Pith. sign in

REVIEW 4 major objections 8 minor 46 references

Beyond path selection: Better LLMs for Scientific Information Extraction with MimicSFT and Relevance and Rule-induced(R$^2$)GRPO

T0 review · 4 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a two-stage post-training recipe — supervised fine-tuning with templated pseudo-reasoning, then a GRPO variant with relevance and rule-pattern rewards — lifts a small instruction-tuned LLM's scientific relation…

desk verdict MimicSFT + R2GRPO is a promising empirical recipe for scientific IE, but the paper's own Best F1@K data undercut the 'beyond path selection' framing. read the letter →

arxiv 2505.22068 v1 pith:PSDXJJZP submitted 2025-05-28 cs.CL cs.AIcs.IR

Ran Li , Shimin Di , Yuchen Liu , Chen Jing , Yu Qiu , Lei Chen This is my paper · ORCID
classification cs.CLcs.AIcs.IR
keywords scientificinformationextractionrelationnamedentityrecognitionreinforcementlearningwithverifiablerewardsGRPOsupervisedfine-tuningreasoningcapacitychain-of-thought
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

The paper takes issue with the claim that reinforcement learning with verifiable rewards only reorders a model's existing reasoning paths without adding capability. It argues that in scientific information extraction, a task that needs both domain memorization and rule-based inference, both supervised fine-tuning and RLVR can genuinely expand what a model can retrieve and reason about. To show this, it introduces MimicSFT, which trains an LLM to emit fixed structured reasoning templates before its answer, and R2GRPO, which adds relevance and rule-pattern rewards to the GRPO objective. On the SciER benchmark, the two-stage combination reaches 66.81 relation-extraction F1, above the best specialized supervised model at 62.32, while the Best F1@K analysis indicates knowledge-boundary expansion, not just better path ranking. If correct, the result offers a cheap post-training route — no high-quality chain-of-thought data needed — for turning a small instruction-tuned model into a competitive scientific information extraction system.

What carries the argument

The machinery is a constrained-generation view of information extraction with schema and factual constraints, a unified gradient framework where SFT and GRPO differ only in the token-level gradient coefficient, and a composite reward $R(o_i,x,y_{\text{gold}}) = w_1R_{\mathrm{F1}} + w_2R_{\mathrm{span}} + w_3R_{\mathrm{relevancy}} + w_4R_{\mathrm{rule}}$. The two-level reasoning structure — a fixed template level $z_1$ inside reasoning tags and an RL-optimized level $z_2$ inside think tags — decomposes constraint satisfaction so the model first addresses schema constraints and then factual grounding. The reward's relevancy term checks that cited evidence appears in the source sentence, and the rule term rewards adherence to domain patterns such as causal or comparative verbs.

What would settle it

Re-run the leading supervised baseline in the same evaluation harness as R2GRPO* — identical prompts, entity-span normalization, relation-type label set, and Rel/Rel+ definitions — on the SciER test split; if the measured Rel F1 is at least as high as 66.81, the paper's central claim of surpassing specialized supervised models is false.

Watch

Extended reading notes

Core claim

The central discovery is that RLVR and SFT are complementary in constrained generation: SFT with pseudo-reasoning templates (MimicSFT) widens the model's knowledge frontier, while reward-driven RL (R2GRPO) sharpens the probability placed on high-quality outputs, so the combination outperforms either alone. On SciER, R2GRPO* reaches 66.81 Rel F1 and 65.95 Rel+ F1, surpassing the best supervised system at 62.32 Rel and all compared LLMs; on out-of-domain data it also improves over the base model and plain GRPO. The paper interprets the Best F1@K curves as evidence that both stages raise the ceiling of the model's reasoning capacity, contradicting the path-selection-only view.

Load-bearing premise

The load-bearing assumption is that the published scores for the specialized supervised baselines are directly comparable to the proposed method's scores, i.e., that entity spans and relation triples are normalized and scored in exactly the same way; if the protocols differ, the claim of surpassing supervised models may be an artifact.

Editorial extensions

If this is right

  • A small instruction-tuned model post-trained with MimicSFT and R2GRPO can exceed specialized supervised extractors on scientific relation extraction without high-quality chain-of-thought data.
  • The Best F1@K analysis implies both SFT and RLVR expand the model's ability frontier, so RLVR's gain on these tasks is not merely re-ranking existing hypotheses.
  • MimicSFT's fixed reasoning template outperforms standard SFT, indicating that the structure of the reasoning process, not its detailed content, drives much of the gain in constrained generation.
  • Combining MimicSFT with R2GRPO outperforms either stage alone, showing that knowledge acquisition and reward-based reasoning refinement are complementary rather than competing.
  • The larger gains on the out-of-domain set suggest structured reasoning improves generalization rather than memorization of the training distribution.

Reading between the lines

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

  • Editorial inference: the rule-pattern reward is built on hand-selected trigger words and phrases, so the method as described may need a new pattern list when moved to another scientific subfield; learning the pattern vocabulary automatically would be a natural testable extension.
  • Editorial inference: because the supervised-baseline scores are quoted from the benchmark paper rather than re-measured, the headline comparison would be strengthened by a same-harness re-run of those baselines on the identical prompts and scoring script.
  • Editorial inference: the Best F1@K analysis is based on a 50-sample subset, so scaling that analysis to the full test set would show whether the reasoning-capacity gain is broad or concentrated in a few hard cases.
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 / 8 minor

Summary. The paper proposes a two-stage post-training method for scientific information extraction (SciIE): MimicSFT, which fine-tunes an LLM on structured pseudo-reasoning templates without requiring high-quality CoT data, and R2GRPO, a GRPO variant with a composite reward combining F1, entity span, relevance, and rule-pattern signals. The authors argue that, contrary to recent claims that RLVR only refines reasoning paths without expanding reasoning capacity, both SFT and RLVR can improve reasoning capacity in SciIE. They report that the combination R2GRPO* (MimicSFT followed by R2GRPO) surpasses baseline LLMs and specialized supervised models on relation extraction, and they analyze Best F1@K and Avg@K to support the capacity-expansion claim. The paper includes a theoretical discussion of hierarchical constraint satisfaction and a unified gradient view of SFT and RL.

Significance. If substantiated, the claim that RLVR can go beyond path selection in a knowledge-intensive task like scientific IE would be an important contribution, as it challenges recent negative results from mathematical reasoning and suggests a task-dependent view of what RL post-training learns. MimicSFT's use of templated reasoning without distilled CoT data is also practically appealing, and the reported improvements over strong supervised BERT baselines on SciER, if valid, would be of broad interest. However, the significance hinges on two load-bearing points: whether the 'beyond path selection' claim is actually supported by the Best F1@K analysis, and whether the comparison to supervised baselines is protocol-comparable. Both need to be addressed before the central claims can be accepted.

major comments (4)
  1. [Section 4.3, Figures 2-3] The 'beyond path selection' claim is not supported by the presented Best F1@K evidence. The text explicitly acknowledges that 'SFT-based models (particularly MimicSFT) achieve slightly higher Best F1@K at very large K values' while RLVR models show better Best F1@1 and Avg@K. Since R2GRPO* is initialized from MimicSFT, the RL stage does not raise the large-K ceiling of its own SFT initialization; it sharpens the output distribution. This is precisely the path re-ranking behavior that the paper claims to rule out. To support the title claim, the paper would need to show that R2GRPO*'s Best F1@K exceeds MimicSFT's at large K with statistical significance, or to provide an alternative operationalization of 'reasoning capacity' that is not confounded with top-1 accuracy. As written, the statement that 'consistent improvement in Best F1@K scores, even at large K values, indicates a genuine expansion' is contradicted by the paper's own observations.
  2. [Section 4.2, Table 1] The comparison to supervised baselines (PURE, PL-Marker, HGERE) uses numbers taken from the original SciER paper [41] rather than re-running those systems in the same evaluation pipeline. The definitions of Rel and Rel+ in Section 4.1 are potentially ambiguous: Rel already requires that 'the types and spans of both entities and the relation type match a gold relation,' while Rel+ 'further requires the entity type is correct in the triples.' If the original benchmark uses a different protocol (e.g., boundary vs. strict matching, or different handling of entity spans), the claimed superiority of R2GRPO* over supervised models could be an artifact. The authors should either reproduce the baselines in their own pipeline or provide a detailed protocol-matching analysis.
  3. [Section 3.3.3 and Appendix A.3] The RL training data selection criterion ('instances where SFT performs poorly but clear reward signals exist') is not operationalized. The RL stage uses only 1K samples selected by this criterion, while SFT/MimicSFT are trained on the full dataset. This distribution mismatch can inflate RL gains if the selected subset is biased toward examples where SFT's weaknesses are most easily corrected by reward optimization. The paper does not report an ablation with a random or representative subset of the same size, nor does it specify the exact rule used to determine 'SFT performs poorly.' This is load-bearing because the central empirical claim is that R2GRPO improves over its SFT initialization.
  4. [Section 3.4, Eq. (10)] The hierarchical constraint-satisfaction inequality in Eq. (10) is assumed and then claimed to be verified by Figures 2-3. However, Figures 2-3 are reported to show the opposite at large K: MimicSFT has higher Best F1@K than R2GRPO*. Thus the 'verification' does not support Eq. (10) as stated. The theoretical discussion should be revised to clarify which quantity is actually being verified (e.g., Best F1@1 or Avg@K) and should not claim support from experiments that contradict the inequality under the paper's own reading.
minor comments (8)
  1. [Section 3.3.1, Eq. (11)] Equation (11) contains a garbled expectation symbol ('E⅁⅁x,o') and the gradient coefficient GC is not rigorously defined. This equation is a restatement of the GRPO objective, not a new derivation; the authors should state this explicitly and fix the notation.
  2. [Abstract and Section 1] There are grammatical errors, e.g., 'Previous study suggest' and 'underperforms small Bert-based models' should be 'Previous studies suggest' and 'underperform small BERT-based models.'
  3. [Section 3.3.2] The rule-pattern reward R_rule uses 'I(pattern_k satisfied...)' without defining the indicator function or explaining how the set of patterns is obtained (e.g., is it manually defined or automatically mined?).
  4. [Section 3.3.2] The relevancy reward uses 'Map(ci, evidence_gold)' without defining what 'Map' is or how evidence_gold is obtained during RL training; clarify whether evidence_gold is available from the dataset or extracted heuristically.
  5. [Section 4.1] The definitions of Rel and Rel+ appear contradictory: Rel is said to require matching 'types and spans of both entities,' while Rel+ 'further requires the entity type is correct.' Please clarify the exact distinction between the two metrics.
  6. [Section 4.3] The Best F1@K analysis is based on only 50 'challenging' samples; the selection of these samples is not described, and no confidence intervals are reported. This makes it difficult to assess whether the observed differences are meaningful.
  7. [Appendix A.1/A.2] The appendix training prompt contains a copy-paste error: the 'Ner Background' section includes a sentence about determining relationships, which belongs in the relation extraction prompt.
  8. [Section 4.3, 'Exploration-Exploitation Balance'] The statement that 'the slightly lower Best F1@K of R2GRPO compared to MimicSFT at very large K values reflects an intentional trade-off' is speculative and not supported by any direct experimental manipulation; it should be phrased as an interpretation, not a result.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central empirical claim is trained on a held-out split, baselines are external, and the only 'theoretical' step is explicitly flagged as an assumption to be tested by the same experiments.

full rationale

The paper's main claim is non-circular. MimicSFT and R2GRPO are trained on the SciER training split (including a 1K RL subset) and evaluated on held-out SciER and OOD test sets; no test labels enter the training objective or the reward functions. The supervised baselines (PURE, PL-Marker, HGERE) are taken from the external SciER paper [41], which is a protocol-comparability risk rather than circularity. The 'theoretical analysis' in Section 3.4 explicitly frames Eq. 10 as an assumption ('we assume that... We will verify this later through experiments'), so it is a hypothesis to be tested, not a derivation that assumes its conclusion. Eq. 11's unified gradient framework is a restatement of the GRPO objective and is not used to manufacture a prediction. There are no load-bearing self-citations: references [39] and [26] are external works with no author overlap. The paper's own admission that MimicSFT achieves slightly higher Best F1@K at very large K values weakens the 'beyond path selection' framing, but that is an evidentiary or interpretation concern, not a circular step by the definitions used here.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim depends on a small set of undisclosed reward parameters and on assumptions about the comparability of external baselines, the representativeness of the 50-sample Best@K subset, an unproven inequality in the theoretical argument, and the neutrality of the RL data selection. No new physical or conceptual entities are introduced beyond the training methods themselves.

free parameters (4)
  • Reward weights w1-w4
    In Eq. 6, the composite reward weights are tunable but their values are not reported, so the reader cannot replicate the exact objective.
  • Rrelevancy penalty lambda_penalty
    In the relevancy reward, lambda_penalty and the length threshold are not specified.
  • Rrelevancy length threshold
    The threshold above which the length penalty applies is not given.
  • Rule-pattern weights w_k
    Rrule is defined as a weighted sum over patterns, but the pattern set and weights are not enumerated.
assumptions (4)
  • domain assumption The supervised baseline results (PURE, PL-Marker, HGERE) reported in the SciER paper [41] were obtained under the same evaluation protocol used here.
    The paper uses these numbers as benchmarks without re-running the models, so the comparison rests on protocol equivalence.
  • ad hoc to paper The 50-sample subset of the test set used for Best F1@K is representative of the full test set.
    The analysis in Section 4.3 draws conclusions about reasoning capacity from a subset of 50 samples without explaining how they were chosen.
  • ad hoc to paper Equation 10, the hierarchical constraint-satisfaction inequality, is assumed and then claimed to be verified by the same experiments.
    Section 3.4 states P(y in C | x; theta_hier) >= P(y in C | x; theta_direct) as an assumption and says it will be verified later, making the theoretical support an assumption rather than a derivation.
  • domain assumption The RL training subset selection based on SFT poor performance preserves the test distribution.
    If the selection over-weights instances that are hard for SFT, the RL gains may not generalize to the full test set; the paper does not compare with a random subset baseline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond path selection: Better LLMs for Scientific Information Extraction with MimicSFT and Relevance and Rule-induced(R$^2$)GRPO." pith.science (2026). https://pith.science/paper/PSDXJJZP

@misc{pith2026250522068,
  author       = {Pith},
  title        = {Pith review of: Beyond path selection: Better LLMs for Scientific Information Extraction with MimicSFT and Relevance and Rule-induced(R$^2$)GRPO},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PSDXJJZP}},
  note         = {Machine review of arXiv:2505.22068}
}
abstract

Previous study suggest that powerful Large Language Models (LLMs) trained with Reinforcement Learning with Verifiable Rewards (RLVR) only refines reasoning path without improving the reasoning capacity in math tasks while supervised-finetuning(SFT) with distillation can. We study this from the view of Scientific information extraction (SciIE) where LLMs and reasoning LLMs underperforms small Bert-based models. SciIE require both the reasoning and memorization. We argue that both SFT and RLVR can refine the reasoning path and improve reasoning capacity in a simple way based on SciIE. We propose two-stage training with 1. MimicSFT, using structured reasoning templates without needing high-quality chain-of-thought data, 2. R$^2$GRPO with relevance and rule-induced rewards. Experiments on scientific IE benchmarks show that both methods can improve the reasoning capacity. R$^2$GRPO with mimicSFT surpasses baseline LLMs and specialized supervised models in relation extraction. Our code is available at https://github.com/ranlislz/R2GRPO.

Figures

Figures reproduced from arXiv: 2505.22068 by the authors.

Figure 1
Figure 1. Our two-stage training for scientific IE (right) and the performance gain (left) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Best F1@K scores representing the reasoning capacity and Avg@K scores representing the [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Best F1@K scores representing the reasoning capacity and Avg@K scores representing the [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Performance v.s temperature Complementary Effects of SFT and RLVR: While SFT-based models (particularly MimicSFT) achieve slightly higher Best F1@K at very large K values, RLVR models demonstrate superior Avg@K and Best F1@1 scores. This reveals a fundamental trade-off…
Figure 5
Figure 5. Figure 5: Response length(a) and Reward(b) v.s. training steps for R2GRPO Exploration-Exploitation Balance: The slightly lower Best F1@K of R2GRPO compared to MimicSFT at very large K values reflects an intentional trade-off. R2GRPO optimizes for high-reward trajectories within …
Figure 6
Figure 6. Figure 6: Performance v.s response token length. The deeper the color, the higher the temperature. Component Contribution Analysis [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: R2GRPO training detail v.s. steps Ner Background Extract specific entities from the following sentence. The entities to be identified are: ’Dataset’, ’Task’, and ’Method’. ### Entity Definitions: - ’Task’: A task in machine learning refers to the specific problem or ty…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 10 canonical work pages

  1. [41]

    Scier: An entity and relation extraction dataset for datasets, methods, and tasks in scientific documents

    Qi Zhang, Zhijia Chen, Huitong Pan, Cornelia Caragea, Longin Jan Latecki, and Eduard Dragut. Scier: An entity and relation extraction dataset for datasets, methods, and tasks in scientific documents. arXiv preprint arXiv:2410.21155, 2024. 12

    arXiv

  2. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  3. [2]

    Training a helpful and harmless assistant with reinforcement learning from human feedback

    Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022

    arXiv 2022

  4. [3]

    Scibert: A pretrained language model for scientific text

    Iz Beltagy, Kyle Lo, and Arman Cohan. Scibert: A pretrained language model for scientific text. arXiv preprint arXiv:1903.10676, 2019

    arXiv 1903

  5. [4]

    Codekgc: Code language model for generative knowledge graph construction

    Zhen Bi, Jing Chen, Yinuo Jiang, Feiyu Xiong, Wei Guo, Huajun Chen, and Ningyu Zhang. Codekgc: Code language model for generative knowledge graph construction. ACM Transac- tions on Asian and Low-Resource Language Information Processing , 23(3):1–16, 2024

    work page 2024

  6. [5]

    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

  7. [6]

    Sft memorizes, rl generalizes: A comparative study of foundation model post-training

    Tianzhe Chu, Yuexiang Zhai, Jihan Yang, Shengbang Tong, Saining Xie, Dale Schuurmans, Quoc V Le, Sergey Levine, and Yi Ma. Sft memorizes, rl generalizes: A comparative study of foundation model post-training. arXiv preprint arXiv:2501.17161, 2025

    arXiv 2025

  8. [7]

    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

Show all 46 references
  1. [8]

    Structured information extraction from scientific text with large language models

    John Dagdelen, Alexander Dunn, Sanghoon Lee, Nicholas Walker, Andrew S Rosen, Gerbrand Ceder, Kristin A Persson, and Anubhav Jain. Structured information extraction from scientific text with large language models. Nature Communications, 15(1):1418, 2024. 10

    2024

  2. [9]

    Gemini 2.0 flash thinking„ 2024

    Google DeepMind. Gemini 2.0 flash thinking„ 2024. URL https://deepmind.google/ technologies/gemini/flash-thinking/. Accessed: 2025-04-15

    2024

  3. [10]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human langu...

    2019

  4. [11]

    Competitive programming with large reasoning models

    Ahmed El-Kishky, Alexander Wei, Andre Saraiva, Borys Minaiev, Daniel Selsam, David Dohan, Francis Song, Hunter Lightman, Ignasi Clavera, Jakub Pachocki, et al. Competitive programming with large reasoning models. arXiv preprint arXiv:2502.06807, 2025

    2025 arXiv

  5. [12]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  6. [13]

    Distilling the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015

    2015 arXiv

  7. [14]

    Training compute-optimal large language models

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556, 2022

    2022 arXiv

  8. [15]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1 (2):3, 2022

    2022

  9. [16]

    Openai o1 system card

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024

    2024 arXiv

  10. [17]

    Scaling laws for neural language models

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020

    2001 arXiv

  11. [18]

    Llms can easily learn to rea- son from demonstrations structure, not content, is what matters!, 2025

    Dacheng Li, Shiyi Cao, Tyler Griggs, Shu Liu, Xiangxi Mo, Eric Tang, Sumanth Hegde, Kourosh Hakhamaneshi, Shishir G Patil, Matei Zaharia, et al. Llms can easily learn to rea- son from demonstrations structure, not content, is what matters!, 2025. URL https://arxiv. org/abs/2502.07374

    2025 arXiv

  12. [19]

    Revisiting large language models as zero-shot relation extractors

    Guozheng Li, Peng Wang, and Wenjun Ke. Revisiting large language models as zero-shot relation extractors. arXiv preprint arXiv:2310.05028, 2023

    2023 arXiv

  13. [20]

    Getting more juice out of the sft data: Reward learning from human demonstration improves sft for llm alignment

    Jiaxiang Li, Siliang Zeng, Hoi-To Wai, Chenliang Li, Alfredo Garcia, and Mingyi Hong. Getting more juice out of the sft data: Reward learning from human demonstration improves sft for llm alignment. Advances in Neural Information Processing Systems , 37:124292–124318, 2024

    2024

  14. [21]

    Pivoine: Instruction tuning for open-world information extraction

    Keming Lu, Xiaoman Pan, Kaiqiang Song, Hongming Zhang, Dong Yu, and Jianshu Chen. Pivoine: Instruction tuning for open-world information extraction. arXiv preprint arXiv:2305.14898, 2023

    2023 arXiv

  15. [22]

    Urbankgent: A unified large language model agent framework for urban knowledge graph construction

    Yansong Ning and Hao Liu. Urbankgent: A unified large language model agent framework for urban knowledge graph construction. arXiv preprint arXiv:2402.06861, 2024

    2024 arXiv

  16. [23]

    Using of jaccard coefficient for keywords similarity

    Suphakit Niwattanakul, Jatsada Singthongchai, Ekkachai Naenudorn, and Supachanun Wanapu. Using of jaccard coefficient for keywords similarity. In Proceedings of the international multiconference of engineers and computer scientists , volume 1, pages 380–384, 2013

    2013

  17. [24]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems , 35:277...

    2022

  18. [25]

    Improving language understanding by generative pre-training

    Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018

    2018

  19. [26]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  20. [27]

    To cot or not to cot? chain- of-thought helps mainly on math and symbolic reasoning

    Zayne Sprague, Fangcong Yin, Juan Diego Rodriguez, Dongwei Jiang, Manya Wadhwa, Prasann Singhal, Xinyu Zhao, Xi Ye, Kyle Mahowald, and Greg Durrett. To cot or not to cot? chain- of-thought helps mainly on math and symbolic reasoning. arXiv preprint arXiv:2409.12183, 2024

    2024 arXiv

  21. [28]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context

    Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024

    2024 arXiv

  22. [29]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timo- thée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  23. [30]

    Gpt-ner: Named entity recognition via large language models

    Shuhe Wang, Xiaofei Sun, Xiaoya Li, Rongbin Ouyang, Fei Wu, Tianwei Zhang, Jiwei Li, and Guoyin Wang. Gpt-ner: Named entity recognition via large language models. arXiv preprint arXiv:2304.10428, 2023

    2023 arXiv

  24. [31]

    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

  25. [32]

    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

  26. [33]

    Zero-shot information extraction via chatting with chatgpt

    Xiang Wei, Xingyu Cui, Ning Cheng, Xiaobin Wang, Xin Zhang, Shen Huang, Pengjun Xie, Jinan Xu, Yufeng Chen, Meishan Zhang, et al. Zero-shot information extraction via chatting with chatgpt. arXiv e-prints, pages arXiv–2302, 2023

    2023

  27. [34]

    Empirical study of zero-shot ner with chatgpt

    Tingyu Xie, Qi Li, Jian Zhang, Yan Zhang, Zuozhu Liu, and Hongwei Wang. Empirical study of zero-shot ner with chatgpt. arXiv preprint arXiv:2310.10035, 2023

    2023 arXiv

  28. [35]

    Joint entity and relation extraction with span pruning and hypergraph neural networks

    Zhaohui Yan, Songlin Yang, Wei Liu, and Kewei Tu. Joint entity and relation extraction with span pruning and hypergraph neural networks. arXiv preprint arXiv:2310.17238, 2023

    2023 arXiv

  29. [36]

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

    2024 arXiv

  30. [37]

    Packed levitated marker for entity and relation extraction

    Deming Ye, Yankai Lin, Peng Li, and Maosong Sun. Packed levitated marker for entity and relation extraction. arXiv preprint arXiv:2109.06067, 2021

    2021 arXiv

  31. [38]

    Zero-shot temporal relation extraction with chatgpt

    Chenhan Yuan, Qianqian Xie, and Sophia Ananiadou. Zero-shot temporal relation extraction with chatgpt. arXiv preprint arXiv:2304.05454, 2023

    2023 arXiv

  32. [39]

    Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model? arXiv preprint arXiv:2504.13837, 2025

    Yang Yue, Zhiqi Chen, Rui Lu, Andrew Zhao, Zhaokai Wang, Shiji Song, and Gao Huang. Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model? arXiv preprint arXiv:2504.13837, 2025

    2025 arXiv

  33. [40]

    Extract, define, canonicalize: An llm-based framework for knowledge graph construction

    Bowen Zhang and Harold Soh. Extract, define, canonicalize: An llm-based framework for knowledge graph construction. arXiv preprint arXiv:2404.03868, 2024

    2024 arXiv

  34. [42]

    Opt: Open pre-trained transformer language models

    Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022

    2022 arXiv

  35. [43]

    A frustratingly easy approach for joint entity and relation extraction

    Zexuan Zhong and Danqi Chen. A frustratingly easy approach for joint entity and relation extraction. arXiv preprint arXiv:2010.12812, 2020

    2010 arXiv

  36. [44]

    Lima: Less is more for alignment

    Chunting Zhou, Pengfei Liu, Puxin Xu, Srinivasan Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, et al. Lima: Less is more for alignment. Advances in Neural Information Processing Systems, 36:55006–55021, 2023

    2023

  37. [45]

    Llms for knowledge graph construction and reasoning: Recent capabilities and future opportunities

    Yuqi Zhu, Xiaohan Wang, Jing Chen, Shuofei Qiao, Yixin Ou, Yunzhi Yao, Shumin Deng, Huajun Chen, and Ningyu Zhang. Llms for knowledge graph construction and reasoning: Recent capabilities and future opportunities. World Wide Web, 27(5):58, 2024

    2024

  38. [46]

    factual, content-bearing

    Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 13 A Appendix A.1 R 2GRPO Training A.2 Dataset Statistics We ...

    1909 arXiv

Pith tools

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