Pith. sign in

REVIEW 4 major objections 6 minor 49 references

Breaking Thought Patterns: A Multi-Dimensional Reasoning Framework for LLMs

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read LADDER claims that a 32B LLM reaches higher generation diversity and creativity than GPT-4o by routing reasoning through CoT, MoE, and multidimensional semantic projection, at a small cost in exact task success.

desk verdict A plausible recombination of known ideas, sunk by internally inconsistent numbers and a gating mechanism that the equations don't deliver. read the letter →

arxiv 2506.13192 v1 pith:7UDAQK3X submitted 2025-06-16 cs.CL cs.AI

Xintong Tang , Meiru Zhang , Shang Xiao , Junzhao Jin , Zihan Zhao , Liwei Li , Yang Zheng , Bangyi Wu This is my paper
classification cs.CLcs.AI
keywords LADDERframeworkChain-of-ThoughtreasoningMixtureofExpertssemanticdimensionalityreductiontextgenerationdiversitycreativelargelanguagemodelsself-BLEU
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 proposes LADDER, a framework that wraps an LLM in three processing stages: semantic lifting into a high-dimensional abstract space, reasoning across multiple Mixture-of-Experts subnetworks, and projection back down to concrete output. Its central claim is that this pipeline breaks the repetitive thought patterns of standard LLMs, letting a 32B model generate more diverse and creative text than much larger models such as GPT-4o and DeepSeek-R1. The authors report that LADDER achieves the lowest Self-BLEU and highest Distinct-2 scores among all compared models, while coming within a few percentage points on task success and semantic consistency. If true, the framework would show that creativity and diversity can be engineered through routing and dimension transformations rather than through sheer parameter scale.

What carries the argument

The load-bearing mechanism is a three-stage transformation. Stage one, Semantic Lifting, applies an expansion projection $C^{\mathrm{up}}_i = W^{\mathrm{up}} x_i + b^{\mathrm{up}}$ to move token or keyword representations into a higher-dimensional semantic space. Stage two, Expert Reasoning, runs a sparsely gated Mixture-of-Experts network where a gating network $G(x) = \mathrm{softmax}(W_g x + b_g)$ computes weights over experts and the output is the weighted sum $y = \sum_i G_i(x) E_i(x)$, with top-k routing to keep only a subset of experts active. Stage three, Dimensional Descent, projects back to the original space via $x_{\mathrm{final}} = W^{\mathrm{down}} y + b^{\mathrm{down}}$, guided by a semantic drift loss and multi-head projections to preserve meaning. The paper's thesis is that this expand-reason-compress cycle is what disrupts rigid reasoning paths and produces more creative, diverse outputs.

What would settle it

Take a random held-out sample from WritingPrompts, CommonsenseQA, and AlpacaEval, fix prompts and decoding seeds, and run LADDER against GPT-4o and DeepSeek-R1; if the Self-BLEU and Distinct-2 advantages shrink to noise or the human creativity preference drops below chance, the central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that combining Chain-of-Thought reasoning, Mixture-of-Experts routing, and multi-dimensional up/down-sampling produces a measurable 'divergent thinking' capability in LLMs. Concretely, Table 1 reports LADDER at Self-BLEU 0.06, Distinct-2 0.46, BERTScore 0.88, task success 89% (writing) and 86% (Q&A), and fluency 4.7, compared with ChatGPT-4o at 0.08, 0.41, 0.90, 92%, and 4.9. The authors interpret these numbers as evidence that a 32B model can outperform much larger commercial systems on diversity and creativity while staying close on fluency and semantic coherence. In creative writing, LADDER received a 48.4% top-1 human preference rate for creativity, versus 15.7% for ChatGPT-4o and 13.6% for DeepSeek-R1. Ablation results attribute drops of roughly nine to fourteen percentage points in task success to the removal of the CoT, MoE, or dimension-mapping stage, supporting the paper's conclusion that all three components contribute to the effect.

Load-bearing premise

The comparison rests on 300 hand-picked 'representative' samples (100 per task) with no documented selection protocol, so if those samples were chosen to favor divergent answers, every reported gain—including the human preference numbers—could be an artifact of sample selection.

Editorial extensions

If this is right

  • A 32B LADDER model would produce more lexically diverse text than GPT-4o (Distinct-2 0.46 versus 0.41) while staying within 0.02 on BERTScore and 0.2 points on fluency.
  • In creative writing, human raters would select LADDER's output as the most creative roughly three times as often as GPT-4o's, based on the reported 48.4% versus 15.7% top-1 preference rate.
  • Removing any one of the three stages—CoT, MoE, or the dimension-mapping module—would cost roughly 9 to 14 percentage points in task success relative to the full framework's 83.7%.
  • On instruction following, LADDER would match open-source baselines such as LLAMA3 near 88% task completion while adding stylistic adaptability across formatting and tone constraints.
  • On commonsense QA, LADDER would trail GPT-4o by about six percentage points (86% versus 92%) but still beat other open-source baselines, suggesting a trade-off between exploratory answers and exact correctness.

Reading between the lines

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

  • If the diversity gains replicate on unbiased samples, the result would imply that creative output can be bought with routing and representation compression instead of parameter count, potentially making smaller models practical for open-ended generation.
  • The same up-expand-reason-compress pattern could apply to other structured generation domains such as code, music, or molecule design, where candidate diversity matters, though the paper only tests three NLP tasks.
  • The reported CommonsenseQA trade-off suggests the framework could serve as an architectural diversity dial: a task-tunable preference for exploring alternative answers over producing the canonical one.
  • Because the human study uses 50 questions and 80 valid responses with a single model trio, a larger preregistered comparison across varied prompts would be needed to confirm the preference gap is stable rather than prompt-specific.
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 proposes LADDER, a framework that combines Chain-of-Thought prompting, a Mixture-of-Experts routing layer, and semantic dimensionality lifting/reduction (``DimMap'') on top of an LLM, with the stated goal of increasing output diversity and creativity while maintaining coherence. The authors claim that a 32B LADDER system outperforms or matches much larger commercial and open models on Self-BLEU, Distinct-2, BERTScore, task success, and fluency. Experiments are reported on creative writing, CommonsenseQA, and instruction following, using 100 ``representative samples'' per task, plus a human preference study. The paper also includes an ablation study that attributes performance drops to the removal of CoT, MoE, and DimMap.

Significance. The architectural idea is timely: combining reasoning prompts with modular expert routing is an active area, and the paper correctly identifies diversity and creativity as underexplored evaluation dimensions. The use of external metrics (Self-BLEU, Distinct-2, BERTScore) and a human preference study are appropriate choices. However, the experimental section as written does not support the paper's central claim. There is no code, data, seed, or sampling protocol; no uncertainty quantification; and, most seriously, the numbers reported for the same LADDER system in Table 1 and Table 2 are mutually incompatible. If the authors can reconcile these inconsistencies and provide a reproducible evaluation, the framework would merit further attention, but in its current form the empirical contribution is not verifiable.

major comments (4)
  1. [§4.3, Table 1 vs §4.5, Table 2] The same LADDER system is reported with incompatible numbers. Table 1 lists Self-BLEU 0.06, Distinct-2 0.46, BERTScore 0.88, and task success 89% (Writing) / 86% (Q&A), while Table 2's "Full LADDER" row lists task success 83.70%, Dist-2 0.36, and BERTScore 0.87. The Distinct-2 gap of 0.10 and the task-success gap of more than five points cannot be explained by rounding, and §4.3.3 additionally reports a completion rate close to 88% for instruction following. Without a statement of which aggregation, subset, or protocol each table refers to, the reader cannot determine which numbers support the headline claim that LADDER improves task completion.
  2. [§1 vs §3.3] Section 1 promises an "attention-based multi-granularity gating network," but Section 3.3 defines only G(x)=softmax(W_g x + b_g) (and a per-expert score z_i), with no attention weights, no multi-granularity structure, and no description of top-k selection. Since this gating network is one of the three listed contributions, the architecture as specified is not the architecture that is claimed, and the experimental results cannot be attributed to it.
  3. [§4.1] The selection of "100 representative samples" per task is not described by any sampling protocol, seed, or inclusion criterion, and its stated purpose is "to encourage divergent thinking." Because every reported number, including the ablation results in Table 2 and the human preference figures in Section 4.4, is computed on these 300 samples, sample-selection bias alone could produce the observed differences. The paper needs either a random sampling procedure with seeds or a release of the sample indices and prompts.
  4. [§4.1-4.4] No error bars, confidence intervals, significance tests, or repeated-run information are reported for any automatic metric or human evaluation. Differences such as 89% vs 88% task success and BERTScore 0.88 vs 0.86 are presented as definitive, and the human evaluation (80 respondents on 50 questions) has no inter-annotator agreement or variance report. Without uncertainty quantification, the claimed "significant improvements" are not established.
minor comments (6)
  1. [§3.1] The framework is described as targeting "multimodal large language pre-trained models," but all experiments are text-only; please clarify the scope.
  2. [§2.4] The threshold "2× 1022 FLOPs" should be typeset as 2×10^22 FLOPs.
  3. [Throughout] The paper alternates between "ChatGPT-4o" and "GPT-4o" (for example, in the Abstract, Table 1, and Figure 1); use one name consistently.
  4. [§4.4 vs §4.3.4] Section 4.4 states that participants selected "the most creative and unexpected answer," but Section 4.3.4 reports a 45.4% "best output" rate across all tasks; the relationship between these two measures should be explained.
  5. [Table 1] The LADDER row reports task success as "89% (Writing) / 86% (Q&A)" but does not give the instruction-following result that is described in §4.3.3; an aggregate row with per-task breakdowns would make the comparison clearer.
  6. [§4.5] The term "DimMap" is introduced only in Section 4.5; it should be defined in Section 3 where the semantic lifting and descent modules are described.

Circularity Check

1 steps flagged · score 2.0 of 10

No derivation-level circularity: the central claims are empirical and the metrics are external. One minor self-definitional step: 'creativity' is scored by the exact diversity objective LADDER is designed and prompted to maximize. The Table 1 versus Table 2 discrepancies are correctness risks, not circularity.

  1. self definitional [Section 4.1 (Evaluation Setup) and Section 4.3.1 (Creative Writing Task)]
    "Creativity is jointly measured using Self-BLEU and Distinct-2 scores to evaluate the diversity and novelty of the generated text—lower Self-BLEU and higher Distinct-2 scores indicate better generation diversity. ... For each task, 100 representative samples are selected to encourage divergent thinking, where the model is prompted to explore multiple possible semantic paths and solutions during generation, rather than relying solely on the most frequent or superficial answers."

    The headline 'significantly improves ... creativity' is evidenced by Self-BLEU/Distinct-2, which quantify exactly the behavior the framework is engineered and prompted to produce: the setup states the model is 'prompted to explore multiple possible semantic paths ... rather than relying solely on the most frequent or superficial answers.' The measured quantity (output diversity) is the design and prompting objective, so the reported creativity gain is the system fulfilling its own instruction rather than an independently derived result. Section 4.3.2 confirms the same mechanism trades away task success in QA ('did not always yield the correct answer'), so 'creative' output is the objective restated in metric form.

full rationale

This paper's central assertion is empirical, not derivational: LADDER's diversity, creativity, fluency, and task-completion gains rest on self-reported experiments using external metrics (Self-BLEU, Distinct-2, BERTScore, human votes), and none of the framework equations (G(x)=softmax(W_g x+b_g), C_up=W_up x+b_up, x_final=W_down y+b_down) is substituted into itself or into any reported number. There is no load-bearing self-citation chain, no imported uniqueness theorem, and no fitted parameter renamed as a prediction; the cited prior work (Shazeer et al., Wei et al., etc.) is used for standard ingredients and constitutes independent support. The one mild circular step is the construct alignment between the 'creativity' claim and the diversity objective, flagged above. Several passages carry correctness risk that belongs outside the circularity score but must be weighed: (1) Table 1 reports LADDER Distinct-2 0.46 and task success 89% (Writing) / 86% (Q&A), while Table 2 reports 'Full LADDER' at Dist-2 0.36 and task success 83.70% — the same system cannot yield both without an explained aggregation change, so the headline superiority numbers are mutually unverifiable; (2) the Introduction promises an 'attention-based multi-granularity gating network,' but Section 3.3 provides only a plain softmax gate, leaving the claimed architecture unsupported; (3) Section 4.3.2 concedes the divergent-thinking strategy 'did not always yield the correct answer,' which directly undermines the Abstract's 'significantly improves task completion' on Q&A; and (4) the '100 representative samples' per task are selected with no protocol or seeds, leaving sample-selection bias possible. These are reliability and internal-consistency problems, not derivation-level circularity; the honest verdict is no significant circularity at score 2.

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

The framework rests on standard ML assumptions and on the unexplained selection of 300 evaluation samples. The learned projection and routing matrices are never specified, so the central empirical claim cannot be separated from the particular training run.

free parameters (4)
  • Semantic lifting projection (W_up, b_up) = not disclosed
    Introduced in Section 3.2, maps input features to a high-dimensional space; its values are learned during fine-tuning but no initialization, regularization, or final values are reported.
  • Gating network weights (W_g, b_g) = not disclosed
    Introduced in Section 3.3, routes inputs to experts; values are learned but unspecified.
  • Expert weights (W_i, b_i) = not disclosed
    Each expert sub-network has its own weights (Section 3.3); learned but unspecified.
  • Dimensional descent projection (W_down, b_down) = not disclosed
    Introduced in Section 3.4; maps high-dimensional outputs back to lower dimension; includes a semantic drift loss but no details on the loss weight or anchors.
assumptions (5)
  • domain assumption Chain-of-thought prompting elicits step-by-step reasoning in large language models and improves task performance.
    Invoked in Section 3.1 and Section 2.1 based on [9]; the paper assumes CoT transfers to the LADDER framework without modification.
  • domain assumption Mixture-of-experts improves computational efficiency and enables modular reasoning.
    Invoked in Section 3.3 based on [10, 16]; the paper assumes MoE routing generalizes to its semantic tasks.
  • domain assumption Dimensionality expansion and reduction improve semantic representation and creativity.
    Invoked in Sections 3.2 and 3.4 based on [17-20, 34]; the paper treats linear projections as sufficient for semantic manipulation.
  • domain assumption Self-BLEU and Distinct-2 are valid measures of creativity and diversity.
    Used as primary creativity metrics in Section 4.1; the paper does not justify these surrogates beyond citing [41].
  • domain assumption The 100 selected samples per task are representative of the full datasets.
    Section 4.1 states samples are selected "to encourage divergent thinking" without a sampling protocol, implying a biased subsample.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Breaking Thought Patterns: A Multi-Dimensional Reasoning Framework for LLMs." pith.science (2026). https://pith.science/paper/7UDAQK3X

@misc{pith2026250613192,
  author       = {Pith},
  title        = {Pith review of: Breaking Thought Patterns: A Multi-Dimensional Reasoning Framework for LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7UDAQK3X}},
  note         = {Machine review of arXiv:2506.13192}
}
read the original abstract

Large language models (LLMs) are often constrained by rigid reasoning processes, limiting their ability to generate creative and diverse responses. To address this, a novel framework called LADDER is proposed, combining Chain-of-Thought (CoT) reasoning, Mixture of Experts (MoE) models, and multi-dimensional up/down-sampling strategies which breaks the limitations of traditional LLMs. First, CoT reasoning guides the model through multi-step logical reasoning, expanding the semantic space and breaking the rigidity of thought. Next, MoE distributes the reasoning tasks across multiple expert modules, each focusing on specific sub-tasks. Finally, dimensionality reduction maps the reasoning outputs back to a lower-dimensional semantic space, yielding more precise and creative responses. Extensive experiments across multiple tasks demonstrate that LADDER significantly improves task completion, creativity, and fluency, generating innovative and coherent responses that outperform traditional models. Ablation studies reveal the critical roles of CoT and MoE in enhancing reasoning abilities and creative output. This work contributes to the development of more flexible and creative LLMs, capable of addressing complex and novel tasks.

Figures

Figures reproduced from arXiv: 2506.13192 by the authors.

Figure 1
Figure 1. Comparative performance of large language models (LLMs) on advertising-copy and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Schematic diagram of the LADDER architecture.Left: the task is tokenized to extract [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Comprehensive Performance Comparison between LADDER and Mainstream Chinese [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: User Voting Proportions for Generated Responses Across Different Models and Tasks [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 13 canonical work pages

  1. [1]

    Attention is all you need

    Ashish Vaswani et al. “Attention is all you need”. In:Proceedings of the 31st International Conference on Neural Information Processing Systems. NIPS’17. Long Beach, California, USA: Curran Associates Inc., 2017, pp. 6000–6010.isbn: 9781510860964

    work page 2017

  2. [2]

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

    Jacob Devlin et al. “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding”. In:Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers).Ed.byJillBurstein,ChristyDoran,andThamarSolorio.Minneapolis,Minnesota: As...

    doi:10.18653/v1/n19- 2019

  3. [3]

    OpenAI et al.GPT-4 Technical Report. 2024. arXiv: 2303 . 08774 [cs.CL].url: https : //arxiv.org/abs/2303.08774

    arXiv 2024

  4. [4]

    Aaron Grattafiori et al.The Llama 3 Herd of Models. 2024. arXiv:2407.21783 [cs.AI].url: https://arxiv.org/abs/2407.21783

    arXiv 2024

  5. [5]

    Mor Geva et al.Transformer Feed-Forward Layers Are Key-Value Memories. 2021. arXiv: 2012.14913 [cs.CL].url:https://arxiv.org/abs/2012.14913

    arXiv 2021

  6. [6]

    Stevenson et al.Can Large Language Models generalize analogy solving like people can?2025

    Claire E. Stevenson et al.Can Large Language Models generalize analogy solving like people can?2025. arXiv:2411.02348 [cs.AI].url:https://arxiv.org/abs/2411.02348

    arXiv 2025

  7. [7]

    Rae et al.Scaling Language Models: Methods, Analysis & Insights from Training Gopher

    Jack W. Rae et al.Scaling Language Models: Methods, Analysis & Insights from Training Gopher. 2022. arXiv:2112.11446 [cs.CL].url:https://arxiv.org/abs/2112.11446

    arXiv 2022

  8. [8]

    DeepSeek-AI et al.DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. 2025. arXiv:2501.12948 [cs.CL].url:https://arxiv.org/abs/2501.12948

    arXiv 2025

Show all 49 references
  1. [9]

    Jason Wei et al.Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

  2. [10]

    Noam Shazeer et al.Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of- Experts Layer. 2017. arXiv:1701.06538 [cs.LG].url: https://arxiv.org/abs/1701.06538

    2017 arXiv

  3. [11]

    Denny Zhou et al.Least-to-Most Prompting Enables Complex Reasoning in Large Language Models. 2023. arXiv:2205.10625 [cs.AI].url:https://arxiv.org/abs/2205.10625

    2023 arXiv

  4. [12]

    Takeshi Kojima et al.Large Language Models are Zero-Shot Reasoners. 2023. arXiv:2205.11916 [cs.CL].url:https://arxiv.org/abs/2205.11916

    2023 arXiv

  5. [13]

    Shunyu Yao et al.Tree of Thoughts: Deliberate Problem Solving with Large Language Models

  6. [14]

    Zhuosheng Zhang et al.Multimodal Chain-of-Thought Reasoning in Language Models. 2024. arXiv:2302.00923 [cs.CL].url:https://arxiv.org/abs/2302.00923

    2024 arXiv

  7. [15]

    arXiv:2305.10601 [cs.CL].url:https://arxiv.org/abs/2305.10601

    arXiv

  8. [16]

    Jiang et al.Mixtral of Experts

    Albert Q. Jiang et al.Mixtral of Experts. 2024. arXiv:2401.04088 [cs.LG] .url: https: //arxiv.org/abs/2401.04088. 14

    2024 arXiv

  9. [17]

    Yanqi Zhou et al.Mixture-of-Experts with Expert Choice Routing. 2022. arXiv:2202.09368 [cs.LG].url:https://arxiv.org/abs/2202.09368

    2022 arXiv

  10. [18]

    Yukun Ma and Erik Cambria.Concept-Based Embeddings for Natural Language Processing

  11. [19]

    Vikas Raunak.Simple and Effective Dimensionality Reduction for Word Embeddings. 2017. arXiv:1708.03629 [cs.CL].url:https://arxiv.org/abs/1708.03629

    2017 arXiv

  12. [20]

    Jesse Mu, Xiang Lisa Li, and Noah Goodman.Learning to Compress Prompts with Gist Tokens

  13. [21]

    Predictability and Surprise in Large Generative Models

    Deep Ganguli et al. “Predictability and Surprise in Large Generative Models”. In:2022 ACM Conference on Fairness Accountability and Transparency. FAccT ’22. ACM, June 2022, pp. 1747–1764.doi: 10 . 1145 / 3531146 . 3533229.url: http : / / dx . doi . org / 10 . 1145 / 3531146.3533229

    2022

  14. [22]

    Word Re-Embedding via Manifold Dimensionality Retention

    Souleiman Hasan and Edward Curry. “Word Re-Embedding via Manifold Dimensionality Retention”. In:Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. Ed. by Martha Palmer, Rebecca Hwa, and Sebastian Riedel. Copenhagen, Denmark: Association for...

    2017 doi

  15. [23]

    Brown et al.Language Models are Few-Shot Learners

    Tom B. Brown et al.Language Models are Few-Shot Learners. 2020. arXiv:2005 . 14165 [cs.CL].url:https://arxiv.org/abs/2005.14165

    2020 arXiv

  16. [24]

    Alethea Power et al.Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets. 2022. arXiv:2201.02177 [cs.LG].url:https://arxiv.org/abs/2201.02177

    2022 arXiv

  17. [25]

    Rishi Bommasani et al.On the Opportunities and Risks of Foundation Models. 2022. arXiv: 2108.07258 [cs.LG].url:https://arxiv.org/abs/2108.07258

    2022 arXiv

  18. [26]

    Jason Wei et al.Emergent Abilities of Large Language Models. 2022. arXiv:2206 . 07682 [cs.CL].url:https://arxiv.org/abs/2206.07682

    2022 arXiv

  19. [27]

    Yuntao Bai et al.Constitutional AI: Harmlessness from AI Feedback. 2022. arXiv:2212.08073 [cs.CL].url:https://arxiv.org/abs/2212.08073

    2022 arXiv

  20. [28]

    Learning Hierarchy-Aware Knowledge Graph Embeddings for Link Prediction

    Zhanqiu Zhang et al. “Learning Hierarchy-Aware Knowledge Graph Embeddings for Link Prediction”. In:Proceedings of the AAAI Conference on Artificial Intelligence34.03 (Apr. 2020), pp. 3065–3072.doi: 10.1609/aaai.v34i03.5701.url: https://ojs.aaai.org/index. php/AAAI/article/view/5701

    2020 doi

  21. [29]

    Multi-level Attention-enhanced Learning for Fine-Grained Visual Classifi- cation

    Kaifeng Ding et al. “Multi-level Attention-enhanced Learning for Fine-Grained Visual Classifi- cation”. In:CSAI. 2024, pp. 105–111.url:https://doi.org/10.1145/3709026.3709034. 15

    2024

  22. [30]

    Yuko Nakagi et al.Triple Phase Transitions: Understanding the Learning Dynamics of Large Language Models from a Neuroscience Perspective. 2025. arXiv:2502.20779 [cs.CL].url: https://arxiv.org/abs/2502.20779

    2025 arXiv

  23. [31]

    Carlos Riquelme et al.Scaling Vision with Sparse Mixture of Experts. 2021. arXiv:2106.05974 [cs.CV].url:https://arxiv.org/abs/2106.05974

    2021 arXiv

  24. [32]

    Xiaofeng Zhang et al.Mixture of Attention Heads: Selecting Attention Heads Per Token. 2022. arXiv:2210.05144 [cs.CL].url:https://arxiv.org/abs/2210.05144

    2022 arXiv

  25. [33]

    DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of- Experts Language Models

    Damai Dai et al. “DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of- Experts Language Models”. In:Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Ed. by Lun-Wei Ku, Andre Martins, and Vivek Srikum...

    2024 doi

  26. [34]

    Building the Directed Semantic Graph for Coherent Long Text Generation

    Ziao Wang, Xiaofeng Zhang, and Hongwei Du. “Building the Directed Semantic Graph for Coherent Long Text Generation”. In:Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. Ed. by Marie-Francine Moens et al. Online and Punta Cana, Dominican R...

    2021 doi

  27. [35]

    Hu et al.LoRA: Low-Rank Adaptation of Large Language Models

    Edward J. Hu et al.LoRA: Low-Rank Adaptation of Large Language Models. 2021. arXiv: 2106.09685 [cs.CL].url:https://arxiv.org/abs/2106.09685

    2021 arXiv

  28. [36]

    ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks

    Jiasen Lu et al. “ViLBERT: Pretraining Task-Agnostic Visiolinguistic Representations for Vision-and-Language Tasks”. In:Advances in Neural Information Processing Systems. Ed. by H. Wallach et al. Vol. 32. Curran Associates, Inc., 2019.url:https://proceedings.neurips. cc/paper_...

    2019

  29. [37]

    Hierarchical Neural Story Generation

    Angela Fan, Mike Lewis, and Yann Dauphin. “Hierarchical Neural Story Generation”. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Ed. by Iryna Gurevych and Yusuke Miyao. Melbourne, Australia: Association for ...

    2018 doi

  30. [38]

    Zelin Zang et al.DMT-HI: MOE-based Hyperbolic Interpretable Deep Manifold Transformation for Unspervised Dimensionality Reduction. 2024. arXiv:2410.19504 [cs.LG] .url: https: //arxiv.org/abs/2410.19504

    2024 arXiv

  31. [39]

    May 2023

    Xuechen Li et al.AlpacaEval: An Automatic Evaluator of Instruction-following Models.https: //github.com/tatsu-lab/alpaca_eval. May 2023

    2023

  32. [40]

    Tian Liang et al.Encouraging Divergent Thinking in Large Language Models through Multi- Agent Debate. 2024. arXiv:2305.19118 [cs.CL].url: https://arxiv.org/abs/2305.19118. 16

    2024 arXiv

  33. [41]

    Yaoming Zhu et al.Texygen: A Benchmarking Platform for Text Generation Models. 2018. arXiv:1802.01886 [cs.CL].url:https://arxiv.org/abs/1802.01886

    2018 arXiv

  34. [42]

    CommonsenseQA: A Question Answering Challenge Targeting Common- sense Knowledge

    Alon Talmor et al. “CommonsenseQA: A Question Answering Challenge Targeting Common- sense Knowledge”. In:Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Pap...

    2019 doi

  35. [43]

    Long Ouyang et al.Training language models to follow instructions with human feedback. 2022. arXiv:2203.02155 [cs.CL].url:https://arxiv.org/abs/2203.02155

    2022 arXiv

  36. [44]

    An Yang et al.Qwen3 Technical Report. 2025. arXiv:2505 . 09388 [cs.CL].url: https : //arxiv.org/abs/2505.09388

    2025 arXiv

  37. [45]

    Team GLM et al.ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools. 2024. arXiv:2406.12793 [cs.CL].url:https://arxiv.org/abs/2406.12793. 17

    2024 arXiv

  38. [46]

    Tianyi Zhang et al.BERTScore: Evaluating Text Generation with BERT. 2020. arXiv:1904. 09675 [cs.CL].url:https://arxiv.org/abs/1904.09675

    2020 arXiv

  39. [2018]

    arXiv:1807.05519 [cs.CL].url:https://arxiv.org/abs/1807.05519

    arXiv

  40. [2023]

    arXiv:2201.11903 [cs.CL].url:https://arxiv.org/abs/2201.11903

    arXiv

  41. [2024]

    arXiv:2304.08467 [cs.CL].url:https://arxiv.org/abs/2304.08467

    arXiv

Pith tools

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