Pith. sign in

REVIEW 5 major objections 5 minor 39 references

ECCoT: A Framework for Enhancing Effective Cognition via Chain of Thought in Large Language Model

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read ECCoT claims that scoring the cosine similarity between a question, its chain of thought, and the answer, then discarding low-scoring chains, raises chain-of-thought accuracy on SVAMP from 78.23% to 92.72%.

desk verdict A reasonable filtering idea buried under a draft with inconsistent math, missing details, and an unvalidated selection step. read the letter →

arxiv 2506.19599 v1 pith:LLR42AXN submitted 2025-06-24 cs.CL cs.AI

Zhenke Duan , Jiqun Pan , Jiani Tu , Xiaoyi Wang , Yanqing Wang This is my paper
classification cs.CLcs.AI
keywords ECCoTchain-of-thoughtreasoningchainvalidationembeddedtopicmodelcausalsentenceembeddingorderstatisticsLLMinterpretability
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

This paper proposes ECCoT, an end-to-end framework that tries to make chain-of-thought reasoning in large language models more accurate and more interpretable by validating the reasoning chain before it is used. The framework first generates topic-conditioned reasoning chains using a Markov-random-field-embedded topic model, then embeds the question, chain, and answer with a causally fine-tuned sentence encoder, and finally removes the chains whose question-chain-answer similarities are statistical outliers. The authors report that this pipeline outperforms six reasoning baselines on all three benchmarks, with the largest gain on SVAMP arithmetic word problems, where accuracy rises from 78.23% for standard step-by-step prompting to 92.72%. The wager is that similarity structure among question, reasoning, and answer is a usable proxy for reasoning quality.

What carries the argument

The load-bearing mechanism is the Rank Framework: for each sample it computes r(1)=CosSim(Q,R), r(2)=CosSim(R,A), and r(3)=CosSim(Q,A) in the CSBert embedding space, then uses the distribution of these coefficients to truncate low-effectiveness chains. The other two components feed this filter: MRF-ETM is an embedded topic model with a Markov random field pairwise similarity term that supplies topic labels and topic-conditioned reasoning prompts, and CSBert is a siamese network fine-tuned with a contrastive loss so that causal triplets (Q, R, A) with genuine causal links are pulled together in vector space.

What would settle it

Take SVAMP, generate chains with the teacher model, then deliberately corrupt the middle steps of half the correct chains while keeping the final answer unchanged. If the Rank Framework still scores the corrupted chains as high as the correct ones and downstream accuracy stays high, the similarity filter is selecting for answer alignment rather than reasoning validity.

Watch

Extended reading notes

Core claim

The paper claims that a reasoning chain's effectiveness can be scored by three cosine similarities — between question and chain, chain and answer, and question and answer — and that keeping chains whose scores rank high under order statistics concentrates training and inference on what the authors call effective cognition. Combined with MRF-ETM topic injection and CSBert causal embedding, this filtering is claimed to outperform Step-by-Step, Curation, Expansion, Feedback, Self-Knowledge, and Vanilla Fine-tuning on ANLI, SVAMP, and CommonQA, and its ablation shows that removing the rank filter hurts the most, dropping SVAMP accuracy from 92.72% to 72.98%.

Load-bearing premise

The Rank Framework assumes that cosine similarity between the reasoning chain, the question, and the answer measures how good the reasoning is, but the paper offers no evidence that a flawed chain cannot hide behind high similarity scores.

Editorial extensions

If this is right

  • If the filter selects effective chains, a small student model fine-tuned only on retained chains should inherit higher accuracy than the same model trained on all generated chains.
  • The ablation claims rank filtering is the largest contributor, so any improvement from topic injection or causal embedding may be contingent on the filter being present.
  • The scaling-law table claims that larger student models (7B to 13B) improve ECCoT accuracy further, suggesting the framework transfers across model sizes.
  • The paper's human interpretability scores are claimed to improve with ECCoT, implying the retained chains are not only more accurate but also more readable.

Reading between the lines

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

  • The similarity filter is vulnerable to pseudo-alignment: a chain that merely paraphrases the final answer could score high on r(2) without containing valid reasoning, so a test that corrupts middle steps while preserving the answer would separate similarity from validity.
  • The order-statistics truncation trades off chain quantity for quality; it may also amplify dataset biases if certain topics or phrasings systematically score higher in the embedding space.
  • The framework's reliance on a larger teacher model to generate chains means its benefit may partly be distillation rather than validation; comparing against a teacher-generated-chain baseline without filtering would isolate the filter's contribution.
  • One could test the mechanism directly by replacing CSBert cosine similarities with random scores: if accuracy stays high, the filter is not doing causal work.
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

5 major / 5 minor

Summary. The paper proposes ECCoT, an end-to-end framework for validating chain-of-thought reasoning in large language models. It combines a Markov Random Field-Embedded Topic Model (MRF-ETM) for topic-aware chain generation, a Causal Sentence-BERT (CSBert) for causal embedding, and a 'Rank Framework' that filters reasoning chains by cosine similarities among question, reasoning, and answer embeddings. The filtered chains are used to fine-tune a student LLM. Experiments on ANLI, SVAMP, and CommonQA report consistent improvements over several baselines, with an ablation study attributing a large performance gain to the Rank module.

Significance. If validated, the framework would address the important problem of filtering unreliable chain-of-thought reasoning and improving LLM interpretability, and the code release is a useful step toward reproducibility. However, as presented, the paper does not establish that the proposed similarity-based filter captures reasoning validity, and the MRF-ETM objective is internally inconsistent. Because the ablation attributes most of the reported improvement to the unvalidated filter and the headline comparisons lack essential experimental controls, the current results do not provide reliable evidence for the framework's effectiveness.

major comments (5)
  1. [Methodology, Eq. (1) and Eqs. (13)-(17)] The generative model and the optimization objective are internally inconsistent. Equation (1) defines p(w_ij=v, z_ij=k) with the MRF similarity term inside the probability, but the term is not normalized and appears nowhere in the ELBO expectation derived in Eq. (13). It is instead reintroduced as a separate additive term in Eq. (14) and Eq. (16), while the KL terms listed in Eq. (17) are omitted from Eq. (16). As a result, the reader cannot determine what objective is actually optimized, how the similarity weight lambda is set, or how the word-pair set P is constructed.
  2. [Rank Framework and Figure 4] The Rank Framework's filtering criterion uses r(1)=cos(Q,R), r(2)=cos(R,A), and r(3)=cos(Q,A), with 'structured ordering statistics' and 'truncation', but the paper provides no argument or experiment showing that these cosine similarities track reasoning validity. The introduction itself warns, citing Greenblatt et al. (2024), that invalid chains can produce correct answers through pseudo-alignment; a chain that simply restates the answer can score high on r(2) while containing bogus intermediate steps. Since Table 4 shows that removing the Rank module drops SVAMP accuracy from 92.72 to 72.98, the paper's central claim rests on this unvalidated filter. A concrete test would be to compare the filter's selections with human validity ratings or to run a control condition with deliberately flawed chains that paraphrase the final answer; neither is reported.
  3. [Pretrained Causal Sentence-Bert, Eq. (18)] The contrastive losses L_qr and L_ra in Eq. (18) are never defined, and the triplet construction, the observation function y, and the fine-tuning data are unspecified. The same section reports that 'eye-tracking experiments' show CSBert aligns with human causal reasoning, but no participants, stimuli, or results are given. These unsupported claims make the CSBert component impossible to assess or reproduce.
  4. [Experiments, Tables 1, 2, 4, 5] Essential experimental details are missing. There is no statement of the number of runs, standard deviations, or significance tests, so the large differences in Tables 2 and 4 cannot be evaluated. The test-model description ('50 iterations' and 'default parameters of the llama-factory') is insufficient, and no threshold or order-statistic rule is given for the Rank filter. The teacher-student distillation protocol and the generation procedure for the reasoning chains are also unspecified. Without these details, the reported accuracies, especially the 92.72% SVAMP result, are not reproducible.
  5. [Comparison of Effective Cognition, Table 3] BLEU and ROUGE are lexical overlap metrics, not measures of 'cognitive process effectiveness'. Using them as outcome measures partially overlaps with the cosine-similarity family used in the Rank filter, creating a circularity concern. The text promises 'human interpretability scores' but no such scores are reported. Thus Table 3 does not independently validate the framework's cognitive claims.
minor comments (5)
  1. [Equation (7) and Equation (13)] The subscript in Eq. (7) appears as 'beta_k,mi_j' instead of 'beta_{k,w_ij}', and the same typo is inherited in Eq. (13).
  2. [Figures 1 and 2] Figures 1 and 2 contain corrupted symbols such as '??' and '∅?', making the proposed pipeline and the MRF-ETM workflow unreadable; the figures and their captions should be regenerated.
  3. [Table 2 and dataset naming] Table 2 contains 'ECCoT72.23' with a missing space, and the paper inconsistently uses 'SV AMP' and 'SVAMP'; these should be unified.
  4. [Reference list] The reference to 'Necva, Burcu, & Harun (2023)' describes a 'Siamese neural network', which is not obviously the 'Causal Sentence-BERT' used in the paper; the relationship should be clarified. The 'Patel, A., et al. (2021)' reference should list the full author list.
  5. [Scaling law verification, Table 5] The scaling-law comparison in Table 5 confounds model family and parameter size: LLama2-7B, LLama3.1-8B, and LLama2-13B differ not only in size but also in architecture and training, so the table does not isolate a scaling law.

Circularity Check

1 steps flagged · score 3.0 of 10

Partial circularity: the 'effective cognition' evaluation restates the Rank filter's similarity objective, but the headline accuracy results are held-out and independent.

  1. self definitional [Methodology, Rank Framework (Figure 4) and Experiments, Table 3]
    "By calculating similarity coefficients for each cognitive process and filtering out low-effectiveness distributions, we retain high-effectiveness cognitive process samples. ... In addition to traditional accuracy metrics, we incorporated cognitive reasonableness metrics, such as human interpret ability scores, to evaluate the effectiveness of the reasoning chains generated by ECCoT and the baseline methods. [Table 3:] Comparison of Cognitive Process Effectiveness Across Different Datasets: BLEU-4 ROUGE-1 ROUGE-2 ROUGE-L"

    The Rank filter selects chains using r(1)=cos(Q,R'), r(2)=cos(R',A), r(3)=cos(Q,A), i.e., cosine similarity between the generated chain, question, and answer. The paper's 'cognitive process effectiveness' evaluation in Table 3 is operationalized with BLEU-4 and ROUGE-1/2/L, which are also similarity/overlap metrics against a reference. Because the same family of similarity-to-reference is both the selection objective and the outcome metric, the reported improvement in 'effective cognition' is to a substantial degree a restatement of the filter's own objective rather than independent evidence. The paper never reports the promised human interpretability scores that could break this overlap, so the 'effectiveness' claim is partly defined by the same similarity construct used for filtering.

full rationale

The central result, Table 2 accuracy (e.g., 92.72% on SVAMP versus 78.23% for Step-by-Step), is an empirical comparison on held-out splits after fine-tuning the student on Rank-filtered chains; it does not reduce by definition to the filter's similarity scores because accuracy is measured against ground-truth labels. The MRF-ETM and CSBert components are trained on the target-task data, but that is standard distillation practice, not circular. No load-bearing self-citation chain or imported uniqueness theorem is present; the only author-overlapping citation (FlexKBQA, Z. Li et al. 2023, with co-author Duan) is a background citation in related work, not used to justify ECCoT. The one genuine circularity concern is in the 'effective cognition' evaluation: the Rank filter's scoring function and Table 3's BLEU/ROUGE metrics are both similarity/overlap measures against the answer/reference, so Table 3's claim that ECCoT improves 'cognitive process effectiveness' is partly self-fulfilling. The paper's own admission of 'pseudo-alignment' (Greenblatt et al., 2024) and the lack of validation that cosine similarity tracks reasoning validity are correctness risks rather than definitional circularity, and therefore do not raise the score above 3.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

No invented physical entities are introduced. The central claim depends on two fitted components (MRF-ETM and CSBert) and an unvalidated filtering criterion; key parameters (lambda, topic count, filter threshold) are unreported.

free parameters (4)
  • lambda (MRF similarity weight)
    Hyperparameter controlling the Markov Random Field similarity term in Eq. (1) and Eq. (15). No value is reported, and the sign of the term conflicts between the generative formulation and the loss.
  • Number of topics K in MRF-ETM
    Not reported in the experiments; central to the topic-conditioning stage of the framework.
  • Rank filtering threshold
    The structured ordering statistics filter requires a cutoff for discarding ineffective chains; no threshold or retention rate is given.
  • CSBert contrastive loss components L_qr and L_ra
    Defined in Eq. (18) but never specified; the fine-tuning of CSBert cannot be reproduced.
assumptions (3)
  • ad hoc to paper Equation (1) defines a valid probability distribution over words and topics despite lacking a normalization constant.
    The exponential MRF term is written without a partition function; all subsequent ELBO manipulations assume a normalized likelihood.
  • standard math Mean-field variational family q(theta, z) = q(theta) times product of q(z_ij) closely approximates the posterior.
    Standard variational inference assumption, invoked in Eqs. (2)-(4).
  • domain assumption Cosine similarity between reasoning chain, answer, and question is a valid measure of reasoning effectiveness.
    Load-bearing premise of the Rank Framework; not empirically established in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ECCoT: A Framework for Enhancing Effective Cognition via Chain of Thought in Large Language Model." pith.science (2026). https://pith.science/paper/LLR42AXN

@misc{pith2026250619599,
  author       = {Pith},
  title        = {Pith review of: ECCoT: A Framework for Enhancing Effective Cognition via Chain of Thought in Large Language Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LLR42AXN}},
  note         = {Machine review of arXiv:2506.19599}
}
read the original abstract

In the era of large-scale artificial intelligence, Large Language Models (LLMs) have made significant strides in natural language processing. However, they often lack transparency and generate unreliable outputs, raising concerns about their interpretability. To address this, the Chain of Thought (CoT) prompting method structures reasoning into step-by-step deductions. Yet, not all reasoning chains are valid, and errors can lead to unreliable conclusions. We propose ECCoT, an End-to-End Cognitive Chain of Thought Validation Framework, to evaluate and refine reasoning chains in LLMs. ECCoT integrates the Markov Random Field-Embedded Topic Model (MRF-ETM) for topic-aware CoT generation and Causal Sentence-BERT (CSBert) for causal reasoning alignment. By filtering ineffective chains using structured ordering statistics, ECCoT improves interpretability, reduces biases, and enhances the trustworthiness of LLM-based decision-making. Key contributions include the introduction of ECCoT, MRF-ETM for topic-driven CoT generation, and CSBert for causal reasoning enhancement. Code is released at: https://github.com/erwinmsmith/ECCoT.git.

Figures

Figures reproduced from arXiv: 2506.19599 by the authors.

Figure 1
Figure 1. (a) Theme Recognition Stage, (b) Theme Explanation Stage, (c) Thought Cognition Stage, (d) Effective Cognition [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Showcased the main workflow and optimization [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) The fine-tuning stage for causal embedding (b) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: (a) Calculate the similarity of the Q, R, and A [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 28 canonical work pages

  1. [1]

    Singla, P., & Garg, D. (2021). Explanations for com- monsenseqa: New dataset and models. InProceedings of the 59th annual meeting of the association for computa- tional linguistics and the 11th international joint conference on natural language processing (volume 1: Long papers) (p. 3050–3065). Online: Association for Computational Linguistics

    work page 2021

  2. [2]

    Allamanis, M., Panthaplackel, S., & Yin, P. (2025). Unsuper- vised evaluation of code llms with round-trip correctness. In Proceedings of the 41st international conference on machine learning (icml’24)(V ol. 235, p. 1050–1066). JMLR.org

    work page 2025

  3. [3]

    Ang, Y ., Bao, Y ., Huang, Q., Tung, A. K. H., & Huang, Z. (2024). Tsgassist: An interactive assistant harness- ing llms and rag for time series generation recommenda- tions and benchmarking. InProc. vldb endow.(V ol. 17, p. 4309–4312). Retrieved from https://doi.org/10 .14778/3685800.3685862

    arXiv 2024

  4. [4]

    Chen, Z., Chen, J., Singh, A., & Sra, M. (2024). Xplain- llm: A knowledge-augmented dataset for reliable grounded explanations in llms. InProceedings of the 2024 confer- ence on empirical methods in natural language processing (p. 7578–7596). Miami, Florida, USA: Association for Computational Linguistics

    work page 2024

  5. [5]

    W., Hou, L., Longpre, S., Zoph, B., Tai, Y ., Fe- dus, W.,

    Chung, H. W., Hou, L., Longpre, S., Zoph, B., Tai, Y ., Fe- dus, W., . . . Wei, J. (2024, January). Scaling instruction- finetuned language models.Journal of Machine Learning Research,25(1), 1–53

    work page 2024

  6. [6]

    Fang, Z., He, Y ., & Procter, R. (2024). Cwtm: Leverag- ing contextualized word embeddings from bert for neural topic modeling. InProceedings of the 2024 joint interna- tional conference on computational linguistics, language resources and evaluation (lrec-coling 2024)(p. 4273–4286)

    work page 2024

  7. [7]

    Hubinger, E

    Greenblatt, R., Denison, C., Wright, B., Roger, F., MacDi- armid, M., Marks, S., . . . Hubinger, E. (2024).Alignment faking in large language models

    work page 2024

  8. [8]

    Guo, R., Xu, W., & Ritter, A. (2024). Meta-tuning llms to leverage lexical knowledge for generalizable language style understanding. InProceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: Long papers)(p. 13708–13731). Bangkok, Thailand: Association for Computational Linguistics

    work page 2024

Show all 39 references
  1. [9]

    Haffari, G. (2024). Direct evaluation of chain-of-thought in multi-hop reasoning with knowledge graphs. InFind- ings of the association for computational linguistics: Acl 2024(p. 2862–2883). Bangkok, Thailand: Association for Computational Linguistics

    2024

  2. [10]

    Ho, N., Schmid, L., & Yun, S.-Y . (2023). Large language models are reasoning teachers. InProceedings of the 61st annual meeting of the association for computational linguis- tics (volume 1: Long papers)(pp. 14852–14882). Toronto, Canada: Association for Computational Linguistics

    2023

  3. [11]

    Huang, X., & Gong, H. (2024). A dual-attention learning network with word and sentence embedding for medical visual question answering.IEEE Transactions on Medical Imaging,43(2), 832-845

    2024

  4. [12]

    Reif, E., & Dixon, L. (2024). Llm comparator: Interactive analysis of side-by-side evaluation of large language models. IEEE transactions on visualization and computer graphics

    2024

  5. [13]

    G., Latif, E., Wu, X., Liu, N., & Zhai, X

    Lee, G. G., Latif, E., Wu, X., Liu, N., & Zhai, X. (2024). Applying large language models and chain-of-thought for automatic scoring.Computers and Education: Artificial Intelligence, 100213-

    2024

  6. [14]

    Lefebvre, G., Elghazel, H., Guillet, T., Aussem, A., & Sonnati, M. (2024). A new sentence embedding framework for the education and professional training domain with applica- tion to hierarchical multi-label text classification.Data & Knowledge Engineering, 102281-

    2024

  7. [15]

    Li, Q., Chen, Z., Ji, C., Jiang, S., & Li, J. (2025). Llm-based multi-level knowledge generation for few-shot knowledge graph completion. InProceedings of the thirty-third in- ternational joint conference on artificial intelligence (ijcai ’24)(p. 2135–2143). Article 236: IJCAI...

    2025 doi

  8. [16]

    Li, Y ., Yang, Y ., Zhu, J., Chen, H., & Wang, H. (2024). Llm-empowered few-shot node classification on incomplete graphs with real node degrees. InProceedings of the 33rd acm international conference on information and knowledge management (cikm ’24)(p. 1306–1315). New York, ...

    2024

  9. [17]

    (2023).Flexkbqa: A flexible llm-powered frame- work for few-shot knowledge base question answering

    Wang, J. (2023).Flexkbqa: A flexible llm-powered frame- work for few-shot knowledge base question answering. arXiv:2308.12060

    2023 arXiv

  10. [18]

    Lin, Z., Chen, H., Lu, Y ., Rao, Y ., Xu, H., & Lai, H. (2024). Hierarchical topic modeling via contrastive learning and hy- perbolic embedding. InProceedings of the 2024 joint inter- national conference on computational linguistics, language resources and evaluation (lrec-col...

    2024

  11. [19]

    Mahmood, A., Wang, J., Yao, B., Wang, D., & Huang, C. M. (2025). User interaction patterns and breakdowns in con- versing with llm-powered voice assistants.International Journal of Human - Computer Studies, 103406-103406

    2025

  12. [20]

    Mondal, D., Modi, S., Panda, S., Singh, R., & Rao, G. (2024). Kam-cot: Knowledge augmented multimodal chain- of-thoughts reasoning. InAaai conference on artificial intelligence

    2024

  13. [21]

    Necva, B., Burcu, C., & Harun, A. (2023). A siamese neu- ral network for learning semantically-informed sentence embeddings.Expert Systems With Applications

    2023

  14. [22]

    Ni, J., Shi, M., Stammbach, D., Sachan, M., Ash, E., & Leip- pold, M. (2024). Afacta: Assisting the annotation of factual claim detection with reliable llm annotators. InProceedings of the 62nd annual meeting of the association for computa- tional linguistics (volume 1: Long p...

    2024

  15. [23]

    Kiela, D. (2020). Adversarial NLI: A new benchmark for natural language understanding. InProceedings of the 58th annual meeting of the association for computational linguis- tics(p. 4885–4901). Association for Computational Lin- guistics. Retrieved from https://aclanthology.or...

    2020

  16. [24]

    (2021).Svamp: Simple variations on arithmetic math word problems.arXiv:2106.05827 [cs.CL]

    Patel, A., et al. (2021).Svamp: Simple variations on arithmetic math word problems.arXiv:2106.05827 [cs.CL]. Retrieved fromhttps://arxiv.org/abs/2106.05827

    2021 arXiv

  17. [25]

    Qiu, C., Xie, Z., Liu, M., & Hu, H. (2024). Explainable knowledge reasoning via thought chains for knowledge- based visual question answering.Inf. Process. Manag.,61, 103726

    2024

  18. [26]

    Shridhar, K., Stolfo, A., & Sachan, M. (2023). Distilling reasoning capabilities into smaller language models. In Findings of the association for computational linguistics: Acl 2023(p. 7059–7073). Toronto, Canada: Association for Computational Linguistics. van Schaik, T. A., &...

    2023

  19. [27]

    Wang, P., Chan, A., Ilievski, F., Chen, M., & Ren, X. (2022). Pinto: Faithful language reasoning using prompt-generated rationales.arXiv:2211.01562

    2022 arXiv

  20. [28]

    Wang, Z., Li, C., Yang, Z., Liu, Q., Hao, Y ., Chen, X., . . . Sui, D. (2024). Analyzing chain-of-thought prompting in black- box large language models via estimated v-information. In Proceedings of the 2024 joint international conference on computational linguistics, language...

    2024

  21. [29]

    Xia, F., . . . Zhou, D. (2022). Chain-of-thought prompting elicits reasoning in large language models. InProceedings of the 36th international conference on neural information processing systems (nips ’22)(pp. 24824–24837). Red

    2022

  22. [31]

    Williams, A., Thrush, T., & Kiela, D. (2022). Anlizing the ad- versarial natural language inference dataset. InProceedings of the 5th annual meeting of the society for computation in linguistics.Association for Computational Linguistics. Re- trieved from https://github.com/fac...

    2022

  23. [32]

    Wu, Y ., Zhang, Z., & Zhao, H. (2024). Mitigating mislead- ing chain-of-thought reasoning with selective filtering. In Proceedings of the 2024 joint international conference on computational linguistics, language resources and evalu- ation (lrec-coling 2024)(p. 11325–11340). T...

    2024

  24. [33]

    Xu, F., Zhou, T., Nguyen, T., Bao, H., Lin, C., & Du, J. (2025). Integrating augmented reality and llm for enhanced cogni- tive support in critical audio communications.International Journal of Human - Computer Studies, 103402-103402

    2025

  25. [34]

    Xu, X., Li, M., Tao, C., Shen, T., Cheng, R., Li, J., . . . Zhou, T. (2024, October 21).A survey on knowledge distillation of large language models.arXiv:2402.13116v4 [cs.CL]. Retrieved from https://arxiv.org/abs/2402.13116v4 (arXiv preprint) yang Lu, H., ci Liu, T., Cong, R.,...

    2025 arXiv

  26. [35]

    Zhang, J., Gao, H., Zhang, P., Feng, B., Deng, W., & Hou, Y . (2024). La-ucl: Llm-augmented unsupervised contrastive learning framework for few-shot text classification. InPro- ceedings of the 2024 joint international conference on com- putational linguistics, language resourc...

    2024

  27. [36]

    (2022, Octo- ber 7).Automatic chain of thought prompting in large lan- guage models.arXiv:2210.03493v1 [cs.CL]

    Zhang, Z., Zhang, A., Li, M., & Smola, A. (2022, Octo- ber 7).Automatic chain of thought prompting in large lan- guage models.arXiv:2210.03493v1 [cs.CL]. Retrieved fromhttps://arxiv.org/abs/2210.03493v1

    2022 arXiv

  28. [37]

    (2024, May 20).Multimodal chain-of-thought rea- soning in language models.arXiv:2302.00923v5 [cs.CL]

    Zhang, Z., Zhang, A., Li, M., Zhao, H., Karypis, G., & Smola, A. (2024, May 20).Multimodal chain-of-thought rea- soning in language models.arXiv:2302.00923v5 [cs.CL]. Retrieved from https://arxiv.org/abs/2302.00923v5 (Reviewed on OpenReview: https://openreview.net/ forum?id=y1...

    2024 arXiv

  29. [38]

    Zhou, D., Scharli, N., Hou, L., Wei, J., Scales, N., Wang, X., . . . Chi, E. H. (2022). Least-to-most prompting enables complex reasoning in large language models.arXiv preprint arXiv:2205.10625

    2022 arXiv

  30. [39]

    (2024).Can language models perform robust reason- ing in chain-of-thought prompting with noisy rationales? arXiv:2410.23856

    Zhou, Z., Tao, R., Zhu, J., Luo, Y ., Wang, Z., & Han, B. (2024).Can language models perform robust reason- ing in chain-of-thought prompting with noisy rationales? arXiv:2410.23856

    2024 arXiv

  31. [40]

    Zhu, X., Li, J., Liu, Y ., Ma, C., & Wang, W. (2024). Distilling mathematical reasoning capabilities into small language models.Neural Networks, 106594-106594

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.