Pith. sign in

REVIEW 5 major objections 6 minor 72 references

This paper claims that the optimal data-tuning-inference recipe for an extractive LLM can be predicted by a quadratic function fit to F1 scores from only 10 of 28 candidate combinations, and demonstrates this on a relation-extraction datase

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 ↗

T0 review · deepseek-v4-flash

2026-08-02 13:18 UTC pith:2AIMFV52

load-bearing objection A plausible AutoML-style planner for extractive LLM configuration with one genuinely interesting held-out result, but the generalization claim rests on a single scenario and the paper's own ablation undercuts it. the 5 major comments →

arxiv 2607.20470 v1 pith:2AIMFV52 submitted 2026-05-22 cs.AI

PlanE: Meta Planning of Data, Tuning, and Inference for Extractive-based LLMs

Jiacheng Wang , Weiyan Zhang , Guangya Yu This is my paper
classification cs.AI
keywords extractive LLMsdata decompositioninstruction tuningprompt inferenceDTI plannermeta-planningquadratic surrogateinformation extraction
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 paper is trying to establish that building a task-specific extractive LLM can be planned rather than brute-forced: a meta-planner learns which combination of data decomposition, tuning method, and inference prompt works best for a given base model and dataset. The proposed PlanE framework defines 28 candidate DTI combinations and a 'DTI planner' — a quadratic score function over three scalar encodings of data/tuning/inference choices. The claim is that fitting this function on 10 measured combinations from one set of scenarios (other models on the same dataset, or other datasets on the same model) suffices to identify the best of all 28 for a new scenario. In the main validation, the planner selects pipeline-based decomposition, SFT+KTO tuning, and intersection inference for an 8-billion-parameter model on a medical relation-extraction dataset, reaching 54.30 F1 — the same best as exhaustive grid search, at a fraction of the search cost. The paper also extends the planner to a performance-plus-efficiency objective.

Core claim

The central discovery is that a single quadratic form in three categorical variables — data decomposition, tuning strategy, and inference strategy — can act as a learned ranking model over DTI combinations. The variables vd, vt, vi are derived from F1 scores of 10 representative combinations, mapped to integers 1–10. Coefficients are solved by gradient descent on a ReLU constraint loss that enforces the observed best combination to outrank all others in the training set. On a held-out scenario (a new base model or a new dataset), the planner evaluates all 28 combinations with the fitted quadratic and picks the argmax. In the main experiment this picks pipeline-based decomposition, SFT+KTO tu

What carries the argument

The DTI planner: a quadratic polynomial y(co) = a·vd² + b·vt² + c·vi² + d·vd·vt + e·vd·vi + f·vt·vi + g·vd + h·vt + i·vi, where vd, vt, vi are 1–10 integer encodings of data decomposition, tuning strategy, and inference strategy, computed from a handful of F1 measurements. The coefficients are fit by minimizing a ReLU constraint loss that requires the known-best combination to score higher than every other combination in the training set; once fit, the planner ranks all 28 combinations for a new model or dataset.

Load-bearing premise

That a single quadratic function of three scalar variables, fit on F1 scores from 10 representative combinations observed on other models or datasets, transfers well enough to correctly rank the 18 unmeasured combinations for a new model or dataset — an assumption the paper's own ablation (Section 5.5) undercuts by finding no patterned effect of DTI factors across datasets.

What would settle it

Measure all 28 DTI combinations on a held-out 8-billion-parameter model and dataset, fit the planner on the other three models (or the other two datasets), and check whether the planner's argmax achieves the true maximum F1. If any unmeasured combination outscores the planner's pick, the quadratic transfer fails. A cheaper check: fit the planner on 10 randomly chosen combinations of the same data and see whether the argmax still matches the grid-search optimum.

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

If this is right

  • Constructing a task-specific extractive LLM may no longer require running all candidate pipelines; a planner fit on a few measurements can predict the best combination.
  • The same planner shape supports multi-objective optimization, trading F1 against construction time via min-max normalized weights.
  • The result suggests that the interaction between data structure, tuning algorithm, and inference prompt over a fixed base LLM has a smooth, low-dimensional structure that can be learned and transferred.
  • Because the planner identifies the same optimum even when the variable mapping range is changed from 1–10 to 1–20 and when coefficients are re-solved non-uniquely, the ranking is claimed to be robust to these implementation details.

Where Pith is reading between the lines

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

  • The transfer claim is only as strong as the fitted scenarios: the paper's own ablation finds no consistent pattern of DTI effects across datasets, so the planner's success may depend on the specific family of 8-billion-parameter models used for fitting.
  • If the quadratic surrogate holds broadly, it suggests that the space of LLM-construction choices is low-rank enough for systematic design-of-experiments approaches — a handful of runs could guide which decompositions to invest in, rather than grid search.
  • A testable extension: fit the planner on 10 randomly chosen combinations instead of the specific 10 representative ones; if the optimum is still recovered, the result would be stronger evidence for the smoothness assumption.
  • The planner approach might extend to other categorical choices in LLM pipelines, such as prompting strategy families or decoding hyperparameters, as long as they can be scalar-encoded.

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 PlanE, a framework for constructing task-specific extractive LLMs by jointly planning data decomposition (none/pipeline/bidirectional), instruction tuning (SFT, SFT+GRPO/DPO/KTO), and inference (direct/intersection/union). A DTI planner is introduced as a quadratic scoring function over three scalar variables vd, vt, vi, with coefficients solved from inequality constraints on historical data. The paper reports two generalization views: fitting on three base LLMs and evaluating on a fourth (Qwen3-8B on CMeIE-V2), and fitting on two datasets and evaluating on a third (Qwen3-8B on CMeIE-V2). In both views the planner selects pipeline-based decomposition + SFT+KTO + intersection inference (ID=15), which is the best of 28 measured combinations. The paper also reports multi-objective optimization and an ablation over data/tuning/inference.

Significance. If the transfer claim held, PlanE would be a useful practical tool: it would reduce the cost of finding a good DTI configuration on a new model/dataset to ten measured combinations rather than a full grid, with code and datasets released. The two-view experimental design is a reasonable way to frame cross-model and cross-dataset transfer, and the fact that the selected combination is not one of the ten measured anchor combinations is a nontrivial result. The main value is therefore empirical: a single quadratic transfer function over constructed scalar variables can rank unseen combinations. However, the evidence for this claim is currently thin, and several internal results weaken it. The paper explicitly acknowledges non-unique coefficient solutions (Appendix F) and reports that DTI effects show no clear pattern across datasets (Section 5.5), which makes the one successful held-out case in each view insufficient to establish the advertised generalization.

major comments (5)
  1. [§5.2, §5.3, §5.5, Table 7] The central generalization claim rests on exactly one held-out model (View 1) and one held-out dataset (View 2). No leave-one-out evaluation is reported, even though the existing data permit LOOCV over the four models in View 1 and the three datasets in View 2. Section 5.5 explicitly states that DTI effects across datasets 'shows no pattern,' which is in tension with a fixed quadratic transfer function. Please report LOOCV ranks for all available folds, not just the favorable fold, and report at least the rank of the true optimum or top-3 coverage for each held-out case. Without this, the single ID=15 success could be a favorable selection from an underdetermined coefficient set.
  2. [§4.2, Eqs. (4)–(6), Appendix F, Table 5] The 9-parameter quadratic is fit only through inequality constraints with epsilon=1e-10, so the coefficient solution is massively underdetermined; Appendix F concedes non-uniqueness. This is not merely a theoretical concern: the score margins on the held-out target are tiny. In Table 5, f(x)dataset gives ID15=229, ID12=220, ID6=227, and the coefficient variant in Table A7 gives ID15=42, ID6=40, ID12=33. Thus the selected optimum is a knife-edge result; an unreported initialization or a different valid coefficient vector could plausibly change the ranking. The Appendix F demonstration that two arbitrary solutions both select ID15 does not establish robustness. Please add a principled selection criterion (e.g., regularize the coefficient norm, maximize margin, or use a validation fold) and report sensitivity to initialization and to the choice of solution among the feasible set.
  3. [Table 4, Table 5] The reported search-time savings do not reconcile with the measured run times. Table 4 lists PlanE's search time as 187,257 s, but the ten representative combinations whose F1s define the variables (IDs 1–7 and 17–19 in Table 5) sum to about 461,000 s of construction time. Likewise, the Grid Search search time of 742,090 s is far below the sum of all 28 run times in Table 5 (about 2,162,000 s). Please define precisely what 'search time' measures (e.g., optimization CPU seconds, wall-clock excluding training, or an estimated cost model) and reconcile the numbers. As written, the headline claim of 'reducing search time by 554,833 seconds' is not verifiable from the paper's own tables.
  4. [§5.2, Table 5] No variance or seed information is reported for any F1 value. The decisive comparison is ID15 at 54.30 F1 versus ID12 at 53.28 F1 and ID6 at 52.70 F1, a margin of about 1.0 point against the closest alternative. Given the known sensitivity of instruction tuning and decoding to random seeds, and the fact that the planner's own score margins are within 1–5 units on an arbitrary scale, the paper should report at least three independent runs (or confidence intervals) for the top configurations. Without this, the 'best of 28' claim is not statistically supported.
  5. [§4.3, Appendix C] The variable mapping in Appendix C uses F1 values measured on the target scenario itself (e.g., vd is computed from None/Pipeline/Bidirectional SFT-direct F1 on the target dataset/model). This is legitimate as a cost-saving protocol, but it means the planner is a post hoc scoring function over partially observed target performance, not a prediction from observable dataset/model metadata. The paper should state this limitation explicitly and discuss what happens when the ten anchor F1s are noisy or when only cheaper proxies (e.g., small-sample estimates) are available. This also affects the claimed reduction in search time: the protocol still requires training and evaluating ten full DTI combinations on the target.
minor comments (6)
  1. [§3.1, Definition 2] The label '(RE)' under Definition 2 should be '(EE)' for event extraction.
  2. [§4.2, Eq. (4)] The term list 'vd, vt, and vt' should read 'vd, vt, and vi'.
  3. [Table 4, Table 6] The column 'Sea.' is used without a definition in the table caption; define it. Some F1 entries in Table 6 appear to have stray digits (e.g., '52.704', '53.282').
  4. [Appendix F] The phrase 'we can still stably obtain the optimal DTI combination' overstates what two arbitrary feasible solutions show; suggest replacing with a statement about sensitivity and feasible-set size.
  5. [Appendix H, Eq. (8)] The symbol Ai in the GRPO objective is used but not defined in the text; define it as the group-relative advantage for output i.
  6. [§6, Conclusion] The claim that PlanE is 'the first to focus on extractive-based LLM construction' is a strong novelty claim and may be hard to verify; consider softening.

Circularity Check

0 steps flagged

No significant circularity: DTI coefficients are fit on held-out scenarios, and the selected optimum ID=15 is not among the 10 measured input combinations.

full rationale

The claimed derivation chain is not circular. The DTI planner's coefficients are learned on source scenarios only: f(x)dataset uses CMeIE-V2 F1 results from GLM4-9B, LLaMA3.1-8B, and InternLM3-8B (Appendix D), while f(x)model uses Qwen3-8B results on ACE05 and 14Lap (Appendix E). The target features vd, vt, vi are F1-derived scalar encodings computed from 10 representative combinations on the target (Appendix C, Table A3). The selected optimum, ID=15 (Pipeline/SFT+KTO/Intersection), is not among those 10 measured combinations; its F1=54.30 enters only in the evaluation table. Therefore Eq. (7) argmax y is not Eq. (3) argmax F1 by construction, and the planner is not re-announcing a target-fitted parameter. The non-uniqueness of the coefficients is disclosed in Appendix F, and a second solution also yields ID=15, so no uniqueness theorem is smuggled in. The only self-citation (Zhang et al. 2024) is for standard evaluation metrics and is not load-bearing. The strongest concerns are evidential, not circular: one held-out scenario per view, no error bars, and Section 5.5 states DTI effects 'across datasets shows no pattern,' which undermines the transfer generality but does not make the derivation circular.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central contribution is a fitted surrogate, not a derivation: nine polynomial coefficients are optimized against observed F1 constraints from two or three training scenarios. The scalar variables themselves are F1-derived, so the planner's 'prediction' is a ranking over configurations informed by 10 measured F1 values. No new physical or architectural entities are introduced.

free parameters (3)
  • Quadratic planner coefficients a-i = e.g., Performance f(x)_dataset: a=-1.75, b=-3.49, c=-2.07, d=2.11, e=0.61, f=0.14, g=1.47, h=39.26, i=26.45; other objec
    Solved by gradient descent on inequality constraints (Eq. 6) from empirical F1 data; they are the central ranking mechanism and are not derived from first principles.
  • Variable mapping range = 1-10 default; 1-20 in robustness check
    Hand-chosen min-max normalization of F1 values onto integer variables; tested for robustness but arbitrary.
  • Multi-objective weights alpha and beta = alpha = beta = 0.5
    Default weighting of normalized F1 and time in the multi-objective score; no data-driven justification is given.
axioms (5)
  • domain assumption A quadratic polynomial in scalar encodings vd, vt, vi can model F1 ranking across DTI combinations and generalize across models/datasets.
    Core to the planner; stated in Sec. 4.2 Eq. 4. No theoretical justification is given, and Sec. 5.5 says DTI effects show no pattern across datasets.
  • domain assumption F1 scores of the 10 representative combinations determine all three variables and are sufficient to predict the best among all 28 combinations.
    Appendix C defines vd/vt/vi from these F1s, and Sec. 4.3 claims a subset of 10 is required.
  • domain assumption The 28 DTI combinations span the practically relevant policy space.
    The Limitations section explicitly notes that many more combinations involving other data, tuning, and inference strategies are not considered.
  • domain assumption Strict Micro-F1 with exact span/type matching is the appropriate evaluation objective.
    Evaluation metric definition in Sec. 5.1; standard for the three IE tasks but task-specific.
  • standard math Gradient descent finds a feasible coefficient solution for the inequality-constrained quadratic.
    Used to solve Eq. 6; the paper itself notes non-uniqueness of solutions in Appendix F.

pith-pipeline@v1.3.0-alltime-deepseek · 26736 in / 14487 out tokens · 154533 ms · 2026-08-02T13:18:21.854198+00:00 · methodology

0 comments
read the original abstract

Enhancing the task-specific capabilities of Large Language Models (LLMs) primarily requires substantial instruction-tuning datasets. However, the sheer volume of such data imposes a considerable annotation cost, and a lack of optimization methods for tailoring LLMs to specific tasks. To address the above issues, we propose a \textbf{Plan}ning framework for constructing \textbf{E}xtractive-based LLMs called \textbf{PlanE}, which includes data decomposition, instruction tuning, and prompt inference. Additionally, we introduce a Data-Tuning-Inference (DTI) planner, aimed at selecting the optimal base-LLM and its DTI combinations for specific datasets to improve construction efficiency. The experimental results demonstrate the effectiveness of our PlanE from two views: (1) across different datasets using the same base-LLM, and (2) on the same dataset using different base-LLMs. Furthermore, we validate the generalizability of the proposed DTI planner under different optimization objectives. The codes are publicly available at https://github.com/gugugu-469/PlanE.

Figures

Figures reproduced from arXiv: 2607.20470 by Guangya Yu, Jiacheng Wang, Weiyan Zhang.

Figure 1
Figure 1. Figure 1: (a) Three key factors in building task-specific [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The framework of PlanE. space, such as latency, accuracy, cost, or effi￾ciency. DeepSpeed-Inference balances latency and accuracy, while AdaCoT (Lou et al., 2025) uses Pareto search to navigate cost-performance trade￾offs. Meta-R1 (Dong et al., 2025) introduces root￾scaled efficiency for cross-scale efficiency normal￾ization, and Chinchilla (Hoffmann et al., 2022) pro￾vides a compute-optimal scaling law th… 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

72 extracted references · 4 canonical work pages

  1. [1]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Direct Preference Optimization: Your Language Model is Secretly a Reward Model , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  2. [2]

    arXiv preprint arXiv:2402.01306 , year=

    Kto: Model alignment as prospect theoretic optimization , author=. arXiv preprint arXiv:2402.01306 , year=

    Pith/arXiv arXiv

  3. [3]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    Seeking Common but Distinguishing Difference, A Joint Aspect-based Sentiment Analysis Model , author=. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    2021

  4. [4]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    Position-Aware Tagging for Aspect Sentiment Triplet Extraction , author=. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    2020

  5. [5]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    Pith/arXiv arXiv

  6. [6]

    Introducing Qwen1.5 , url =

    Qwen Team , month =. Introducing Qwen1.5 , url =

  7. [7]

    Qwen3: Think Deeper, Act Faster , url =

    Qwen Team , month =. Qwen3: Think Deeper, Act Faster , url =

  8. [8]

    arXiv preprint arXiv:2403.17297 , year=

    Internlm2 technical report , author=. arXiv preprint arXiv:2403.17297 , year=

    Pith/arXiv arXiv

  9. [9]

    ai , author=

    Yi: Open foundation models by 01. ai , author=. arXiv preprint arXiv:2403.04652 , year=

    Pith/arXiv arXiv

  10. [10]

    arXiv preprint arXiv:2401.02954 , year=

    Deepseek llm: Scaling open-source language models with longtermism , author=. arXiv preprint arXiv:2401.02954 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2405.04434 , year=

    Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model , author=. arXiv preprint arXiv:2405.04434 , year=

    Pith/arXiv arXiv

  12. [12]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    GLM: General Language Model Pretraining with Autoregressive Blank Infilling , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  13. [13]

    arXiv preprint arXiv:2406.12793 , year=

    Chatglm: A family of large language models from glm-130b to glm-4 all tools , author=. arXiv preprint arXiv:2406.12793 , year=

    Pith/arXiv arXiv

  14. [14]

    arXiv preprint arXiv:2411.07140 , year=

    Chinese simpleqa: A chinese factuality evaluation for large language models , author=. arXiv preprint arXiv:2411.07140 , year=

    Pith/arXiv arXiv

  15. [15]

    arXiv preprint arXiv:2411.04368 , year=

    Measuring short-form factuality in large language models , author=. arXiv preprint arXiv:2411.04368 , year=

    Pith/arXiv arXiv

  16. [16]

    a is b" fail to learn

    The reversal curse: Llms trained on" a is b" fail to learn" b is a" , author=. arXiv preprint arXiv:2309.12288 , year=

    Pith/arXiv arXiv

  17. [17]

    arXiv preprint arXiv:2411.16353 , year=

    The Two-Hop Curse: LLMs trained on A-> B, B-> C fail to learn A--> C , author=. arXiv preprint arXiv:2411.16353 , year=

    arXiv

  18. [18]

    arXiv preprint arXiv:2311.07468 , year=

    Are we falling in a middle-intelligence trap? an analysis and mitigation of the reversal curse , author=. arXiv preprint arXiv:2311.07468 , year=

    Pith/arXiv arXiv

  19. [19]

    arXiv preprint arXiv:2403.13799 , year=

    Reverse training to nurse the reversal curse , author=. arXiv preprint arXiv:2403.13799 , year=

    Pith/arXiv arXiv

  20. [20]

    arXiv preprint arXiv:2403.00758 , year=

    Mitigating Reversal Curse via Semantic-aware Permutation Training , author=. arXiv preprint arXiv:2403.00758 , year=

    Pith/arXiv arXiv

  21. [21]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Rethinking the Reversal Curse of LLMs: a Prescription from Human Knowledge Reversal , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  22. [22]

    arXiv e-prints , pages=

    Two-stage Generative Question Answering on Temporal Knowledge Graph Using Large Language Models , author=. arXiv e-prints , pages=

  23. [23]

    Proceedings of the Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region , pages=

    Retrieving supporting evidence for generative question answering , author=. Proceedings of the Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region , pages=

  24. [24]

    arXiv preprint arXiv:2308.03296 , year=

    Studying large language model generalization with influence functions , author=. arXiv preprint arXiv:2308.03296 , year=

    Pith/arXiv arXiv

  25. [25]

    Findings of the Association for Computational Linguistics ACL 2024 , pages=

    Unexpected phenomenon: Llms’ spurious associations in information extraction , author=. Findings of the Association for Computational Linguistics ACL 2024 , pages=

    2024

  26. [26]

    Journal of biomedical informatics , volume=

    Development of a benchmark corpus to support the automatic extraction of drug-related adverse effects from medical case reports , author=. Journal of biomedical informatics , volume=. 2012 , publisher=

    2012

  27. [27]

    ACL , pages=

    CBLUE: A Chinese Biomedical Language Understanding Evaluation Benchmark , author=. ACL , pages=

  28. [28]

    arXiv preprint arXiv:2309.10305 , year=

    Baichuan 2: Open large-scale language models , author=. arXiv preprint arXiv:2309.10305 , year=

    Pith/arXiv arXiv

  29. [29]

    IEEE Transactions on Knowledge and Data Engineering , year=

    A Bidirectional Extraction-Then-Evaluation Framework for Complex Relation Extraction , author=. IEEE Transactions on Knowledge and Data Engineering , year=

  30. [30]

    Oscar Sainz and Iker Garc. Go. The Twelfth International Conference on Learning Representations , year=

  31. [31]

    arXiv preprint arXiv:2312.15548 , year=

    YAYI-UIE: A Chat-Enhanced Instruction Tuning Framework for Universal Information Extraction , author=. arXiv preprint arXiv:2312.15548 , year=

    Pith/arXiv arXiv

  32. [32]

    arXiv preprint arXiv:2304.08085 , year=

    Instructuie: Multi-task instruction tuning for unified information extraction , author=. arXiv preprint arXiv:2304.08085 , year=

    Pith/arXiv arXiv

  33. [33]

    arXiv preprint arXiv:2406.14745 , year=

    Relation extraction with fine-tuned large language models in retrieval augmented generation frameworks , author=. arXiv preprint arXiv:2406.14745 , year=

    Pith/arXiv arXiv

  34. [34]

    Proceedings of the conference

    Revisiting relation extraction in the era of large language models , author=. Proceedings of the conference. Association for Computational Linguistics. Meeting , volume=. 2023 , organization=

    2023

  35. [35]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Unified Structure Generation for Universal Information Extraction , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  36. [36]

    K now C oder: Coding Structured Knowledge into LLM s for Universal Information Extraction

    Li, Zixuan and Zeng, Yutao and Zuo, Yuxin and Ren, Weicheng and Liu, Wenxuan and Su, Miao and Guo, Yucan and Liu, Yantao and Lixiang, Lixiang and Hu, Zhilei and Bai, Long and Li, Wei and Liu, Yidan and Yang, Pan and Jin, Xiaolong and Guo, Jiafeng and Cheng, Xueqi. K now C oder: Coding Structured Knowledge into LLM s for Universal Information Extraction. P...

    doi:10.18653/v1/2024.acl- 2024

  37. [37]

    RUIE : Retrieval-based Unified Information Extraction using Large Language Model

    Liao, Xincheng and Duan, Junwen and Huang, Yixi and Wang, Jianxin. RUIE : Retrieval-based Unified Information Extraction using Large Language Model. Proceedings of the 31st International Conference on Computational Linguistics. 2025

    2025

  38. [38]

    Knowledge Graph Unlearning with Schema

    Xiao, Yang and Ye, Ruimeng and Hui, Bo. Knowledge Graph Unlearning with Schema. Proceedings of the 31st International Conference on Computational Linguistics. 2025

    2025

  39. [39]

    Commonsense Subgraph for Inductive Relation Reasoning with Meta-learning

    Zhao, Feng and Zhang, Zhilu and Yan, Cheng and Liu, Xianggan. Commonsense Subgraph for Inductive Relation Reasoning with Meta-learning. Proceedings of the 31st International Conference on Computational Linguistics. 2025

    2025

  40. [40]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Disentangled Counterfactual Learning for Physical Audiovisual Commonsense Reasoning , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  41. [41]

    OpenEQA: Embodied Question Answering in the Era of Foundation Models , year=

    Majumdar, Arjun and Ajay, Anurag and Zhang, Xiaohan and Putta, Pranav and Yenamandra, Sriram and Henaff, Mikael and Silwal, Sneha and Mcvay, Paul and Maksymets, Oleksandr and Arnaud, Sergio and Yadav, Karmesh and Li, Qiyang and Newman, Ben and Sharma, Mohit and Berges, Vincent and Zhang, Shiqi and Agrawal, Pulkit and Bisk, Yonatan and Batra, Dhruv and Kal...

  42. [42]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    Mastering Context-to-Label Representation Transformation for Event Causality Identification with Diffusion Models , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2024 , month=. doi:10.1609/aaai.v38i17.29840 , abstractNote=

    work page doi:10.1609/aaai.v38i17.29840 2024

  43. [43]

    Empowering Large Language Models for Textual Data Augmentation

    Li, Yichuan and Ding, Kaize and Wang, Jianling and Lee, Kyumin. Empowering Large Language Models for Textual Data Augmentation. Findings of the Association for Computational Linguistics: ACL 2024. 2024. doi:10.18653/v1/2024.findings-acl.756

    work page doi:10.18653/v1/2024.findings-acl.756 2024

  44. [44]

    Data Augmentation using LLM s: Data Perspectives, Learning Paradigms and Challenges

    Ding, Bosheng and Qin, Chengwei and Zhao, Ruochen and Luo, Tianze and Li, Xinze and Chen, Guizhen and Xia, Wenhan and Hu, Junjie and Luu, Anh Tuan and Joty, Shafiq. Data Augmentation using LLM s: Data Perspectives, Learning Paradigms and Challenges. Findings of the Association for Computational Linguistics: ACL 2024. 2024. doi:10.18653/v1/2024.findings-acl.97

    work page doi:10.18653/v1/2024.findings-acl.97 2024

  45. [45]

    Proceedings of the 61st ACM/IEEE Design Automation Conference , pages=

    Data is all you need: Finetuning llms for chip design via an automated design-data augmentation framework , author=. Proceedings of the 61st ACM/IEEE Design Automation Conference , pages=

  46. [46]

    and Kapanipathi, Pavan

    Abdelaziz, Ibrahim and Basu, Kinjal and Agarwal, Mayank and Kumaravel, Sadhana and Stallone, Matthew and Panda, Rameswar and Rizk, Yara and Bhargav, G P Shrivatsa and Crouse, Maxwell and Gunasekara, Chulaka and Ikbal, Shajith and Joshi, Sachindra and Karanam, Hima and Kumar, Vineet and Munawar, Asim and Neelam, Sumit and Raghu, Dinesh and Sharma, Udit and...

    2024

  47. [47]

    Mixing It Up: The Cocktail Effect of Multi-Task Fine-Tuning on

    Menachem Brief and Oded Ovadia and Gil Shenderovitz and Noga BenYoash and Rachel Lemberg and Eitam Sheetrit , year=. Mixing It Up: The Cocktail Effect of Multi-Task Fine-Tuning on

  48. [48]

    LLM s Know What They Need: Leveraging a Missing Information Guided Framework to Empower Retrieval-Augmented Generation

    Wang, Keheng and Duan, Feiyu and Li, Peiguang and Wang, Sirui and Cai, Xunliang. LLM s Know What They Need: Leveraging a Missing Information Guided Framework to Empower Retrieval-Augmented Generation. Proceedings of the 31st International Conference on Computational Linguistics. 2025

    2025

  49. [49]

    DRAGIN : Dynamic Retrieval Augmented Generation based on the Real-time Information Needs of Large Language Models

    Su, Weihang and Tang, Yichen and Ai, Qingyao and Wu, Zhijing and Liu, Yiqun. DRAGIN : Dynamic Retrieval Augmented Generation based on the Real-time Information Needs of Large Language Models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. doi:10.18653/v1/2024.acl-long.702

    doi:10.18653/v1/2024.acl-long.702 2024

  50. [50]

    Retrieval Augmented Generation or Long-Context LLM s? A Comprehensive Study and Hybrid Approach

    Li, Zhuowan and Li, Cheng and Zhang, Mingyang and Mei, Qiaozhu and Bendersky, Michael. Retrieval Augmented Generation or Long-Context LLM s? A Comprehensive Study and Hybrid Approach. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track. 2024. doi:10.18653/v1/2024.emnlp-industry.66

    doi:10.18653/v1/2024.emnlp-industry.66 2024

  51. [51]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Mastering context-to-label representation transformation for event causality identification with diffusion models , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  52. [52]

    arXiv preprint arXiv:2407.21783 , year=

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    Pith/arXiv arXiv

  53. [53]

    V ae D iff- D oc RE : End-to-end Data Augmentation Framework for Document-level Relation Extraction

    Tran, Khai Phan and Hua, Wen and Li, Xue. V ae D iff- D oc RE : End-to-end Data Augmentation Framework for Document-level Relation Extraction. Proceedings of the 31st International Conference on Computational Linguistics. 2025

    2025

  54. [54]

    Fine-tuning Language Models for Triple Extraction with Data Augmentation

    Zhang, Yujia and Sadler, Tyler and Taesiri, Mohammad Reza and Xu, Wenjie and Reformat, Marek. Fine-tuning Language Models for Triple Extraction with Data Augmentation. Proceedings of the 1st Workshop on Knowledge Graphs and Large Language Models (KaLLM 2024). 2024. doi:10.18653/v1/2024.kallm-1.12

    work page doi:10.18653/v1/2024.kallm-1.12 2024

  55. [55]

    CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge , author=. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , pages=

    2019

  56. [56]

    Database , volume=

    Overview of DrugProt task at BioCreative VII: data and methods for large-scale text mining and knowledge graph generation of heterogenous chemical--protein relations , author=. Database , volume=. 2023 , publisher=

    2023

  57. [57]

    Linguistic Data Consortium, Philadelphia , volume=

    ACE 2005 multilingual training corpus , author=. Linguistic Data Consortium, Philadelphia , volume=

    2005

  58. [58]

    Journal of artificial intelligence research , volume=

    Reinforcement learning: A survey , author=. Journal of artificial intelligence research , volume=

  59. [59]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    Pith/arXiv arXiv

  60. [60]

    SC22: International Conference for High Performance Computing, Networking, Storage and Analysis , pages=

    Deepspeed-inference: enabling efficient inference of transformer models at unprecedented scale , author=. SC22: International Conference for High Performance Computing, Networking, Storage and Analysis , pages=. 2022 , organization=

    2022

  61. [61]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    AutoPrompt: Eliciting Knowledge from Language Models with Automatically Generated Prompts , author=. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    2020

  62. [62]

    Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    Dynamic Strategy Planning for Efficient Question Answering with Large Language Models , author=. Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    2025

  63. [63]

    Second Conference on Language Modeling , year=

    Meta-Learning for Speeding Up Large Model Inference in Decentralized Environments , author=. Second Conference on Language Modeling , year=

  64. [64]

    Sirui Hong and Mingchen Zhuge and Jonathan Chen and Xiawu Zheng and Yuheng Cheng and Jinlin Wang and Ceyao Zhang and Zili Wang and Steven Ka Shing Yau and Zijuan Lin and Liyang Zhou and Chenyu Ran and Lingfeng Xiao and Chenglin Wu and J. Meta. The Twelfth International Conference on Learning Representations , year=

  65. [65]

    arXiv preprint arXiv:2505.11896 , year=

    AdaCoT: Pareto-Optimal Adaptive Chain-of-Thought Triggering via Reinforcement Learning , author=. arXiv preprint arXiv:2505.11896 , year=

    Pith/arXiv arXiv

  66. [66]

    The Twelfth International Conference on Learning Representations , year=

    PromptAgent: Strategic Planning with Language Models Enables Expert-level Prompt Optimization , author=. The Twelfth International Conference on Learning Representations , year=

  67. [67]

    arXiv preprint arXiv:2505.17895 , year=

    DataRater: Meta-Learned Dataset Curation , author=. arXiv preprint arXiv:2505.17895 , year=

    arXiv

  68. [68]

    International Conference on Machine Learning , pages=

    Grad-match: Gradient matching based data subset selection for efficient deep model training , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    2021

  69. [69]

    arXiv preprint arXiv:2507.17702 , year=

    Towards greater leverage: Scaling laws for efficient mixture-of-experts language models , author=. arXiv preprint arXiv:2507.17702 , year=

    arXiv

  70. [70]

    arXiv preprint arXiv:2508.17291 , year=

    Meta-R1: Empowering Large Reasoning Models with Metacognition , author=. arXiv preprint arXiv:2508.17291 , year=

    Pith/arXiv arXiv

  71. [71]

    and Sifre, Laurent , title =

    Hoffmann, Jordan and Borgeaud, Sebastian and Mensch, Arthur and Buchatskaya, Elena and Cai, Trevor and Rutherford, Eliza and de Las Casas, Diego and Hendricks, Lisa Anne and Welbl, Johannes and Clark, Aidan and Hennigan, Tom and Noland, Eric and Millican, Katie and van den Driessche, George and Damoc, Bogdan and Guy, Aurelia and Osindero, Simon and Simony...

    2022

  72. [72]

    arXiv preprint arXiv:2307.09288 , year=

    Llama 2: Open foundation and fine-tuned chat models , author=. arXiv preprint arXiv:2307.09288 , year=

    Pith/arXiv arXiv