learning: An information compression perspec- tive for in-context example selection and ordering,

    Z Wu, Y Wang, J Ye, and L Kong, “Self-adaptive in- context learning: An information compression perspec- tive for in-context example selection and ordering,” in ACL, 2022

    work page 2022

  3. [3]

    Finding support examples for in- context learning,

    X Li and X Qiu, “Finding support examples for in- context learning,” inEMNLP, 2023

    work page 2023

  4. [4]

    An end-to-end submodular framework for data-efficient in-context learning,

    L Kumari, S Wang, AM Das, T Zhou, and J Bilmes, “An end-to-end submodular framework for data-efficient in-context learning,” inNAACL-HLT, 2024

    work page 2024

  5. [5]

    IDEAL: Influence-Driven Selective Annotations Empower In-Context Learners in Large Language Models

    S Zhanget al., “Ideal: Influence-driven selective anno- tations empower in-context learners in large language models,” arXiv, vol. abs/2310.10873, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  6. [6]

    Selective annotation makes language models better few-shot learners,

    H Suetal., “Selective annotation makes language models better few-shot learners,”arXiv, vol. abs/2209.01975, 2022

    Pith/arXiv arXiv 2022

  7. [7]

    Which examples to annotate for in-context learning? towards effective and efficient selection,

    C Mavromatis et al., “Which examples to annotate for in-context learning? towards effective and efficient selection,” arXiv, vol. abs/2310.20046, 2023

    Pith/arXiv arXiv 2023

  8. [8]

    Compositional exemplars for in-context learning,

    J Ye, Z Wu, J Feng, T Yu, and L Kong, “Compositional exemplars for in-context learning,” inICML, 2023

    work page 2023

  9. [9]

    Learning to retrieve prompts for in-context learning,

    O Rubin, J Herzig, and J Berant, “Learning to retrieve prompts for in-context learning,” arXiv, vol. abs/2112.08633, 2021

    Pith/arXiv arXiv 2021

  10. [10]

    Generalized submodular information measures: Theo- retical properties, examples, optimization algorithms, and applications,

    RK Iyer, N Khargonkar, JA Bilmes, and H Asnani, “Generalized submodular information measures: Theo- retical properties, examples, optimization algorithms, and applications,” IEEE Trans. Inf. Theory, vol. 68, pp. 752–781, 2022

    work page 2022

  11. [11]

    Prism: A rich class of parameterized submodular information measures for guided data subset selection,

    S Kothawade, V Kaushal, G Ramakrishnan, JA Bilmes, and RK Iyer, “Prism: A rich class of parameterized submodular information measures for guided data subset selection,” inAAAI, 2021

    work page 2021

  12. [12]

    A sequential algorithm for training text classifiers,

    DD Lewis and WA Gale, “A sequential algorithm for training text classifiers,” inSIGIR, 1994

    work page 1994

  13. [13]

    An analysis of approximations for maximizing submodular set functions—i,

    GL Nemhauser, LA Wolsey, and ML Fisher, “An analysis of approximations for maximizing submodular set functions—i,”Math. Program., vol. 14, pp. 265–294, 1978

    work page 1978

  14. [14]

    Sentence-bert: Sentence embeddings using siamese bert-networks,

    N Reimers and I Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,” inEMNLP, 2019

    work page 2019

  15. [15]

    Submodularity beyond submodular energies: Coupling edges in graph cuts,

    S Jegelka and JA Bilmes, “Submodularity beyond submodular energies: Coupling edges in graph cuts,” CVPR, pp. 1897–1904, 2011

    work page 1904

  16. [16]

    A class of submodular functions for document summarization,

    HC Lin and JA Bilmes, “A class of submodular functions for document summarization,” inACL, 2011

    work page 2011

  17. [17]

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

    Z Yanget al., “Hotpotqa: A dataset for diverse, explain- able multi-hop question answering,” inEMNLP, 2018

    work page 2018

  18. [18]

    Active prompting with chain-of-thought for large language models,

    S Diao, P Wang, Y Lin, X Liu, and T Zhang, “Active prompting with chain-of-thought for large language models,” inACL, 2023

    work page 2023

  19. [19]

    Rethinking the role of demonstra- tions: What makes in-context learning work?

    S Min et al., “Rethinking the role of demonstra- tions: What makes in-context learning work?”arXiv, vol. abs/2202.12837, 2022

    Pith/arXiv arXiv 2022

  20. [20]

    Ground-truth labels matter: A deeper look into input-label demonstrations,

    J Kim et al., “Ground-truth labels matter: A deeper look into input-label demonstrations,” arXiv, vol. abs/2205.12685, 2022

    Pith/arXiv arXiv 2022

  21. [21]

    An explanation of in-context learning as implicit bayesian inference,

    SM Xie, A Raghunathan, P Liang, and T Ma, “An explanation of in-context learning as implicit bayesian inference,” arXiv, vol. abs/2111.02080, 2021

    Pith/arXiv arXiv 2021

  22. [22]

    A theory of emergent in- context learning as implicit structure induction,

    M Hahn and N Goyal, “A theory of emergent in- context learning as implicit structure induction,”arXiv, vol. abs/2303.07971, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Unified demonstration retriever for in- context learning,

    X Li et al., “Unified demonstration retriever for in- context learning,”arXiv, vol. abs/2305.04320, 2023

    Pith/arXiv arXiv 2023

  24. [24]

    Understanding in-context learning via supportive pretraining data,

    X Han, D Simig, T Mihaylov, Y Tsvetkov, A Celikyilmaz, and T Wang, “Understanding in-context learning via supportive pretraining data,” inACL, 2023

    work page 2023

  25. [25]

    Rethinking the Role of Scale for In-Context Learning: An Interpretability-based Case Study at 66 Billion Scale

    H Bansal, K Gopalakrishnan, S Dingliwal, S Bodapati, K Kirchhoff, and D Roth, “Rethinking the role of scale for in-context learning: An interpretability-based case study at 66 billion scale,”arXiv, vol. abs/2212.09095, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  26. [26]

    Larger language models do in-context learning differently,

    JW Weiet al., “Larger language models do in-context learning differently,”arXiv, vol. abs/2303.03846, 2023

    Pith/arXiv arXiv 2023

  27. [27]

    In-contextlearningandinductionheads,

    COlsson etal.,“In-contextlearningandinductionheads,” arXiv, vol. abs/2209.11895, 2022

    Pith/arXiv arXiv 2022

  28. [28]

    Active learning for bert: An empirical study,

    L Ein-Doret al., “Active learning for bert: An empirical study,” inEMNLP, 2020

    work page 2020

  29. [29]

    Revisiting uncertainty-based query strategies for active learning with transformers,

    C Schröder, A Niekler, and M Potthast, “Revisiting uncertainty-based query strategies for active learning with transformers,” inFindings, 2021

    work page 2021

  30. [30]

    Attention is all you need,

    A Vaswaniet al., “Attention is all you need,” inNeurIPS, 2017

    work page 2017

  31. [31]

    Transformers learn in-context by gradient descent,

    J von Oswaldet al., “Transformers learn in-context by gradient descent,” inICML, 2022

    work page 2022

  32. [32]

    Trained transformers learn linear models in-context,

    R Zhang, S Frei, and PL Bartlett, “Trained transformers learn linear models in-context,”J. Mach. Learn. Res., vol. 25, 49:1–49:55, 2023

    work page 2023

  33. [33]

    Don’t make your llm an evaluation benchmark cheater,

    K Zhou et al., “Don’t make your llm an evaluation benchmark cheater,”arXiv, vol. abs/2311.01964, 2023

    Pith/arXiv arXiv 2023

  34. [34]

    On the opportunities and risks of foundation models,

    R Bommasaniet al., “On the opportunities and risks of foundation models,”arXiv, vol. abs/2108.07258, 2021

    Pith/arXiv arXiv 2021

  35. [35]

    A survey of large language models,

    WX Zhaoet al., “A survey of large language models,” arXiv, vol. abs/2303.18223, 2023

    Pith/arXiv arXiv 2023

  36. [36]

    A survey on in-context learning,

    Q Dong et al., “A survey on in-context learning,” in EMNLP, 2022

    work page 2022

  37. [37]

    Emergent abilities of large language models,

    J Wei et al., “Emergent abilities of large language models,” arXiv, vol. abs/2206.07682, 2022

    Pith/arXiv arXiv 2022

  38. [38]

    Data distributional properties drive emergent in-context learning in transformers,

    SCY Chanet al., “Data distributional properties drive emergent in-context learning in transformers,”arXiv, vol. abs/2205.05055, 2022

    Pith/arXiv arXiv 2022

  39. [39]

    Submodular functions and optimization,

    S Fujishige, “Submodular functions and optimization,” 1991

    work page 1991

  40. [40]

    AUTOMATA: Gradient Based Data Subset Selection for Compute-Efficient Hyper-parameter Tuning

    K Killamsetty et al., “Automata: Gradient based data subset selection for compute-efficient hyper-parameter tuning,” arXiv, vol. abs/2203.08212, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  41. [41]

    Talisman: Targeted active learning for object detection with rare classes and slices using submodular mutual information,

    S Kothawade, S Ghosh, S Shekhar, Y Xiang, and RK Iyer, “Talisman: Targeted active learning for object detection with rare classes and slices using submodular mutual information,” inECCV, 2021

    work page 2021

  42. [42]

    Efficient Data Subset Selection to Generalize Training Across Models: Transductive and Inductive Networks

    E Jain, T Nandy, G Aggarwal, A Tendulkar, RK Iyer, and A De, “Efficient data subset selection to generalize training across models: Transductive and inductive networks,” arXiv, vol. abs/2409.12255, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  43. [43]

    Demystifying multi-faceted video summarization: Tradeoff between diversity, representa- tion, coverage and importance,

    V Kaushal et al., “Demystifying multi-faceted video summarization: Tradeoff between diversity, representa- tion, coverage and importance,”WACV, pp. 452–461, 2019

    work page 2019

  44. [44]

    How good is a video summary? a new benchmarking dataset and evaluation frame- work towards realistic video summarization,

    V Kaushal, S Kothawade, A Tomar, RK Iyer, and G Ramakrishnan, “How good is a video summary? a new benchmarking dataset and evaluation frame- work towards realistic video summarization,” arXiv, vol. abs/2101.10514, 2021

    Pith/arXiv arXiv 2021

  45. [45]

    Lazier than lazy greedy,

    B Mirzasoleiman, A Badanidiyuru, A Karbasi, J Von- drák, and A Krause, “Lazier than lazy greedy,”arXiv, vol. abs/1409.7938, 2014

    Pith/arXiv arXiv 2014

  46. [46]

    Submodular point processes with applications to machine learning,

    RK Iyer and JA Bilmes, “Submodular point processes with applications to machine learning,” inAISTATS, 2015

    work page 2015

  47. [47]

    Score: Submodular combinatorial representation learn- ing,

    A Majee, S Kothawade, K Killamsetty, and RK Iyer, “Score: Submodular combinatorial representation learn- ing,” inICML, 2023

    work page 2023

  48. [48]

    SMILe: Leveraging Submodular Mutual Information For Robust Few-Shot Object Detection

    A Majee, R Sharp, and R Iyer, “Smile: Leveraging submodular mutual information for robust few-shot object detection,”arXiv, vol. abs/2407.02665, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  49. [49]

    Submodular combinatorial information measures with applications in machine learning,

    RK Iyer, N Khargoankar, JA Bilmes, and H Asanani, “Submodular combinatorial information measures with applications in machine learning,” inALT, 2020

    work page 2020

  50. [50]

    Learning mixtures of submod- ular shells with application to document summarization,

    HC Lin and JA Bilmes, “Learning mixtures of submod- ular shells with application to document summarization,” in UAI, 2012

    work page 2012

  51. [51]

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

    Y Lu, M Bartolo, A Moore, S Riedel, and P Stenetorp, “Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity,” inACL, 2021

    work page 2021

  52. [52]

    Squad: 100,000+ questions for machine comprehension of text,

    P Rajpurkar, J Zhang, K Lopyrev, and P Liang, “Squad: 100,000+ questions for machine comprehension of text,” in EMNLP, 2016

    work page 2016

  53. [53]

    Training language models to follow instructions with human feedback,

    L Ouyanget al., “Training language models to follow instructions with human feedback,”NeurIPS, 2022

    work page 2022

  54. [54]

    Questions are all you need to train a dense passage retriever,

    DS Sachan, M Lewis, D Yogatama, L Zettlemoyer, J Pineau, and M Zaheer, “Questions are all you need to train a dense passage retriever,”Trans. ACL, vol. 11, pp. 600–616, 2022

    work page 2022

  55. [55]

    Ask to Understand: Question Generation for Multi-hop Question Answering

    J Li, M Ren, Y Gao, and Y Yang, “Ask to understand: Question generation for multi-hop question answering,” arXiv, vol. abs/2203.09073, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  56. [56]

    MuSiQue: Multihop questions via single-hop question composition,

    H Trivedi, N Balasubramanian, T Khot, and A Sab- harwal, “MuSiQue: Multihop questions via single-hop question composition,”Trans. ACL, 2022

    work page 2022

  57. [57]

    Reasoning over public and private data in retrieval-based systems,

    S Arora, P Lewis, A Fan, J Kahn, and C Ré, “Reasoning over public and private data in retrieval-based systems,” Trans. ACL, 2023

    work page 2023

  58. [58]

    Exploiting se- mantic annotations and q-learning for constructing an ef- ficient hierarchy/graph texts organization,

    AM el-Said, AI Eldesoky, and HA Arafat, “Exploiting se- mantic annotations and q-learning for constructing an ef- ficient hierarchy/graph texts organization,”Sci. World J., vol. 2015, 2015

    work page 2015

  59. [59]

    The pas- cal recognising textual entailment challenge,

    I Dagan, O Glickman, and B Magnini, “The pas- cal recognising textual entailment challenge,” in ML Challenges Workshop, 2005

    work page 2005

  60. [60]

    A broad- coverage challenge corpus for sentence understanding through inference,

    A Williams, N Nangia, and SR Bowman, “A broad- coverage challenge corpus for sentence understanding through inference,” inNAACL, 2017

    work page 2017

  61. [61]

    Recursive deep models for semantic compositionality over a sentiment treebank,

    R Socheret al., “Recursive deep models for semantic compositionality over a sentiment treebank,” inEMNLP, 2013

    work page 2013

  62. [62]

    Dbpedia - a large-scale, mul- tilingual knowledge base extracted from wikipedia,

    J Lehmann et al., “Dbpedia - a large-scale, mul- tilingual knowledge base extracted from wikipedia,” Semant. Web, vol. 6, pp. 167–195, 2015

    work page 2015

  63. [63]

    Hellaswag: Can a machine really finish your sentence?

    R Zellers, A Holtzman, Y Bisk, A Farhadi, and Y Choi, “Hellaswag: Can a machine really finish your sentence?” In ACL, 2019

    work page 2019

  64. [64]

    Multiwoz - a large-scale multi- domain wizard-of-oz dataset for task-oriented dialogue modelling,

    P Budzianowskiet al., “Multiwoz - a large-scale multi- domain wizard-of-oz dataset for task-oriented dialogue modelling,” inEMNLP, 2018

    work page 2018

  65. [65]

    Learning to parse database queries using inductive logic programming,

    JM Zelle and RJ Mooney, “Learning to parse database queries using inductive logic programming,” inAAAI, 1996

    work page 1996

  66. [66]

    Don’t give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization,

    S Narayan, SB Cohen, and M Lapata, “Don’t give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization,” arXiv, vol. abs/1808.08745, 2018

    Pith/arXiv arXiv 2018

  67. [67]

    Metaicl: Learning to learn in context,

    S Min, M Lewis, L Zettlemoyer, and H Ha- jishirzi, “Metaicl: Learning to learn in context,”arXiv, vol. abs/2110.15943, 2021

    Pith/arXiv arXiv 2021

  68. [68]

    Gemma: Open models based on gemini research and technology,

    GTT Mesnard et al., “Gemma: Open models based on gemini research and technology,” arXiv, vol. abs/2403.08295, 2024

    Pith/arXiv arXiv 2024

  69. [69]

    Language models are few-shot learners,

    TB Brown et al., “Language models are few-shot learners,” arXiv, vol. abs/2005.14165, 2020

    Pith/arXiv arXiv 2005

  70. [70]

    Decoupled weight decay regularization,

    I Loshchilov and F Hutter, “Decoupled weight decay regularization,” inICLR, 2017

    work page 2017

  71. [71]

    Mpnet: Masked and permuted pre-training for language under- standing,

    K Song, X Tan, T Qin, J Lu, and TY Liu, “Mpnet: Masked and permuted pre-training for language under- standing,” arXiv, vol. abs/2004.09297, 2020

    Pith/arXiv arXiv 2004

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.