Pith. sign in

REVIEW 4 major objections 6 minor 4 cited by

Adaptive Graph of Thoughts: Test-Time Adaptive Reasoning Unifying Chain, Tree, and Graph Structures

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read AGoT recursively decomposes queries into a DAG of subproblems and reports scientific-reasoning gains that match training-based distillation, using only test-time compute.

desk verdict AGoT is a real but over-claimed extension: the recursive graph framework shows some objective gains, yet the headline GPQA parity with distillation rests on a shuffled baseline whose hardness is asserted, not proven. read the letter →

arxiv 2502.05078 v1 pith:D2ZMI6OE submitted 2025-02-07 cs.AI cs.CL

Tushar Pandey , Ara Ghukasyan , Oktay Goktas , Santosh Kumar Radha This is my paper · ORCID
classification cs.AIcs.CL
keywords AdaptiveGraphofThoughtstest-timereasoningrecursivedecompositionDAGcomplexity-guidedexpansionLLMmulti-hopretrievalanswer-positionbias
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 introduces Adaptive Graph of Thoughts (AGoT), a test-time inference framework that recursively decomposes a query into subproblems arranged as a directed acyclic graph and expands only the subproblems a complexity check deems difficult. Its central claim is that this dynamic decomposition improves a frozen LLM's accuracy across reasoning, multi-hop retrieval, and puzzle benchmarks, with the largest reported gain being +46.2% on a shuffled graduate-level science question set. The authors compare that gain to improvements obtained by reinforcement-learning distillation, arguing that careful inference structuring can rival post-training methods without modifying the model. If the claim holds, it would make strong reasoning improvements accessible to anyone who can run an LLM, not only to those who can fine-tune or distill one.

What carries the argument

The load-bearing object is the recursive DAG of thoughts. Each node is indexed by a heritage, a sequence of layer and node positions that traces the node through any nested graphs, and holds a thought, a layer strategy, an answer, and possibly an attached nested graph. The complexity check $C$ is the gate: nodes flagged as complex spawn a fresh AGoT process on their own thought, while non-complex nodes are evaluated directly. This selective expansion is what lets AGoT allocate extra computation only where the LLM itself signals difficulty, and the final-thought check gives the framework a self-termination condition that avoids unnecessary branching.

What would settle it

Re-run AGoT and a direct input-output baseline on the same shuffled GPQA items with answer keys counterbalanced in both directions; if the AGoT gain over direct IO on the shuffled set is no larger than its gain on the unshuffled set, the +46.2% figure is not evidence that recursive decomposition matches distillation-grade reasoning gains.

Watch

Extended reading notes

Core claim

Stated on its own terms, the paper claims that a general, task-agnostic inference framework built from recursively nested graphs can lift a fixed model's accuracy more than fixed chain or tree prompting can, and that on the hardest scientific questions tested the lift approaches what model distillation achieves. The framework starts from an empty graph and generates thoughts layer by layer; each node is checked for complexity, complex nodes trigger nested AGoT processes, and the process ends when a final thought is recognized. Across the three task categories the paper reports AGoT as the best-scoring framework in every row of its comparison table, including a +400% relative gain on a hard arithmetic-puzzle set and consistent accuracy gains on multi-hop retrieval.

Load-bearing premise

The premise that makes the headline comparison work is that the shuffled GPQA subset is at least as hard as the unshuffled set, despite its much lower direct-IO baseline; if the shuffle mainly removes an answer-position cue, the +46.2% gain is not comparable to distillation gains.

Editorial extensions

If this is right

  • If AGoT's reported gains generalize, inference-framework design becomes a first-class performance lever alongside model scale and post-training.
  • Compute during a query is routed dynamically: easy subproblems are answered directly and hard ones receive nested decomposition, so users can trade tokens for accuracy.
  • Because the largest gains appear on the shuffled answer-position version of the science benchmark, benchmark reporting that ignores answer-order effects can understate or overstate a framework's true capability.
  • AGoT yields its biggest relative jumps on explorative tasks, suggesting combinatorial problems benefit most from recursive exploration, while the paper also notes absolute scores there remain low.

Reading between the lines

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

  • As an editorial extension, the intermediate thoughts AGoT generates could be collected into synthetic reasoning traces; if such traces train a student model well, test-time decomposition and distillation would be complementary routes rather than competing ones.
  • As an editorial extension, the near-zero unshuffled gain alongside the large shuffled gain suggests part of AGoT's headline effect is to neutralize a first-answer bias, and a direct shuffled-data comparison across all baselines would determine which part of the gain is general reasoning improvement.
  • As an editorial extension, the complexity check is currently an LLM judgment call; replacing it with a cheaper, less opaque difficulty estimator would make the framework's compute allocation more predictable and its gains easier to isolate.
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 / 6 minor

Summary. The paper introduces Adaptive Graph of Thoughts (AGoT), a test-time prompting framework that dynamically decomposes queries into a directed acyclic graph of thought nodes, using LLM-driven complexity checks to recursively spawn nested graphs. The authors provide a mathematical formalism, an illustrative example, and experiments with gpt-4o-mini and gpt-4o on GPQA, HotpotQA, MoreHopQA, HybridQA, mini-crosswords, and Game of 24. The central claim is that AGoT achieves up to +46.2% improvement on GPQA, comparable to gains from DeepSeek-R1 distillation, and outperforms IO, CoT, and AIoT across reasoning, retrieval, and explorative tasks.

Significance. If the headline result were fully supported, demonstrating that a training-free prompting framework can match computationally intensive RL-based distillation would be a substantial contribution. The unified graph-based formalism and the public implementation are useful assets, and the results on objective tasks such as Game of 24 and shuffled GPQA are promising. However, the central comparison to distillation is built on an unsupported equivalence between shuffled and unshuffled GPQA, and the retrieval gains rely on a self-scored metric. As it stands, the evidence does not justify the abstract's parity claim.

major comments (4)
  1. [Section 5.1.2, Table 1] The headline claim that AGoT achieves gains 'comparable to' DeepSeek-R1 distillation rests on the assertion that 'GPQA S is at least as hard as GPQAD.' This assertion is unsupported and is contradicted by the authors' own data: the IO baseline on GPQAS (37.4 for gpt-4o-mini, 39.4 for gpt-4o) is far below the GPQAD IO baseline (54.6), which the authors attribute to answer-position bias and 'saturation of position-specific patterns' (Section 5.1.1 and Discussion). A relative gain over a position-debiased, lower baseline is not comparable to distillation gains measured on the standard, position-inflated GPQAD, so the abstract's statement that AGoT is comparable to RL-based training is not established.
  2. [Section 4, implementation details] AGoT is evaluated with temperature=0.3, whereas IO, CoT, and AIoT use 'default settings' (for OpenAI models, default temperature is 1.0). Lower temperature reduces stochasticity and can improve accuracy on deterministic reasoning tasks. This confound means that the reported improvements of AGoT over the baselines may be partly due to decoding temperature rather than the framework itself; the authors should rerun all baselines at the same temperature or otherwise control for this variable.
  3. [Section 5.2, Table 3] The retrieval gains that appear in the abstract and Figure 1 are based on the LLM-assisted accuracy score (LAAS), which is computed by gpt-4o-mini, the same model that generates the answers. On objective string metrics, AGoT does not consistently improve: HybridQA EM drops from 55 (IO) to 45 (AGoT) and F1 from 68.4 to 67.8, while MoreHopQA EM and F1 are below AIoT. The LAAS metric is not validated against human judgments or an independent model, so the retrieval superiority claim is not robustly supported.
  4. [Section 5, Tables 2-3] Many of the comparisons are based on small samples (e.g., 100 HotpotQA questions, 20 Game of 24 problems) and are reported without confidence intervals, significance tests, or multiple runs. Several headline improvements are small absolute differences (e.g., GPQAD +0.9%, MoreHopQA LAAS 72 vs 70), so the claim that AGoT 'outperforms' existing methods across all tasks is not statistically established.
minor comments (6)
  1. [Algorithm 1, lines 13-22] The pseudocode is internally inconsistent. Lines 13-15 immediately output and return when a final thought is found, making the subsequent for loop unreachable; additionally, line 21 assigns Fh' but h' is the starting heritage, not the heritage of the final node. Since Section 4 states that Algorithm 1 provides 'an exact definition' of AGoT, this should be corrected.
  2. [Section 2.1, Eq. (1)] The set of heritages H is defined as the union of S^d over d=0..dmax, but some notation, such as σli in Eq. (5), is introduced without a precise domain; clarifying these definitions would improve reproducibility.
  3. [Section 5.2.1] The comparison of AGoT's 80% LAAS on a 100-question hard subset to Gao et al.'s 81% on the entire HotpotQA dataset is not apples-to-apples; the authors acknowledge the subset but the phrasing could mislead readers.
  4. [Figures 1 and 5] The y-axis labels do not identify which metric is being averaged for each category; specifying that retrieval uses LAAS and reasoning uses GPQA accuracy would aid interpretation.
  5. [Throughout] Typos such as 'an dynamic' (Abstract and Introduction), 'GQPA' (Section 3), 'form' for 'from' (Section 6), and 'came' for 'game' (Section 5.3.1) should be corrected.
  6. [Section 4] The lack of any sensitivity analysis for the free parameters dmax, lmax, and nmax limits the practical guidance, even though the authors disclaim optimality.

Circularity Check

1 steps flagged · score 4.0 of 10

Objective reasoning benchmarks are not circular, but the retrieval results are graded by the same model being tested, and the distillation-parity headline relies on an unproven GPQA difficulty assumption.

  1. other [Section 5.2, Table 3 (LLM-assisted accuracy score, LAAS)]
    "In addition to exact match (EM) and F1 scores, we report for this category an LLM-assisted accuracy score (LAAS) obtained by requesting a binary response that indicates the semantic equivalence of two input strings. This metric is implemented under the reasonable assumption that gpt-4o-mini is capable of highly accurate text classification in this setting."

    LAAS is the metric that drives the paper's retrieval-superiority claims, yet the judge is gpt-4o-mini, the same model whose outputs AGoT organizes. By construction the evaluator is therefore part of the system being evaluated: AGoT's final answer is produced via Eval(th', Gh) in Eq. (12), and LAAS is another query to the same model asking whether that output is semantically equivalent to the reference. The paper asserts, rather than validates, that gpt-4o-mini is a reliable classifier in this setting. Table 3 makes the self-referentiality load-bearing: AGoT is best on LAAS, but EM is most often best for IO and F1 for AIoT, so the framework's retrieval advantage reduces to the model judging its own outputs favorably.

full rationale

The headline reasoning results are not circular: GPQA multiple-choice accuracy and Game of 24 are objective ground-truth tasks, and the AGoT graph construction (Eqs. 8-13, Algorithm 1) is an explicit mechanism with no fitted parameter that is later renamed as a prediction. The paper also self-cites Radha and Goktas (2025) to justify answer-position shuffling, but its own Table 1 independently shows a large IO drop under shuffling, so that self-citation is not by itself load-bearing. The main non-circular weakness is external: Section 5.1.2 asserts 'GPQAS is at least as hard as GPQAD' to compare +46.2% on the shuffled set with distillation gains on unshuffled sets, but this is an unproven benchmark-difficulty assumption, not a circular derivation. The one genuine self-referential element is LAAS, where the same gpt-4o-mini model judges semantic equivalence of answers it helped produce. Because objective benchmarks anchor the central claim, the overall circularity is partial rather than total.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The framework rests on several domain assumptions about LLM reliability, plus three hyperparameters chosen by the authors. The most fragile assumption is that the shuffled GPQA subset is at least as hard as the standard set, a claim stated without evidence and used to validate the RL comparison. No new physical or mathematical entities are introduced; heritages, nested graphs, and strategies are software constructs.

free parameters (4)
  • dmax (maximum recursion depth) = 1
    Set for all experiments; authors state it was chosen for satisfactory computation time, not optimality (Sections 4 and 6.1).
  • lmax (maximum number of layers) = 3
    Set for all experiments; authors state it was chosen for satisfactory computation time, not optimality (Sections 4 and 6.1).
  • nmax (maximum nodes per layer) = 3
    Set for all experiments; authors state it was chosen for satisfactory computation time, not optimality (Sections 4 and 6.1).
  • generation temperature for AGoT = 0.3
    Used for AGoT generation while other frameworks use default settings, which may affect relative performance (Section 4).
assumptions (5)
  • domain assumption LLM actions T, C, Eval, and Phi reliably generate thoughts, classify complexity, and evaluate answers.
    Algorithm 1 and Section 4 assume these mappings are implemented by LLM agents with sufficient accuracy for recursion control and final evaluation.
  • domain assumption gpt-4o-mini can accurately classify semantic equivalence of two strings for the LAAS metric.
    Section 5.2 states this as a 'reasonable assumption' and uses LAAS to report retrieval gains.
  • ad hoc to paper GPQAS is at least as hard as GPQAD.
    Section 5.1.2 asserts this without evidence to justify comparing AGoT's shuffled-subset gain to DeepSeek-R1 distillation on standard sets.
  • domain assumption Thoughts are accessible, combinable units of information.
    Section 1.2 defines thoughts as graph nodes; the entire graph representation depends on this assumption.
  • standard math Acyclic graphs guarantee finite thought processes.
    Section 1.2 uses the DAG property to ensure termination; this is a standard graph theory fact.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Graph of Thoughts: Test-Time Adaptive Reasoning Unifying Chain, Tree, and Graph Structures." pith.science (2026). https://pith.science/paper/D2ZMI6OE

@misc{pith2026250205078,
  author       = {Pith},
  title        = {Pith review of: Adaptive Graph of Thoughts: Test-Time Adaptive Reasoning Unifying Chain, Tree, and Graph Structures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D2ZMI6OE}},
  note         = {Machine review of arXiv:2502.05078}
}
read the original abstract

Large Language Models (LLMs) have demonstrated impressive reasoning capabilities, yet their performance is highly dependent on the prompting strategy and model scale. While reinforcement learning and fine-tuning have been deployed to boost reasoning, these approaches incur substantial computational and data overhead. In this work, we introduce Adaptive Graph of Thoughts (AGoT), a dynamic, graph-based inference framework that enhances LLM reasoning solely at test time. Rather than relying on fixed-step methods like Chain of Thought (CoT) or Tree of Thoughts (ToT), AGoT recursively decomposes complex queries into structured subproblems, forming an dynamic directed acyclic graph (DAG) of interdependent reasoning steps. By selectively expanding only those subproblems that require further analysis, AGoT unifies the strengths of chain, tree, and graph paradigms into a cohesive framework that allocates computation where it is most needed. We validate our approach on diverse benchmarks spanning multi-hop retrieval, scientific reasoning, and mathematical problem-solving, achieving up to 46.2% improvement on scientific reasoning tasks (GPQA) - comparable to gains achieved through computationally intensive reinforcement learning approaches and outperforming state-of-the-art iterative approaches. These results suggest that dynamic decomposition and structured recursion offer a scalable, cost-effective alternative to post-training modifications, paving the way for more robust, general-purpose reasoning in LLMs.

Figures

Figures reproduced from arXiv: 2502.05078 by the authors.

Figure 1
Figure 1. Performance comparison of reasoning frameworks [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Architectural comparison of inference frameworks showing structural evolution from linear (CoT) to more complex [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Schematic showing the layer-wise evolution of [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Diagram illustrating a final AGoT state after evaluation of a technical problem from GPQA. Grey labels on the [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Average absolute difference in performance score [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Total nodes and the percentage of complex nodes [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

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

  1. Framework of Thoughts: A Foundation Framework for Dynamic and Optimized Reasoning based on Chains, Trees, and Graphs

    cs.AI 2026-02 conditional novelty 6.0 of 10

    A framework that wraps chain/tree/graph reasoning schemes with parallel execution, caching, and hyperparameter/prompt optimization, reporting 1.9–35× speedups and modest accuracy gains.

  2. MARCO: Meta-Reflection with Cross-Referencing for Code Reasoning

    cs.CL 2025-05 conditional novelty 6.0 of 10

    MARCO combines cross-problem knowledge accumulation with cross-agent lesson sharing to improve LLM code reasoning at inference time.

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

    cs.IR 2026-03 conditional novelty 5.0 of 10

    Structuring questions into knowledge-driven keypoint groups before retrieval and reasoning improves small-model accuracy on MedQA.

  4. Large Language Models for Planning: A Comprehensive and Systematic Survey

    cs.AI 2025-05 conditional novelty 3.0 of 10

    A structured survey of LLM planning methods, benchmarks, and interpretability work, organized around a three-way taxonomy.

Reference graph

Works this paper leans on

35 extracted references · 7 canonical work pages · cited by 4 Pith papers

  1. [1]

    multi-agent-llm-0.1.3, 2024

    Github AgnostiqHQ. multi-agent-llm-0.1.3, 2024. URL https://github.com/AgnostiqHQ/multi-agent-llm. Implementation of the AIoT/GIoT and AGoT methods for multi-agent LLM architectures

    work page 2024

  2. [2]

    Llm-generated black-box explanations can be adversarially helpful, 2024

    Rohan Ajwani, Shashidhar Reddy Javaji, Frank Rudzicz, and Zining Zhu. Llm-generated black-box explanations can be adversarially helpful, 2024. URL https://arxiv.org/abs/2405.06800

    arXiv 2024

  3. [3]

    Palm 2 technical report

    Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. URL https://arxiv.org/abs/2305.10403

    arXiv 2023

  4. [4]

    Graph of thoughts: Solving elaborate problems with large language models

    Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17682--17690, 2024. URL https...

    doi:10.1609/aaai.v38i16.29720 2024

  5. [5]

    Hybridqa: A dataset of multi-hop question answering over tabular and textual data, 2021

    Wenhu Chen, Hanwen Zha, Zhiyu Chen, Wenhan Xiong, Hong Wang, and William Wang. Hybridqa: A dataset of multi-hop question answering over tabular and textual data, 2021. URL https://arxiv.org/abs/2004.07347

    arXiv 2021

  6. [6]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, and Peiyi Wang et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025. URL https://arxiv.org/abs/2501.12948

    arXiv 2025

  7. [7]

    The llama 3 herd of models

    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 preprint arXiv:2407.21783, 2024. URL https://arxiv.org/abs/2407.21783

    arXiv 2024

  8. [8]

    Meta reasoning for large language models, 2024

    Peizhong Gao, Ao Xie, Shaoguang Mao, Wenshan Wu, Yan Xia, Haipeng Mi, and Furu Wei. Meta reasoning for large language models, 2024. URL https://arxiv.org/abs/2406.11698

    arXiv 2024

Show all 35 references
  1. [10]

    Understanding the effects of iterative prompting on truthfulness, 2024

    Satyapriya Krishna, Chirag Agarwal, and Himabindu Lakkaraju. Understanding the effects of iterative prompting on truthfulness, 2024. URL https://arxiv.org/abs/2402.06625

    2024 arXiv

  2. [11]

    More agents is all you need

    Junyou Li, Qin Zhang, Yangbin Yu, Qiang Fu, and Deheng Ye. More agents is all you need. arXiv preprint arXiv:2402.05120, 2024. URL https://arxiv.org/abs/2402.05120

    2024 arXiv

  3. [12]

    Search-o1: Agentic search-enhanced large reasoning models, 2025

    Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. Search-o1: Agentic search-enhanced large reasoning models, 2025. URL https://arxiv.org/abs/2501.05366

    2025 arXiv

  4. [13]

    Dgot: Dynamic graph of thoughts for scientific abstract generation, 2024

    Xinyu Ning, Yutong Zhao, Yitong Liu, and Hongwen Yang. Dgot: Dynamic graph of thoughts for scientific abstract generation, 2024. URL https://arxiv.org/abs/2403.17491

    2024 arXiv

  5. [14]

    Gpt-4 technical report

    OpenAI. Gpt-4 technical report. OpenAI Technical Report, 2023. URL https://cdn.openai.com/papers/gpt-4.pdf. Accessed: 2023-09-13

    2023

  6. [15]

    Gpt-4o system card, 2024

    OpenAI. Gpt-4o system card, 2024. URL https://arxiv.org/abs/2410.21276

    2024 arXiv

  7. [16]

    Introducing openai o1-preview

    OpenAI . Introducing openai o1-preview. https://openai.com/index/introducing-openai-o1-preview/, September 2024. URL https://openai.com/index/introducing-openai-o1-preview/. Accessed: 2024-09-17

    2024

  8. [17]

    Adapt: As-needed decomposition and planning with language models, 2024

    Archiki Prasad, Alexander Koller, Mareike Hartmann, Peter Clark, Ashish Sabharwal, Mohit Bansal, and Tushar Khot. Adapt: As-needed decomposition and planning with language models, 2024. URL https://arxiv.org/abs/2311.05772

    2024 arXiv

  9. [18]

    Language models are unsupervised multitask learners

    Alec Radford, Rewon Child Jeffrey Wu, Dario Amodei David Luan, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI Technical Report, 2024. URL https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf

    2024

  10. [19]

    Composite learning units: Generalized learning beyond parameter updates to transform llms into adaptive reasoners

    Santosh Kumar Radha and Oktay Goktas. Composite learning units: Generalized learning beyond parameter updates to transform llms into adaptive reasoners. arXiv preprint arXiv:2410.08037, 2024

    2024 arXiv

  11. [20]

    On the reasoning capacity of ai models and how to quantify it

    Santosh Kumar Radha and Oktay Goktas. On the reasoning capacity of ai models and how to quantify it. arXiv preprint arXiv:2501.13833, 2025

    2025 arXiv

  12. [21]

    Iteration of thought: Leveraging inner dialogue for autonomous large language model reasoning, 2024

    Santosh Kumar Radha, Yasamin Nouri Jelyani, Ara Ghukasyan, and Oktay Goktas. Iteration of thought: Leveraging inner dialogue for autonomous large language model reasoning, 2024. URL https://arxiv.org/abs/2409.12618

    2024 arXiv

  13. [22]

    Gpqa: A graduate-level google-proof q&a benchmark

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. Gpqa: A graduate-level google-proof q&a benchmark. arXiv preprint arXiv:2311.12022, 2023. URL https://arxiv.org/abs/2311.12022

    2023 arXiv

  14. [23]

    Morehopqa: More than multi-hop reasoning, 2024

    Julian Schnitzler, Xanh Ho, Jiahao Huang, Florian Boudin, Saku Sugawara, and Akiko Aizawa. Morehopqa: More than multi-hop reasoning, 2024. URL https://arxiv.org/abs/2406.13397

    2024 arXiv

  15. [24]

    Hsu, Richard Antonello, Shailee Jain, Alexander G

    Chandan Singh, Aliyah R. Hsu, Richard Antonello, Shailee Jain, Alexander G. Huth, Bin Yu, and Jianfeng Gao. Explaining black box text modules in natural language with language models, 2023. URL https://arxiv.org/abs/2305.09863

    2023 arXiv

  16. [25]

    BB ox-adapter: Lightweight adapting for black-box large language models

    Haotian Sun, Yuchen Zhuang, Wei Wei, Chao Zhang, and Bo Dai. BB ox-adapter: Lightweight adapting for black-box large language models. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors, Proceedi...

    2024

  17. [26]

    Gemini: a family of highly capable multimodal models

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. URL https://arxiv.org/abs/2312.11805

    2023 arXiv

  18. [27]

    M Towhidul Islam Tonmoy, S M Mehedi Zaman, Vinija Jain, Anku Rani, Vipula Rawte, Aman Chadha, and Amitava Das

    S. M Towhidul Islam Tonmoy, S M Mehedi Zaman, Vinija Jain, Anku Rani, Vipula Rawte, Aman Chadha, and Amitava Das. A comprehensive survey of hallucination mitigation techniques in large language models, 2024. URL https://arxiv.org/abs/2401.01313

    2024 arXiv

  19. [28]

    Reinforcement learning enhanced llms: A survey, 2024 a

    Shuhe Wang, Shengyu Zhang, Jie Zhang, Runyi Hu, Xiaoya Li, Tianwei Zhang, Jiwei Li, Fei Wu, Guoyin Wang, and Eduard Hovy. Reinforcement learning enhanced llms: A survey, 2024 a . URL https://arxiv.org/abs/2412.10400

    2024 arXiv

  20. [29]

    Self-consistency improves chain of thought reasoning in language models

    Xuezhi Wang, Jason Wei, Dale Schuurmans, 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, 2022

    2022 arXiv

  21. [30]

    Tdag: A multi-agent framework based on dynamic task decomposition and agent generation, 2024 b

    Yaoxiang Wang, Zhiyong Wu, Junfeng Yao, and Jinsong Su. Tdag: A multi-agent framework based on dynamic task decomposition and agent generation, 2024 b . URL https://arxiv.org/abs/2402.10178

    2024 arXiv

  22. [31]

    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: 0 24824--24837, 2022

    2022

  23. [32]

    Qwen2.5 technical report

    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. URL https://arxiv.org/pdf/2412.15115

    2024 arXiv

  24. [33]

    Cohen, Ruslan Salakhutdinov, and Christopher D

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering, 2018. URL https://arxiv.org/abs/1809.09600

    2018 arXiv

  25. [34]

    Tree of thoughts: Deliberate problem solving with large language models

    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 Information Processing Systems, 36, 2024. URL https://arxiv.org/abs/2305.10601

    2024 arXiv

  26. [35]

    Redel: A toolkit for llm-powered recursive multi-agent systems, 2024

    Andrew Zhu, Liam Dugan, and Chris Callison-Burch. Redel: A toolkit for llm-powered recursive multi-agent systems, 2024. URL https://arxiv.org/abs/2408.02248

    2024 arXiv

  27. [36]

    Hydra: Model factorization framework for black-box llm personalization, 2024

    Yuchen Zhuang, Haotian Sun, Yue Yu, Rushi Qiang, Qifan Wang, Chao Zhang, and Bo Dai. Hydra: Model factorization framework for black-box llm personalization, 2024. URL https://arxiv.org/abs/2406.02888

    2024 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.