Pith. sign in

REVIEW 4 major objections 5 minor 29 references

This paper argues that explicitly structuring a question into knowledge-driven keypoint groups, then reasoning over each group separately, lets a small language model outperform both chain-of-thought and standard retrieval on medical exam q

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 17:23 UTC pith:X4KWOPZO

load-bearing objection A plausible modular RAG pipeline for small models, but the headline claim that structuring—not compute/supervision—drives the gain is not supported by the current experiments. the 4 major comments →

arxiv 2603.26807 v2 pith:X4KWOPZO submitted 2026-03-26 cs.IR cs.AIcs.CL

GroupRAG: Cognitively Inspired Group-Aware Retrieval and Reasoning via Knowledge-Driven Problem Structuring

Xinyi Duan , Yuanrong Tang , Jiangtao Gong This is my paper
classification cs.IR cs.AIcs.CL
keywords retrieval-augmented generationchain-of-thought promptingquestion decompositionkeypoint groupingmedical question answeringsmall language modelsproblem structuringcognitive science
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 claims that the main bottleneck in complex real-world question answering is not missing knowledge or weak reasoning alone, but the failure to model the latent structure of the question. It proposes GroupRAG, which extracts key information points, groups them by the external knowledge they share, and runs retrieval and reasoning separately for each group before fusing the results into one chain of thought. On a 400-question medical exam set, a fine-tuned 8-billion-parameter model reaches about 72% accuracy, compared with about 61.5% for chain-of-thought and 58.25% for naive retrieval-augmented generation on the same base model. The authors interpret this as evidence that problem structuring, not longer reasoning chains, is the key to reliable performance.

Core claim

The central discovery is that decomposing a question into knowledge-driven groups of key points enables a small model to reason more accurately than either linear chain-of-thought or conventional retrieval-augmented generation. Each group becomes an independent starting point for retrieval and inference, and the local conclusions are later categorized as core, support, or noise, with core and support conclusions fused into a single global chain of thought before answer selection. The authors call this a 'convergent reasoning net' in contrast to the single-start linear or divergent chains used in CoT. Ablations show that the grouping and local-reasoning modules contribute the largest accuracy

What carries the argument

Knowledge-driven keypoint grouping. Each keypoint in the question is first grounded by retrieving external knowledge; keypoints whose retrieved knowledge overlaps are grouped together, so the grouping reflects shared underlying concepts rather than surface similarity. This transforms a monolithic prompt into a structured problem space where each group is a separate retrieval-and-reasoning unit. The second piece is the Weighted Inference F-score (WIF), a reward that trains the global-selection model to preserve all core local conclusions, avoid noise, and tolerate support conclusions when assembling the final chain of thought.

Load-bearing premise

The training and evaluation of all five modules rely on labels and gold groups produced by the same large teacher language model, so the 'true' latent structure of a problem is never independently validated.

What would settle it

Run GroupRAG with a semantic-similarity grouping instead of the knowledge-driven grouping on a fresh 400-question medical set using the same trained modules; if accuracy does not drop materially, the central claim that knowledge-driven grouping is load-bearing collapses. A sharper version: construct questions by concatenating independent fact bundles, verify the pipeline recovers the intended bundles exactly, and test on those.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If problem structuring is the key bottleneck, further scaling retrieval volume or chain length should yield diminishing returns on dense real-world QA; the paper's comparisons on a small model support this.
  • The five-stage pipeline (keypoint extraction, grouping, local reasoning, global reasoning, answer alignment) provides a modular structure whose intermediate labels can be generated once by a large teacher and used to train cheap specialized small models.
  • The WIF reward formalizes the intuition that a good global reasoning step must be complete (cover all core points), faithful (avoid noise), and balanced (tolerate supports), giving a quantitative training signal for selection.
  • Because the method mainly helps small models, it offers a practical route to deploying competent QA systems under tight compute budgets without sacrificing accuracy.

Where Pith is reading between the lines

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

  • A natural extension, not tested in the paper, is whether the knowledge-driven grouping criterion transfers to other dense domains such as legal arguments, radiology reports, or equipment troubleshooting; the Bar Exam mention in the abstract suggests the authors expect it to.
  • If grouping is truly causal rather than merely predictive, then replacing the teacher's groups with random or semantics-only groups on the same test set should degrade accuracy; this is a direct experiment that could be run without new labels.
  • The pipeline's reliance on a single teacher for both labels and evaluation means the strongest version of the claim—that the discovered groups match human 'true' problem structure—remains unverified; a human-annotation study of group quality would be the missing check.
  • The authors suggest multi-agent collaboration as future work; a cheaper test of the same idea would be to let the five modules be trained jointly with a final-answer reward, which would show whether the decomposition itself, or merely the staged supervision, drives the gain.

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

4 major / 5 minor

Summary. The paper proposes GroupRAG, a multi-stage pipeline that extracts keypoints from a question, groups them using retrieval-informed knowledge-driven grouping, performs per-group local reasoning, selects and integrates local conclusions into a global chain-of-thought, and aligns the final answer. The pipeline is implemented as five fine-tuned LLaMA 3.1-8B sub-models trained on GPT-4o-generated intermediate labels, with policy-gradient optimization for the global selection step. On a 400-question MedQA test set, the authors report 71.75% accuracy for GroupRAG with the trained 8B model versus 61.50% for trained CoT and 58.25% for trained Naive RAG. They also report leave-one-out and progressive ablations, and conclude that explicit problem structuring, rather than longer chains, drives the gains.

Significance. If the central claim were established, GroupRAG would be a meaningful contribution to retrieval-augmented reasoning: it offers a modular, interpretable pipeline inspired by cognitive problem-space theory and demonstrates large gains for small language models on a knowledge-intensive QA benchmark. The stage-wise evaluation and cumulative ablation design are transparent, and the end-to-end accuracy comparison is externally grounded by the dataset's answer labels. However, the headline comparison is confounded by differences in supervision, parameter count, and inference budget, and the module-level evaluations rely on the same GPT-4o teacher that generated the training labels. The contribution is therefore promising but not yet proven at the level claimed.

major comments (4)
  1. [Table 3; §3.2, §4.3] The headline comparison is not controlled for the claimed independent variable. GroupRAG consists of five fine-tuned sub-models (Keypoint Extraction, Grouping, Local Reasoning, Global Reasoning selection and synthesis, Answer Alignment), all trained on GPT-4o-generated intermediate labels, plus a policy-gradient selection step with rollouts; inference performs per-keypoint retrieval, per-group retrieval, selection rollouts, and option-level retrieval. The trained CoT baseline is a single model fine-tuned on CoT traces, and Naive RAG is a single model. The 71.75 vs 61.50/58.25 gap therefore conflates problem structuring with roughly 5x model capacity, richer supervision, and far larger inference/retrieval budget. The conclusion in the Abstract and §5 that 'knowledge-driven problem structuring, rather than longer chains, is critical' is not supported by this comparison. Add matched control
  2. [Abstract vs Section 4] The supplied abstract states 'Experiments on MedQA (medical) and Bar Exam QA (legal) show...' but the full text contains no Bar Exam QA dataset, table, or analysis; the paper's abstract in the body says only MedQA. This is a missing-support issue: either add the Bar Exam experiments or remove the claim from the abstract. As it stands, the claimed generalizability across legal reasoning is unsupported.
  3. [§4.1, §4.2] The intermediate module evaluations are self-referential. GPT-4o generates the gold keypoints, gold groupings, local reasoning conclusions, and Core/Support/Noise labels; the same GPT-4o outputs are then used as ground truth to compute Extract F1, Group F1, Local Acc, and Global WIF. Thus Table 1's intermediate metrics measure agreement with the teacher, not fidelity to any task-external ground truth. The end-to-end answer accuracy is non-circular, but the modular interpretability claims are not independently validated. Please add human evaluation or an external gold standard (e.g., expert-annotated keypoints and groups) for at least a sample.
  4. [§4.1, Table 3] The empirical basis is a single 400-question test set with no confidence intervals, bootstrap, or significance tests. With n=400, the reported 71.75% versus 61.50% is about 10.25 percentage points, roughly 2.3 standard errors, but the absence of any variance estimate across seeds or test subsets makes the robustness claims unmeasurable. Also, the test set composition is described only as 'some of which include distracting or irrelevant details'; please specify how the 400 questions were sampled and whether results are stable across random splits.
minor comments (5)
  1. [§3.2] The statement that 'each stage has independent inputs and outputs' is ambiguous, since modules are executed sequentially. Clarify that this refers to independence of module training, not functional independence.
  2. [Table 3] The column header '-(%)' is confusing; specify that the first column denotes no augmentation (neither CoT, RAG, nor GroupRAG).
  3. [§1 and §5] The contribution statement claims comparison with 'a wide range of RAG-based and CoT-based methods,' but the experiments include only Naive RAG and CoT prompting. Either broaden the baseline set or soften the wording.
  4. [§3.3] Retrieval details are missing: corpus, retriever model, top-k, chunk size, embedding method, and number of retrieved passages per keypoint/group/option. These are necessary for reproducibility and for interpreting the ablation results.
  5. [§3.2] The WIF hyperparameters alpha=2.5, beta=2, gamma=0.5 are chosen heuristically with no sensitivity analysis. Since WIF is used both as a training reward and an evaluation metric, please report robustness to these choices or justify them empirically.

Circularity Check

2 steps flagged

Intermediate metrics are self-referential (GPT-4o is both teacher and gold standard); final answer accuracy remains externally grounded.

specific steps
  1. self definitional [§4.1 (Dataset and Model); §4.2 (Evaluation Metrics)]
    "each training question is sequentially processed by the five modules of GroupRAG, each instantiated with GPT-4o [OpenAI, 2024], with the module-wise outputs recorded as intermediate supervision signals for training. ... For each question, the keypoints extracted by the trained lightweight model are compared against a gold standard set of keypoints extracted by GPT-4o."

    The 'latent structure' the paper claims to discover (keypoints, groupings, Core/Support/Noise, local inferences) is operationalized as whatever GPT-4o produces. The same GPT-4o outputs are used as hard training labels (§3.2, §4.1) and then as the 'gold standard' for evaluation (§4.2). Extract F1, Group BCubed F1, Local Acc, and Global WIF therefore measure how closely the trained student imitates its own teacher, not whether the discovered structure is correct. Because final answer accuracy uses external MedQA labels, this is partial rather than total circularity.

  2. self definitional [§3.2 (Policy Optimization); §4.2 (Evaluation Metrics)]
    "This training procedure ensures that the model learns to select local reasoning conclusions that maximize WIF. ... For global reasoning, we adopt the WIF function defined in Subsection 3.2 to assess the model’s ability to distinguish Core, Support, and Noise local conclusions."

    The Global Reasoning selection model is trained by policy gradient against WIF as the reward (§3.2), and the same WIF is then reported as the evaluation metric for Global Reasoning (§4.2). The Core/Support/Noise labels defining WIF are GPT-4o-generated, and the metric hyperparameters (α=2.5, β=2, γ=0.5) are hand-picked. Thus a high Global WIF shows only that the model optimized the hand-defined, teacher-derived objective it was trained on; it is not an independent measure of reasoning quality. Again, the final answer metric is external and gives the paper independent content.

full rationale

GroupRAG's headline comparison (Table 3) is not definitional: final answer accuracy is measured against MedQA's correct options, so the 71.75% vs 61.50%/58.25% gap is an externally grounded outcome. I am not scoring the comparison's confoundedness (unequal supervision, number of trained sub-models, and inference/test-time compute) as circularity. However, the paper's internal validation loop is partially self-referential: five sub-modules are trained on GPT-4o's module outputs (§3.2, §4.1), and the same GPT-4o outputs are then called 'gold standard' for keypoint, grouping, local, and global metrics (§4.2). Extract F1, Group F1, Local Acc, and Global WIF are therefore imitation-fidelity scores, not validated evidence that the latent structure is correct. Global WIF is particularly circular because it is both the policy-gradient training reward and the reported evaluation metric. The Abstract's Bar Exam QA claim has no corresponding experiment in the full text, but that is an unsupported-evidence issue, not circularity. Overall, the central final-accuracy claim has independent content, so the paper is only partially circular.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 1 invented entities

Everything the central claim rests on that the reader didn't pay for upstream: hand-picked WIF hyperparameters, GPT-4o as universal labeler, an unstated retrieval corpus, and a custom dataset split.

free parameters (1)
  • WIF reward hyperparameters = α=2.5, β=2, γ=0.5
    Chosen by hand in Section 3.2 to prioritize Core recall, penalize Noise, and mildly encourage Supports. The final answer accuracy depends on this reward shaping.
axioms (4)
  • domain assumption Keypoint grouping improves problem solving
    The central premise is taken from cognitive science (Newell and Simon) and not independently validated on MedQA beyond the pipeline's own comparison.
  • domain assumption GPT-4o provides valid ground truth for all intermediate stages
    Section 4.1 uses GPT-4o outputs as gold labels for keypoints, groups, local reasoning, and global reasoning; if these are wrong, training and intermediate metrics are corrupted.
  • domain assumption The 2,000-train / 400-test MedQA split is representative and fair
    The paper says 'randomly select' but does not disclose the split methodology or compare against the standard MedQA test set, making selection bias possible.
  • domain assumption An appropriate external retrieval corpus is available
    The retrieval source is never named; the group-driven retrieval performance depends on an unstated corpus.
invented entities (1)
  • Core/Support/Noise categories for local reasoning conclusions no independent evidence
    purpose: Select which intermediate conclusions to feed into global reasoning
    These categories are introduced by the paper and assigned using GPT-4o; no external evidence shows these three categories are the correct decomposition.

pith-pipeline@v1.3.0-alltime-deepseek · 246 in / 7336 out tokens · 108197 ms · 2026-08-02T17:23:06.645213+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of GroupRAG: Cognitively Inspired Group-Aware Retrieval and Reasoning via Knowledge-Driven Problem Structuring." pith.science (2026). https://pith.science/paper/X4KWOPZO

@misc{pith2026260326807,
  author       = {Pith},
  title        = {Pith review of: GroupRAG: Cognitively Inspired Group-Aware Retrieval and Reasoning via Knowledge-Driven Problem Structuring},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X4KWOPZO}},
  note         = {Machine review of arXiv:2603.26807}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The performance of language models is commonly limited by insufficient knowledge and constrained reasoning. Prior approaches such as Retrieval-Augmented Generation (RAG) and Chain-of-Thought (CoT) address these issues by incorporating external knowledge or enforcing linear reasoning chains, but often degrade in real-world settings. Inspired by cognitive science, which characterizes human problem solving as search over structured problem spaces rather than single inference chains, we argue that inadequate awareness of problem structure is a key overlooked limitation. We propose GroupRAG, a cognitively inspired, group-aware retrieval and reasoning framework based on knowledge-driven keypoint grouping. GroupRAG identifies latent structural groups within a problem and performs retrieval and reasoning from multiple conceptual starting points, enabling fine-grained interaction between the two processes. Experiments on MedQA (medical) and Bar Exam QA (legal) show that GroupRAG outperforms representative RAG- and CoT-based baselines. These results suggest that explicitly modeling problem structure, as inspired by human cognition, is a promising direction for robust retrieval-augmented reasoning.

Figures

Figures reproduced from arXiv: 2603.26807 by Jiangtao Gong, Xinyi Duan, Yuanrong Tang.

Figure 1
Figure 1. Figure 1: Reasoning Paradigms Comparison. Traditional CoT fol [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the GroupRAG reasoning process on a clinical case, featuring keypoint extraction, knowledge-driven grouping, local [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: An Abstract Overview of GroupRAG. 3.3 Retrieval-Augmented Generation in GroupRAG Retrieval-Augmented Generation (RAG) plays a crucial role in GroupRAG and is integrated into three key modules: Knowledge-Driven Grouping, Local Reasoning, and Answer Alignment. Rather than serving as a standalone retrieval com￾ponent, RAG is tightly coupled with the groupwise reason￾ing structure and operates at different gra… 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

29 extracted references · 12 linked inside Pith

  1. [1]

    Entity-based cross-document coreferencing using the vec- tor space model

    [Bagga and Baldwin, 1998] Amit Bagga and Breck Baldwin. Entity-based cross-document coreferencing using the vec- tor space model. InCOLING 1998 V olume 1: The 17th in- ternational conference on computational linguistics,

    1998

  2. [8]

    How the mind creates structure: Hierarchical learning of action sequences

    [Eckstein and Collins, 2021] Maria K Eckstein and Anne GE Collins. How the mind creates structure: Hierarchical learning of action sequences. InCogsci... annual confer- ence of the cognitive science society. cognitive science so- ciety (us). conference, volume 43, page 618,

    2021

  3. [10]

    Lightrag: Simple and fast retrieval-augmented generation.arXiv preprint arXiv:2410.05779,

    [Guoet al., 2024 ] Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. Lightrag: Simple and fast retrieval-augmented generation.arXiv preprint arXiv:2410.05779,

    Pith/arXiv arXiv 2024

  4. [11]

    People construct simplified mental representa- tions to plan.Nature, 606(7912):129–136,

    [Hoet al., 2022 ] Mark K Ho, David Abel, Carlos G Correa, Michael L Littman, Jonathan D Cohen, and Thomas L Griffiths. People construct simplified mental representa- tions to plan.Nature, 606(7912):129–136,

    2022

  5. [12]

    Dis- tilling step-by-step! outperforming larger language mod- els with less training data and smaller model sizes

    [Hsiehet al., 2023 ] Cheng-Yu Hsieh, Chun-Liang Li, Chih- Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alex Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Dis- tilling step-by-step! outperforming larger language mod- els with less training data and smaller model sizes. In Findings of the Association for Computational Linguistics: ACL 2023, pages 8003–8017,

    2023

  6. [14]

    What dis- ease does this patient have? a large-scale open domain question answering dataset from medical exams.Applied Sciences, 11(14):6421,

    [Jinet al., 2021 ] Di Jin, Eileen Pan, Nassim Oufattole, Wei- Hung Weng, Hanyi Fang, and Peter Szolovits. What dis- ease does this patient have? a large-scale open domain question answering dataset from medical exams.Applied Sciences, 11(14):6421,

    2021

  7. [15]

    Disentangling questions from query generation for task-adaptive retrieval.arXiv preprint arXiv:2409.16570,

    [Leeet al., 2024 ] Yoonsang Lee, Minsoo Kim, and Seung- won Hwang. Disentangling questions from query generation for task-adaptive retrieval.arXiv preprint arXiv:2409.16570,

    Pith/arXiv arXiv 2024

  8. [16]

    Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural infor- mation processing systems, 33:9459–9474,

    [Lewiset al., 2020 ] Patrick Lewis, Ethan Perez, Aleksan- dra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K ¨uttler, Mike Lewis, Wen-tau Yih, Tim Rockt¨aschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural infor- mation processing systems, 33:9459–9474,

    2020

  9. [18]

    Chain of thought with explicit evidence reasoning for few-shot re- lation extraction.arXiv preprint arXiv:2311.05922,

    [Maet al., 2023 ] Xilai Ma, Jing Li, and Min Zhang. Chain of thought with explicit evidence reasoning for few-shot re- lation extraction.arXiv preprint arXiv:2311.05922,

    Pith/arXiv arXiv 2023

  10. [19]

    Retrieval-augmented chain-of- thought in semi-structured domains.arXiv preprint arXiv:2310.14435,

    [Maviet al., 2023 ] Vaibhav Mavi, Abulhair Saparov, and Chen Zhao. Retrieval-augmented chain-of- thought in semi-structured domains.arXiv preprint arXiv:2310.14435,

    Pith/arXiv arXiv 2023

  11. [20]

    Si- mon.Human Problem Solving

    [Newell and Simon, 1972] Allen Newell and Herbert A. Si- mon.Human Problem Solving. Prentice-Hall, Englewood Cliffs, NJ,

    1972

  12. [22]

    Adaptive graph of thoughts: Test-time adaptive reasoning unify- ing chain, tree, and graph structures.arXiv preprint arXiv:2502.05078,

    [Pandeyet al., 2025 ] Tushar Pandey, Ara Ghukasyan, Ok- tay Goktas, and Santosh Kumar Radha. Adaptive graph of thoughts: Test-time adaptive reasoning unify- ing chain, tree, and graph structures.arXiv preprint arXiv:2502.05078,

    Pith/arXiv arXiv 2025

  13. [23]

    Interleaving retrieval with chain-of-thought reasoning for knowledge- intensive multi-step questions

    [Trivediet al., 2023 ] Harsh Trivedi, Niranjan Balasubrama- nian, Tushar Khot, and Ashish Sabharwal. Interleaving retrieval with chain-of-thought reasoning for knowledge- intensive multi-step questions. InProceedings of the 61st annual meeting of the association for computational lin- guistics (volume 1: long papers), pages 10014–10037,

    2023

  14. [24]

    Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171,

    [Wanget al., 2022 ] Xuezhi Wang, Jason Wei, Dale Schu- urmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171,

    Pith/arXiv arXiv 2022

  15. [25]

    Rat: Retrieval aug- mented thoughts elicit context-aware reasoning in long- horizon generation.arXiv preprint arXiv:2403.05313,

    [Wanget al., 2024 ] Zihao Wang, Anji Liu, Haowei Lin, Jiaqi Li, Xiaojian Ma, and Yitao Liang. Rat: Retrieval aug- mented thoughts elicit context-aware reasoning in long- horizon generation.arXiv preprint arXiv:2403.05313,

    Pith/arXiv arXiv 2024

  16. [26]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837,

    [Weiet al., 2022 ] Jason Wei, Xuezhi Wang, Dale Schuur- mans, 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

  17. [27]

    Tree of thoughts: Deliberate problem solving with large language models.Advances in neural informa- tion processing systems, 36:11809–11822,

    [Yaoet al., 2023 ] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.Advances in neural informa- tion processing systems, 36:11809–11822,

    2023

  18. [28]

    Reasoning over hierarchical question decomposition tree for explainable question answering

    [Zhanget al., 2023 ] Jiajie Zhang, Shulin Cao, Tingjian Zhang, Xin Lv, Juanzi Li, Lei Hou, Jiaxin Shi, and Qi Tian. Reasoning over hierarchical question decomposition tree for explainable question answering. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 14556–14570,

    2023

  19. [29]

    Least- to-most prompting enables complex reasoning in large lan- guage models.arXiv preprint arXiv:2205.10625, 2022

    [Zhouet al., 2022 ] Denny Zhou, Nathanael Sch¨arli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuur- mans, Claire Cui, Olivier Bousquet, Quoc Le, et al. Least- to-most prompting enables complex reasoning in large lan- guage models.arXiv preprint arXiv:2205.10625, 2022

    Pith/arXiv arXiv 2022

  20. [1972]

    GPT-4o system card

    [OpenAI, 2024] OpenAI. GPT-4o system card. https:// openai.com/index/gpt-4o-system-card/,

    2024

  21. [1998]

    Behavioral hierarchy: exploration and representation

    [Bartoet al., 2013 ] Andrew G Barto, George Konidaris, and Christopher Vigorito. Behavioral hierarchy: exploration and representation. InComputational and robotic models of the hierarchical organization of behavior, pages 13–46. Springer,

    2013

  22. [2012]

    The llama 3 herd of models.arXiv e-prints, pages arXiv–2407,

    [Dubeyet al., 2024 ] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models.arXiv e-prints, pages arXiv–2407,

    2024

  23. [2013]

    Forest-of-thought: Scaling test- time compute for enhancing llm reasoning.arXiv preprint arXiv:2412.09078,

    [Biet al., 2024 ] Zhenni Bi, Kai Han, Chuanjian Liu, Yehui Tang, and Yunhe Wang. Forest-of-thought: Scaling test- time compute for enhancing llm reasoning.arXiv preprint arXiv:2412.09078,

    Pith/arXiv arXiv 2024

  24. [2020]

    Rt: a retrieving and chain-of-thought frame- work for few-shot medical named entity recognition.Jour- nal of the American Medical Informatics Association, 31(9):1929–1938,

    [Liet al., 2024 ] Mingchen Li, Huixue Zhou, Han Yang, and Rui Zhang. Rt: a retrieving and chain-of-thought frame- work for few-shot medical named entity recognition.Jour- nal of the American Medical Informatics Association, 31(9):1929–1938,

    2024

  25. [2021]

    Minirag: Towards extremely sim- ple retrieval-augmented generation.arXiv preprint arXiv:2501.06713,

    [Fanet al., 2025 ] Tianyu Fan, Jingyuan Wang, Xubin Ren, and Chao Huang. Minirag: Towards extremely sim- ple retrieval-augmented generation.arXiv preprint arXiv:2501.06713,

    Pith/arXiv arXiv 2025

  26. [2022]

    To- wards reasoning era: A survey of long chain-of-thought for reasoning large language models.arXiv preprint arXiv:2503.09567,

    [Chenet al., 2025 ] Qiguang Chen, Libo Qin, Jinhao Liu, Dengyun Peng, Jiannan Guan, Peng Wang, Mengkang Hu, Yuhang Zhou, Te Gao, and Wanxiang Che. To- wards reasoning era: A survey of long chain-of-thought for reasoning large language models.arXiv preprint arXiv:2503.09567,

    Pith/arXiv arXiv 2025

  27. [2023]

    Atlas: Few-shot learning with retrieval augmented language models.Journal of Machine Learn- ing Research, 24(251):1–43,

    [Izacardet al., 2023 ] Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. Atlas: Few-shot learning with retrieval augmented language models.Journal of Machine Learn- ing Research, 24(251):1–43,

    2023

  28. [2024]

    Program of thoughts prompting: Disentangling computation from reason- ing for numerical reasoning tasks.arXiv preprint arXiv:2211.12588,

    [Chenet al., 2022 ] Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W Cohen. Program of thoughts prompting: Disentangling computation from reason- ing for numerical reasoning tasks.arXiv preprint arXiv:2211.12588,

    Pith/arXiv arXiv 2022

  29. [2025]

    Cushen and Jennifer Wiley

    [Cushen and Wiley, 2012] Patrick J. Cushen and Jennifer Wiley. Cues to solution, restructuring patterns, and reports of insight in creative problem solving.Consciousness and Cognition, 21(3):1166–1175,

    2012