Pith. sign in

REVIEW 4 major objections 5 minor 40 references

Graph-of-Causal Evolution: Challenging Chain-of-Model for Reasoning

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper claims that replacing chain-of-model's linear subchains with a sparse, differentiable causal graph over token latents improves long-range causal reasoning and self-evolution in transformers, with consistent gains reported on…

desk verdict The headline claim—GoCE surpasses chain-of-model—is never tested, and the missing artifacts and baseline mismatch make the reported gains uninterpretable. read the letter →

arxiv 2506.07501 v1 pith:MKPHJVHG submitted 2025-06-09 cs.LG cs.AIcs.CL

classification cs.LGcs.AIcs.CL
keywords causalreasoningchain-of-modelgraphofevolutiondifferentiablecausal-maskedattentionmixtureexpertsself-evolutioninterventionconsistencyloss
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 targets a weakness in chain-of-model reasoning, where each subchain reads only the previous subchain and a causal mask blocks global context flow, so long-range dependencies can get lost. It proposes Graph-of-Causal Evolution (GoCE), a transformer variant that builds a sparse, differentiable causal graph over token latents and pushes causal constraints through attention and expert routing. An intervention-consistency loss and a self-evolution gate are added so the model can check and update itself as it reasons. Running this variant inside three code-executing LLM sandboxes, the paper reports consistent gains over the base models on CLUTRR, CLadder, EX-FEVER, and CausalQA across accuracy, calibration, and robustness metrics. The intended significance is a design principle: replace linear reasoning chains with an evolving causal graph to make long-range causal capture and self-improvement work together.

What carries the argument

The load-bearing object is a sparse causal adjacency matrix built from seed token latents: a small MLP scores each pair of latent nodes, a hard-concrete sampler turns those scores into discrete edges, and a topological sort keeps the resulting graph acyclic. That matrix is reused as a mask in causal-masked attention, as a vicinity mask in causal mixture-of-experts routing, and as the intervention target set in the counterfactual module. The second mechanism is the self-evolution gate, which mutates the current best parameters and accepts the mutation only if task reward and intervention-consistency loss both improve.

What would settle it

Run the released GoCE code on CLadder with an open-weight transformer and check whether the Rung-2 and Rung-3 accuracy plus the intervention-consistency metrics actually improve over the base model; if the execution logs show no parameter updates, the claimed self-evolution loop is not real.

Watch

Extended reading notes

Core claim

The paper's central claim is that a transformer organized by an evolving sparse causal graph instead of a linear chain captures long-range causal dependencies better and can self-evolve. GoCE turns token latents into a causal adjacency matrix, applies that matrix as masks in multi-head attention and in a causal mixture-of-experts router, and trains with an intervention-consistency loss that compares normal and intervened forward passes. A self-evolution gate then accepts or rejects mutated parameters based on held-out task reward and causal-consistency loss. On the four benchmarks, the paper reports that the GoCE variant improved the measured task metrics and reduced error and calibration losses relative to the three baseline models.

Load-bearing premise

The findings stand or fall on whether the uploaded GoCE code actually ran and updated the model parameters in the three LLM sandboxes, and on whether the open transformer variant used is close enough to the closed commercial baselines for the comparisons to be valid.

Editorial extensions

If this is right

  • Long-range causal dependencies can be enforced inside a transformer by a differentiable sparse graph rather than by a linear chain, without giving up end-to-end training.
  • A single causal graph can gate both attention heads and expert routing, coupling causal structure with parameter sparsity in one mechanism.
  • Intervention consistency loss supplies a training signal that penalizes representations that shift too much when one latent node is clamped.
  • A self-evolution gate can combine held-out task reward with causal-consistency loss to accept parameter mutations, embedding an adaptive loop in the architecture.
  • On the tested benchmarks, the reported gains cover both task quality (accuracy, exact match, ROC-AUC) and calibration and robustness metrics (ECE, NLL, BS, IRS, MR, CS).

Reading between the lines

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

  • If the sandbox results hold up in real training runs, the same graph-mask recipe should be tested head-to-head against a true chain-of-model implementation, which would isolate whether the graph structure is the source of the gains.
  • The paper does not ablate the intervention-consistency loss; a natural extension is to keep only the causal masks and see whether the reported robustness improvements persist.
  • The self-evolution gate is essentially a local evolutionary search over parameters, so a direct test is whether the acceptance rule improves perplexity on held-out language modeling as well as causal-reasoning scores.
  • The framework predicts that the benefit of a graph over a chain should grow with task depth, so the accuracy gap on CLadder Rung-3 should be larger than on Rung-2; that gap is measurable.
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 / 5 minor

Summary. The paper proposes Graph-of-Causal Evolution (GoCE), a transformer modification intended to replace the chain-of-model (CoM) paradigm for reasoning. The method maps token representations to a differentiable sparse causal adjacency matrix, applies causal-masked attention and causal-MoE layers, adds an intervention-consistency loss, and uses a self-evolution gate to update parameters. The authors report experiments run inside LLM chat sandboxes (Claude Sonnet 4, o4-mini-high, DeepSeek R1) on CLUTRR, CLadder, EX-FEVER, and CausalQA, comparing a GoCE-modified transformer against the stock LLMs. They claim GoCE strengthens long-range causal dependency capture, improves self-evolution, and 'surpasses the design of CoM.'

Significance. If the claimed results were valid, GoCE would be a substantial contribution to causal reasoning in LLMs, offering a concrete way to inject sparse causal structure and self-adaptation into transformer layers without retraining from scratch. The paper also has some credible design elements: the use of hard-concrete sampling for differentiable sparsity, the causal-masked attention idea, and the explicit acknowledgment of limitations in Section 5. However, the significance cannot be assessed because the experimental evaluation does not test the paper's headline claim against CoM, the implementation is prompt-driven execution inside chat sandboxes with no released code or logs, the reported baseline mismatch is admitted in Section 5, and the central equations in Sections 3.1-3.5 are garbled. The paper currently provides no verifiable evidence for its central claims.

major comments (4)
  1. [Section 4, Tables 1 and 2] The paper's central claim, stated in the abstract and Section 6, is that GoCE 'surpasses the design of CoM.' However, Section 4 contains no CoM baseline: Table 1 reports the GoCE-modified transformer and Table 2 is explicitly labeled 'original baseline models' (stock Claude Sonnet 4, o4-mini-high, DeepSeek R1). The reported deltas are therefore GoCE-vs-vanilla, not GoCE-vs-CoM. Without a CoM implementation or at least a clear specification and reproducible comparison, the headline claim is untested.
  2. [Sections 4.3 and 5] The experimental pipeline is prompt-driven execution inside LLM chat sandboxes: Section 4.3 states that the researcher 'used prompt commands from pre-processing to the final output' after uploading code and datasets, and Section 5 concedes that the GoCE transformer 'may still have a gap with the original baseline model in basic performance and resource scheduling.' No code, logs, seeds, or error bars are provided. Consequently, even if every table entry is taken at face value, the reported improvements are uninterpretable because the baseline and the modified model are not comparable implementations.
  3. [Section 3.5, fitness function F(theta)] The self-evolution gate's acceptance criterion is F(θ) = R(θ) − αLcf(θ) − βS(θ), where R(θ) is described as 'held-out task reward.' If the same held-out reward metric is also used in the reported evaluations (Section 4, Tables 1 and 2), then parameters are selected on the evaluation metric, which introduces selection bias and makes the final numbers optimistic. The paper must clarify whether R(θ) uses a separate validation split and whether the reported tables come from an untouched test split; if not, the experimental comparisons are circular.
  4. [Sections 3.1-3.4, Equations] The formal definitions of the core algorithm are garbled or incomplete: the hard-concrete edge-sampler equation in Section 3.1 is unreadable, the CSAIL algorithm block in Section 3.2.1 contains undefined symbols and broken equations (e.g., the KL term, the top-k notation, and the meaning of τcf are not clear), the C-MoE algorithm in Section 3.3 mixes notation inconsistently (h_t, g_η, E_vt, FNN_e), and the intervention-loss equation in Section 3.4 is syntactically malformed. Because these equations define the method, a reader cannot verify the claims independently.
minor comments (5)
  1. [Section 4] The subsection numbering is inconsistent: after Section 4 begins, the following subsections are numbered 3.1, 3.2, and 3.3 instead of 4.1, 4.2, and 4.3.
  2. [Section 4, Result & Discussion, first paragraph] The text describing the CausalQA metrics appears under the EX-FEVER sentence: 'the metrics for measuring EX-FEVER are: Interventional robustness score (IRS), mutational robustness (MR), precision in estimation of heterogeneous effects (PEHE)...' These are the metrics for CausalQA, not EX-FEVER, and the EX-FEVER metric list is duplicated.
  3. [Abstract and text] The dataset name is spelled inconsistently as 'CLADDER' in the abstract and 'CLadder' elsewhere; the reference [30] also uses 'CLADDER.' Please standardize to the dataset's official spelling.
  4. [Section 3.1, first sentence] The sentence 'The components of are closely connected to each other' is missing a noun after 'of' and should be reworded.
  5. [Figure 1 reference] Figure 1 is referenced in the introduction and Section 3, but the text does not explain the figure's content or how the example is solved by CoM versus GoCE; please add a caption explanation or refer to it more substantively.

Circularity Check

1 steps flagged · score 6.0 of 10

The self-evolution gate selects parameters by maximizing the same held-out task reward that is then reported as evidence, making the claimed improvements partly circular; the absent CoM baseline is a validity gap, not circularity.

  1. fitted input called prediction [Section 3.5 (Self-Evolution Gate), F(θ) and CAER acceptance equations; results reported in Section 4, Tables 1-2]
    "The task reward is responsible for quantifying the performance of θ′ on the target task, such as accuracy or prediction score, with positive feedback as the optimization goal; ... where R(θ) represents held-out task reward ... If θ′ is better than or acceptable than the existing best parameters in terms of both performance and causal consistency, the CAER gate triggers an acceptance signal and updates θ′ to the new θbest."

    The acceptance criterion for mutating parameters is an aggregate F that includes R(θ), and R(θ) is explicitly the held-out target-task reward, i.e., accuracy/NLL/etc. — the same quantities reported in Table 1 as evidence that GoCE improves causal reasoning. Thus θbest is fitted to the evaluation metric before the metric is presented as an experimental finding. Without a described separate test split, the reported GoCE-vs-baseline deltas are partly selection artifacts: improving R(θ) on the held-out set is expected by construction of the gate, so it cannot by itself prove that GoCE strengthens long-range causal dependencies. This is pattern 2: a fitted selection objective renamed as an empirical prediction.

full rationale

The paper's derivation chain is not equation-level circular in most of its architecture: the causal graph construction, masked attention, causal-MoE, and intervention module are defined from the model's own latents and are trained objectives rather than disguised restatements of the conclusion. The one clear circular step is the self-evolution gate: the CAER acceptance criterion uses F(θ) = -Lcf(θ) + αR(θ) + βS(θ) with R the held-out task reward, and the same reward-type metrics (Accuracy@1, NLL, ECE, etc.) are then reported in Table 1 as the experimental demonstration. Because the parameter vector is selected to improve R on the held-out set, the subsequent table entries are partly the result of fitting to the evaluation criterion, not independent predictions. I did not count the self-referential intervention loss as a separate circularity step: it is an internal consistency objective and cannot provide external causal validation, but the paper's empirical claim rests on the task-reward metrics, not on the KL term alone. I also did not count the absence of a chain-of-model baseline as circularity: failing to compare against CoM undermines the headline 'surpasses CoM' claim as a validity matter, but it is not a reduction of the derivation to its inputs. Overall, the central stated improvement is partially forced by construction, so the circularity score is 6 rather than 0-2.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The ledger shows that the paper's quantitative claims rest on several unreported hyperparameters, an unvalidated causal-structure assumption, and the integrity of sandbox execution. No new physical or ontological entity is introduced; the causal graph is an architectural construction.

free parameters (6)
  • hard-concrete temperature tau
    Controls the differentiable discrete edge sampling in the latent-causal graph builder in Section 3.1; no value or schedule is reported.
  • counterfactual temperature tau_cf
    Used in Section 3.4 to sharpen the post-intervention softmax; value is not reported.
  • l0 sparsity penalty weight lambda_L0
    Weights the sparsity penalty in the CSAIL objective in Section 3.2.1, described only as a small hyperparameter.
  • top-k expert count k = 1 (often)
    Number of experts selected by the masked top-k router in the causal-MoE in Section 3.3; when set to 1, expert routing reduces to selecting a single expert.
  • intervention loss weight lambda_delta
    Combines the KL term and the absolute-difference metric in the intervention loss in Section 3.4; value is not reported.
  • mutation noise and acceptance temperature T, gamma, gamma_epsilon
    Govern the self-evolution gate's random perturbations and stochastic acceptance in Section 3.5; no values or annealing schedules are reported.
assumptions (3)
  • ad hoc to paper The causal adjacency matrix learned by TinyMLP and hard-concrete sampling, with acyclicity enforced by considering only j < i, faithfully represents the causal structure needed for reasoning.
    Invoked in Section 3.1 without external validation; the paper's performance claims depend on this structure being informative rather than arbitrary.
  • domain assumption KL divergence between standard and sharpened post-intervention output distributions is a valid measure of causal consistency.
    Used in Sections 3.2.1 and 3.4; no justification or prior evidence is given that this loss tracks true causal structure rather than temperature sensitivity.
  • domain assumption Commercial LLM sandboxes execute the uploaded Python code and weight files faithfully, producing genuine training runs.
    Section 4.3 describes prompt-driven execution in Claude Sonnet 4, o4-mini-high, and DeepSeek R1 sandboxes; the result tables depend entirely on this assumption, which the author partially concedes in Section 5.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Graph-of-Causal Evolution: Challenging Chain-of-Model for Reasoning." pith.science (2026). https://pith.science/paper/MKPHJVHG

@misc{pith2026250607501,
  author       = {Pith},
  title        = {Pith review of: Graph-of-Causal Evolution: Challenging Chain-of-Model for Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MKPHJVHG}},
  note         = {Machine review of arXiv:2506.07501}
}
read the original abstract

In view of the problem that each subchain in the chain-of-model (CoM) relies only on the information of the previous subchain and may lose long-range dependencies due to the causal mask blocking the global context flow between multi-level subchains, this work proposes a graph of causal evolution (GoCE). Its core principle is to map the implicit token representation into a differentiable and sparse causal adjacency matrix, then permeate causal constraints through each layer of calculation using causal-masked attention and causal-MoE. By combining intervention consistency loss test and self-evolution gate, the dynamic balance between causal structure learning and adaptive updating of transformer architecture is realized. The researcher built experimental environments in sandboxes built with Claude Sonnet 4, o4-mini-high, and DeepSeek R1 respectively with the transformer variant architecture introduced in GoCE. It is evaluated on publicly available datasets including CLUTRR, CLADDER, EX-FEVER, and CausalQA and compared with the baseline LLMs. The finding proves that GoCE strengthens the transformer's ability to capture long-range causal dependencies, while the ability to self-evolve is improved. It not only surpasses the design of CoM in terms of design principles, but also provides experience for future research on causal learning and continuous adaptive improvement.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 27 canonical work pages

  1. [1]

    Graph-of-Causal Evolution (GoCE)

    Introduction As the technology in the field of reasoning matures, large language models (LLMs) have evolved from simply predicting the next word to a complex system that deeply combines attention heads (Tikhonov & Ryabinin, 2021; Zhang et al., 2023; Chen et al., 2024). The self-attention mechanism calculates the correlation strength between different repr...

  2. [2]

    Chain-of-Model (CoM)

    Related Work The previously mentioned Song et al. (2025) proposed "Chain-of-Model (CoM)" as a learning paradigm that splits the hidden vector into multiple "Chain-of-Representation (CoR)" sub-vectors. It ensures that the output of the i-th chain depends only on the previous i chains of the input hidden representation through the "Chain-of-Layer (CoL)" mec...

  3. [3]

    KL+(Δmetric)

    Graph-of-Causal Evolution As a self-evolving causal reasoning paradigm, Graph-of Causal Evolution (GoCE) aims to ensure causal dependencies in the reasoning process and continuously update the structure in multiple rounds of reasoning. Compared with CoM, it dynamically builds and iteratively evolves the causal graph through text prompts, so that LLMs can ...

  4. [4]

    Experiments The positivist paradigm is adopted because CoCE needs to be tested through quantifiable and repeatable experiments to accurately examine the effects and causal mechanisms of LLMs after integration into transformers (Park et al., 2020). Positivism emphasizes that theoretical claims must be based on observable facts and objective data, and hypot...

  5. [5]

    The following table 1 and table 2 show the performance of the introduced GoCE and baseline models respectively

    Result & Discussion After the experiment was executed, the researcher evaluated the corresponding metrics for the characteristics of each dataset by calculating in corresponding sandboxs. The following table 1 and table 2 show the performance of the introduced GoCE and baseline models respectively. The metrics for measuring CLUTRR are: Accuracy@k, brier s...

  6. [6]

    Limitation & Future Research The transformer architecture used in the experiment is designed based on open source codes such as academic papers, GitHub, and Hugging Face, and it is difficult to keep up with the timely internal detail updates and optimization strategies of the cutting-edge Claude Sonnet 4, o4-mini-high, and DeepSeek R1. In particular, the ...

  7. [7]

    Win, win, double! Far, far, ahead!

    Conclusion The GoCE proposed in this work shows a design that is more advantageous than CoM. The latent representation of tokens is transformed into a differentiable and sparse causal neighbor matrix through a latent-causal graph. Then, causal-masked attention, causal-MoE, intervention and counterfactual modules are combined in the transformer architectur...

  8. [8]

    Bondarenko, M

    A. Bondarenko, M. Wolska, S. Heindorf, L. Blübaum, A. C. N. Ngomo, B. Stein, and M. Potthast. CausalQA: A benchmark for causal question answering. In Proceedings of the 29th International Conference on Computational Linguistics, pp. 3296–3308, 2022

Show all 40 references
  1. [9]

    A. M. Lipsky and S. Greenland. Causal directed acyclic graphs. JAMA, 327(11):1083–1084, 2022

  2. [10]

    Tikhonov and M

    A. Tikhonov and M. Ryabinin. It’s all in the heads: Using attention heads as a baseline for cross- lingual transfer in commonsense reasoning. arXiv preprint arXiv:2106.12066, 2021

  3. [11]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, … and I. Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, pages 5998–6008, 2017

  4. [12]

    A. Zeng, M. Cafarella, B. Kenig, M. Markakis, B. Youngmann, and B. Salimi. Causal DAG summarization (full version). arXiv preprint arXiv:2504.14937, 2025

  5. [13]

    Mitra, B

    C. Mitra, B. Huang, T. Darrell, and R. Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420–14431, 2024

  6. [14]

    D. Liu, Y. Feng, Y. Yan, and B. Xu. Towards understanding bugs in Python interpreters. Empirical Software Engineering, 28(1):19, 2023

  7. [15]

    H. Ma, W. Xu, Y. Wei, L. Chen, L. Wang, Q. Liu, and S. Wu. EX-FEVER: A dataset for multi-hop explainable fact verification. arXiv preprint arXiv:2310.09754, 2023

  8. [16]

    I. M. Ali. A guide for positivist research paradigm: From philosophy to methodology. Idealogy Journal, 9(2), 2024

  9. [17]

    J. Liu, P. Tang, W. Wang, Y. Ren, X. Hou, P. A. Heng, … and C. Li. A survey on inference optimization techniques for mixture of experts models. arXiv preprint arXiv:2412.14219, 2024

  10. [18]

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, … and D. Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, volume 35, pages 24824–24837, 2022

  11. [19]

    Zhang, S

    J. Zhang, S. Hu, C. Lu, R. Lange, and J. Clune. Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents. arXiv preprint arXiv:2505.22954, 2025

  12. [20]

    Grotov, S

    K. Grotov, S. Titov, V. Sotnikov, Y. Golubev, and T. Bryksin. A large-scale comparison of Python code in Jupyter notebooks and scripts. In Proceedings of the 19th International Conference on Mining Software Repositories, pages 353–364, 2022

  13. [21]

    Sinha, S

    K. Sinha, S. Sodhani, J. Dong, J. Pineau, and W. L. Hamilton. CLUTRR: A diagnostic benchmark for inductive reasoning from text. arXiv preprint arXiv:1908.06177, 2019

  14. [22]

    K. Song, X. Wang, X. Tan, H. Jiang, C. Zhang, Y. Shen, … and L. Qiu. Chain-of-model learning for language model. arXiv preprint arXiv:2505.11820, 2025

  15. [23]

    L. Yu, D. Chen, S. Xiong, Q. Wu, Q. Liu, D. Li, … and L. Pan. Improving causal reasoning in large language models: A survey. arXiv preprint arXiv:2410.16676, 2024

  16. [24]

    M. L. Ghahroud, N. Ebrahimiyan, S. B. Abadi, F. Jafari, and S. Tajdini. The sandbox approach. Journal of FinTech and Artificial Intelligence, 1(1):20–22, 2021

  17. [25]

    Stephens

    M. Stephens. Sandbox. In Encyclopedia of Cryptography, Security and Privacy, pages 1–5. Springer Berlin Heidelberg, Berlin, Heidelberg, 2024

  18. [26]

    Karupiah

    P. Karupiah. Positivism. In Principles of Social Research Methodology, pages 73–82. Springer Nature Singapore, Singapore, 2022

  19. [27]

    S. Diao, P. Wang, Y. Lin, R. Pan, X. Liu, and T. Zhang. Active prompting with chain-of-thought for large language models. arXiv preprint arXiv:2302.12246, 2023

  20. [28]

    Shafayat, F

    S. Shafayat, F. Tajwar, R. Salakhutdinov, J. Schneider, and A. Zanette. Can large reasoning models self-train? arXiv preprint arXiv:2505.21444, 2025

  21. [29]

    V. P. Dwivedi and X. Bresson. A generalization of transformer networks to graphs. arXiv preprint arXiv:2012.09699, 2020. 11

  22. [30]

    W. Cai, J. Jiang, F. Wang, J. Tang, S. Kim, and J. Huang. A survey on mixture of experts in large language models. IEEE Transactions on Knowledge and Data Engineering, 2025

  23. [31]

    W. Gan, Z. Ning, Z. Qi, and P. S. Yu. Mixture of Experts (MoE): A Big Data Perspective. arXiv preprint arXiv:2501.16352, 2025

  24. [32]

    Wright, D

    W. Wright, D. Schroh, P. Proulx, A. Skaburskis, and B. Cort. The sandbox for analysis: Concepts and methods. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pages 801–810, 2006

  25. [33]

    Y. S. Park, L. Konge, and A. R. Artino Jr. The positivism paradigm of research. Academic Medicine, 95(5):690–694, 2020

  26. [34]

    Y. Yao, Z. Li, and H. Zhao. Beyond chain-of-thought, effective graph-of-thought reasoning in language models. arXiv preprint arXiv:2305.16582, 2023

  27. [35]

    Y. Zhou, T. Lei, H. Liu, N. Du, Y. Huang, V. Zhao, … and J. Laudon. Mixture-of-experts with expert choice routing. In Advances in Neural Information Processing Systems, volume 35, pages 7103–7114, 2022

  28. [36]

    Z. Chen, Y. Li, and K. Wang. Optimizing reasoning abilities in large language models: A step-by-step approach. Authorea Preprints, 2024

  29. [37]

    Z. Jin, Y. Chen, F. Leeb, L. Gresele, O. Kamal, Z. Lyu, and B. Schölkopf. CLADDER: Assessing causal reasoning in language models. In Advances in Neural Information Processing Systems, vol. 36, pp. 31038–31065, 2023

  30. [38]

    Z. Yu, L. He, Z. Wu, X. Dai, and J. Chen. Towards better chain-of-thought prompting strategies: A survey. arXiv preprint arXiv:2310.04959, 2023

  31. [39]

    Zhang, A

    Z. Zhang, A. Zhang, M. Li, H. Zhao, G. Karypis, and A. Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023

  32. [40]

    Zheng, Y

    Z. Zheng, Y. Wang, Y. Huang, S. Song, M. Yang, B. Tang, … and Z. Li. Attention heads of large language models: A survey. arXiv preprint arXiv:2409.03752, 2024

Pith tools

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