Pith. sign in

REVIEW 4 major objections 6 minor 61 references

Natural language inference improves when models stack token-by-token interactions from every transformer layer instead of relying on the final layer alone.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 19:06 UTC pith:INOIRIM4

load-bearing objection Solid incremental NLI architecture paper with real tables and ablations, but the multi-granularity story is under-isolated and the write-up has sloppy spots. the 4 major comments →

arxiv 2606.05181 v2 pith:INOIRIM4 submitted 2026-04-18 cs.CL cs.AI

Multi-Granularity Reasoning for Natural Language Inference

Chunling Xi , Di Liang This is my paper
classification cs.CL cs.AI
keywords Multi-Granularity ReasoningNatural Language InferenceSemantic InteractionTransformer LayersDenseNetSentence Pair MatchingNLI Robustness
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Standard transformer models for natural language inference mostly use only their last-layer token representations. The paper argues this collapses fine-grained word cues, phrase structure, and higher-level meaning into one mixed space and therefore misses hierarchical semantic interactions. MGRN builds an interaction tensor by taking the element-wise product of premise and hypothesis token states at every layer, stacks those products, and runs the result through DenseNet before classification. The authors report consistent gains over strong BERT and RoBERTa baselines on SNLI, MultiNLI, paraphrase-style tasks, and several adversarial suites. A sympathetic reader cares because the method offers a simple, reusable way to recover intermediate-layer signals that ordinary fine-tuning throws away.

Core claim

Explicitly constructing multi-layer interaction tensors via element-wise products of premise and hypothesis token representations, stacking them across all transformer layers, and extracting features with DenseNet yields a multi-granularity reasoning space that systematically outperforms final-layer-only and several knowledge-enhanced baselines on NLI and related sentence-pair benchmarks.

What carries the argument

The multi-layer interaction tensor M: at each transformer layer the model forms the element-wise product of every premise token with every hypothesis token, then stacks these products across layers into a four-dimensional tensor that DenseNet processes for classification.

Load-bearing premise

The paper assumes that stacking element-wise products of token states from all transformer layers and feeding them to DenseNet produces a faithful multi-granularity reasoning space that is systematically better than final-layer or single-layer alternatives.

What would settle it

Retrain the identical pipeline using only the final-layer interaction matrix (or randomly chosen layers) and test whether the accuracy and robustness gains on MultiNLI, SNLI, and the reported adversarial transformations disappear.

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

If this is right

  • Fusing intermediate-layer interactions should raise accuracy on NLI and sentence-pair tasks relative to final-layer-only fine-tuning.
  • Hierarchical interaction modeling can reduce dependence on external lexical or syntactic knowledge for certain adversarial perturbations.
  • The same stacking-plus-DenseNet pattern can be reused for paraphrase identification framed as binary NLI.
  • Removing multi-layer stacking, the interaction matrix, or DenseNet each lowers accuracy, so the three components are jointly required for the reported gains.

Where Pith is reading between the lines

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

  • Intermediate-layer products may already carry enough compositional signal that explicit syntactic parsers become less necessary for many NLI cases.
  • The same construction could transfer to other pairwise reasoning tasks such as fact verification or multi-hop QA where both shallow and deep cues matter.
  • If some layers prove noisy, selective layer weighting or gating would be a natural refinement that isolates which granularities actually help.
  • DenseNet feature reuse may be doing as much hierarchical work as the interaction construction itself; swapping the extractor would separate the two contributions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes Multi-Granularity Reasoning Network (MGRN) for natural language inference. Starting from a pretrained transformer (BERT/RoBERTa), it extracts token representations from all L layers, forms per-layer interaction tensors via element-wise products M^(l)_i,j = h^(l,i)_1 ⊙ h^(l,j)_2, stacks them into a 4-D tensor M ∈ R^{n×m×d×L}, and feeds M through DenseNet blocks before a softmax classifier. The authors claim this progressive multi-layer interaction mimics human multi-granularity reasoning and yields consistent gains over strong PLM baselines on SNLI, MultiNLI, GLUE-style sentence-pair tasks, and robustness suites (Tables I–III), with ablations in Table II.

Significance. If the hierarchical-interaction story holds, the work offers a simple, architecture-level way to exploit intermediate transformer layers for NLI without external knowledge graphs or syntax parsers. The reported average gains (roughly 0.5–1.5 points over BERT/RoBERTa bases and several knowledge-enhanced variants) and the robustness numbers under TextFlint-style perturbations would be of practical interest to the NLI community. The contribution is empirical and incremental rather than foundational: element-wise cross-sentence products and DenseNet-style feature reuse are known ingredients; the novelty lies in stacking all layers and claiming multi-granularity superiority. Strengths include multi-dataset evaluation, an ablation table, and qualitative cases. The significance is therefore moderate and contingent on cleaner isolation of the multi-layer claim and resolution of reporting inconsistencies.

major comments (4)
  1. The central architectural claim (§III.C–E, abstract, §I) is that stacking element-wise products from all L layers into M and processing them with DenseNet yields a multi-granularity reasoning advantage over final-layer baselines. Table II’s only direct support is “w/o multi-layer interaction” (85.1→84.6 matched). The drop is tiny, the replacement is unspecified (last layer only? mean of layers? random subset?), and there is no controlled experiment that keeps the identical DenseNet head while using solely the final-layer interaction tensor M^(L). Without that isolation, the hierarchical-advantage narrative is not established; gains could arise from the interaction matrix + DenseNet capacity applied to the last layer alone, or from hyperparameter differences.
  2. §IV (“Experiments Setting”) repeatedly names the model “CIRN” (“performance of CIRN”, “our proposed CIRN”) while the title, abstract, method, and tables use MGRN. This is not a typographical slip confined to one sentence; it appears in the experimental protocol description itself and undermines confidence that the reported numbers correspond to the architecture defined in §III.
  3. Table I, RoBERTa-Base row: RTE jumps from 73.6 to 82.5 (+8.9 points) under MGRN while SNLI remains essentially flat (90.8→91.2) and several other columns move by <1 point. No error bars, no multiple-run statistics, and no analysis of this outlier appear in §V. An unexplained double-digit gain on a small dataset (RTE) while the primary NLI benchmarks barely move is load-bearing for the “consistent outperformance” claim and requires either multi-seed reporting or an error analysis.
  4. §III.E applies DenseNet (originally defined for 2-D/3-D image feature maps with channel-wise concatenation) directly to the 4-D tensor M ∈ R^{n×m×d×L} without specifying the convolution kernel shapes, how the layer dimension L is treated (extra channel? separate spatial axis?), growth rates, number of Dense Blocks, or transition-layer design. These free parameters are listed nowhere; reproducibility and the claim that DenseNet is the appropriate inductive bias for this interaction tensor therefore cannot be assessed.
minor comments (6)
  1. Abstract and §I assert that final-layer representations “entangle or dilute” fine-grained cues; a short citation or layer-wise probing reference (or a simple diagnostic) would ground this motivation.
  2. Equation (3) and surrounding text use both “interaction matrix” and “interaction tensor”; consistent terminology would help.
  3. Table I header mixes “Sci”, “SICK”, “Twi” without expansion or citation; full dataset names and sizes belong in §IV.
  4. Table III caption lists many transformation acronyms; a one-line definition or pointer to TextFlint would aid readers.
  5. Several references (e.g., [3], [21], [22], [29]–[61]) are concurrent arXiv notes sharing co-authors; while not circular for the numbers, a clearer separation of prior published baselines from concurrent work would improve transparency.
  6. Typographical issues: “Neural Language Inference” (§II heading), duplicated citation “[46], [46]” (§I), and occasional missing spaces around math.

Circularity Check

0 steps flagged

No significant circularity: empirical multi-layer interaction architecture evaluated on public NLI benchmarks, with no derivation that reduces by construction to its inputs.

full rationale

The paper proposes an architecture (multi-layer element-wise interaction tensors M^(l) stacked into M then fed to DenseNet) and reports empirical accuracy gains on SNLI, MultiNLI, QQP and related public datasets (Tables I–III). There is no mathematical derivation claiming to force a first-principles result; the “prediction” is simply the classifier output after standard fine-tuning. Ablations and baselines are experimental comparisons, not algebraic identities. Self-citations to concurrent Liang-group arXiv notes appear in Related Work and the reference list but are not load-bearing for uniqueness theorems, forced ansatzes, or fitted parameters renamed as predictions. The central claim therefore rests on external benchmarks rather than circular reduction, so the circularity score is zero.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The central claim rests on standard transformer fine-tuning practice plus three modeling choices that are not independently validated outside this paper: (i) that every intermediate layer’s token states should be retained, (ii) that element-wise multiplication is the right interaction, and (iii) that DenseNet is the right aggregator. No free physical constants appear; free parameters are architectural and training choices. Invented entity is the MGRN pipeline itself.

free parameters (3)
  • DenseNet block depth / growth rate / transition design
    Not specified numerically; chosen by authors and load-bearing for the reported F representation.
  • Which of the L BERT layers enter the stack M
    Paper writes all L layers; no ablation of layer subsets beyond “last layer only,” so the full stack is an un-tuned design choice that affects the claim.
  • Fine-tuning hyperparameters (LR, batch, epochs, dropout)
    Never reported; any reproduction must re-search them, and they can move NLI scores by more than the claimed gains.
axioms (4)
  • domain assumption Intermediate transformer layers encode complementary granularities (lexical → phrasal → abstract) that remain useful after fine-tuning.
    Stated in abstract and §I–II; used to justify multi-layer stacking. Supported by prior probing literature but not re-verified here.
  • ad hoc to paper Element-wise product h1 ⊙ h2 is a sufficient interaction operator for cross-sentence reasoning at every layer.
    Eq. (3); no comparison to concatenation, bilinear, or attention-based alternatives inside the paper.
  • ad hoc to paper DenseNet feature reuse is an appropriate inductive bias for n×m×d×L interaction tensors in NLI.
    §III.E; motivated by image-processing DenseNet, not derived for language.
  • domain assumption Standard NLI label definitions (entailment/contradiction/neutral) and public benchmark splits are valid evaluation targets.
    Implicit throughout §IV–V.
invented entities (1)
  • Multi-Granularity Reasoning Network (MGRN) interactive tensor + DenseNet pipeline no independent evidence
    purpose: Aggregate hierarchical premise–hypothesis interactions for NLI classification.
    The named system is the paper’s contribution; independent evidence is only the authors’ own tables, not external replications.

pith-pipeline@v1.1.0-grok45 · 16838 in / 3067 out tokens · 35102 ms · 2026-07-12T19:06:47.823235+00:00 · methodology

0 comments
read the original abstract

Natural Language Inference (NLI) is a fundamental task in natural language understanding that requires determining the logical relationship between a premise and a hypothesis. Despite the remarkable success of transformer-based pre-trained models, most existing approaches primarily rely on the final-layer token representations, which are often insufficient for capturing the complex and hierarchical semantic interactions required for effective reasoning. In particular, fine-grained lexical cues, phrasal compositions, and higher-level contextual semantics are typically entangled or diluted in a single representation space. To address these limitations, we propose a novel \emph{Multi-Granularity Reasoning Network} (MGRN) that explicitly leverages hierarchical semantic features within an interactive reasoning space. The proposed framework mimics the human cognitive process of language understanding, which naturally progresses from shallow lexical matching to deeper semantic abstraction and logical reasoning. By integrating semantic information across multiple granularities in a progressive and structured manner, MGRN is able to uncover intricate semantic relationships underlying natural language expressions. Extensive experiments on multiple public benchmarks demonstrate that MGRN consistently outperforms strong baseline models, validating the effectiveness and robustness of the proposed approach.

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

61 extracted references · 20 linked inside Pith

  1. [1]

    What is the Jeopardy Model? A Quasi-Synchronous Grammar for QA,

    M. Wang, N. A. Smith, and T. Mitamura, “What is the Jeopardy Model? A Quasi-Synchronous Grammar for QA,” inProceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL), pp. 22–32, 2007

    2007

  2. [2]

    A Large Anno- tated Corpus for Learning Natural Language Inference,

    S. R. Bowman, G. Angeli, C. Potts, and C. D. Manning, “A Large Anno- tated Corpus for Learning Natural Language Inference,” inProceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 632–642, 2015

    2015

  3. [3]

    Chain of evidence: Pixel-level visual attribution for iterative retrieval-augmented generation.arXiv preprint arXiv:2605.01284, 2026

    Peiyang Liu, Ziqiang Cui, Xi Wang, Di Liang, and Wei Ye. Chain of evidence: Pixel-level visual attribution for iterative retrieval-augmented generation.arXiv preprint arXiv:2605.01284, 2026

    Pith/arXiv arXiv 2026

  4. [4]

    Deep Residual Learning for Image Recognition,

    K. He, X. Zhang, R. Shaoqing, and J. Sun, “Deep Residual Learning for Image Recognition,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016

    2016

  5. [5]

    Enhanced LSTM for Natural Language Inference,

    Q. Chen, X. Zhu, Z. Ling, S. Wei, H. Jiang, and D. Inkpen, “Enhanced LSTM for Natural Language Inference,”arXiv preprint arXiv:1609.06038, 2016

    Pith/arXiv arXiv 2016

  6. [6]

    Adversarial Examples for Evaluating Reading Comprehension Systems,

    R. Jia and P. Liang, “Adversarial Examples for Evaluating Reading Comprehension Systems,” inProceedings of the 2017 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 2021–2031, 2017

    2017

  7. [7]

    Su- pervised Learning of Universal Sentence Representations from Natural Language Inference Data,

    A. Conneau, D. Kiela, H. Schwenk, L. Barrault, and A. Bordes, “Su- pervised Learning of Universal Sentence Representations from Natural Language Inference Data,”arXiv preprint arXiv:1705.02364, 2017

    Pith/arXiv arXiv 2017

  8. [8]

    Natural Language Inference over Interaction Space,

    Y . Gong, H. Luo, and J. Zhang, “Natural Language Inference over Interaction Space,”arXiv preprint arXiv:1709.04348, 2017

    Pith/arXiv arXiv 2017

  9. [9]

    Learning to Compose Task-Specific Tree Structures,

    J. Choi, K. M. Yoo, and S.-g. Lee, “Learning to Compose Task-Specific Tree Structures,” inProceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, 2018

    2018

  10. [10]

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

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,” in Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics (NAACL-HLT), pp. 4171– 4186, 2019

    2019

  11. [11]

    RoBERTa: A Robustly Optimized BERT Pretraining Approach,

    Y . Liu, M. Ott, N. Goyal, et al., “RoBERTa: A Robustly Optimized BERT Pretraining Approach,”arXiv preprint arXiv:1907.11692, 2019

    Pith/arXiv arXiv 1907

  12. [12]

    XLNet: Generalized Autoregressive Pretraining for Language Understanding,

    Z. Yang, Z. Dai, Y . Yang, J. Carbonell, R. Salakhutdinov, and Q. V . Le, “XLNet: Generalized Autoregressive Pretraining for Language Understanding,” inAdvances in Neural Information Processing Systems (NeurIPS), pp. 5753–5763, 2019

    2019

  13. [13]

    Asynchronous deep interaction network for natural language inference,

    D. Liang, F. Zhang, Q. Zhang, and X.-J. Huang, “Asynchronous deep interaction network for natural language inference,” inProceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP-IJCNLP), pp. 2692–2700, 2019

    2019

  14. [14]

    Graph convolutional networks for text classification,

    L. Yao, C. Mao, and Y . Luo, “Graph convolutional networks for text classification,” inProceedings of the AAAI Conference on Artificial Intelligence, 2019

    2019

  15. [15]

    Sentence-BERT: Sentence embeddings using siamese BERT-networks,

    N. Reimers and I. Gurevych, “Sentence-BERT: Sentence embeddings using siamese BERT-networks,” inProceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2019

    2019

  16. [16]

    Semantics-Aware BERT for Language Understanding,

    Z. Zhang, Y . Wu, H. Zhao, L. Zuchao, S. Zhang, X. Zhou, and X. Zhou, “Semantics-Aware BERT for Language Understanding,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 34, pp. 9628– 9635, 2020

    2020

  17. [17]

    Enhanced-RCNN: An Efficient Method for Learning Sentence Similarity,

    S. Peng, H. Cui, N. Xie, S. Li, J. Zhang, and X. Li, “Enhanced-RCNN: An Efficient Method for Learning Sentence Similarity,” inProceedings of The Web Conference (WWW), pp. 2500–2506, 2020

    2020

  18. [18]

    CharBERT: Character-Aware Pre-Trained Language Model,

    W. Ma, Y . Cui, C. Si, T. Liu, S. Wang, and G. Hu, “CharBERT: Character-Aware Pre-Trained Language Model,” inProceedings of the 28th International Conference on Computational Linguistics (COLING), pp. 39–50, 2020

    2020

  19. [19]

    Heterogeneous Graph Transformer for Graph-to-Text Generation,

    Z. Hu, B. Tan, and J. Luo, “Heterogeneous Graph Transformer for Graph-to-Text Generation,” inAdvances in Neural Information Process- ing Systems (NeurIPS), pp. 1–12, 2020

    2020

  20. [20]

    Graph neural networks for natural language processing: A survey,

    Q. Guo, X. Xiao, T. Ma, et al., “Graph neural networks for natural language processing: A survey,” inProceedings of the International Joint Conference on Artificial Intelligence (IJCAI), 2020

    2020

  21. [21]

    Parameter importance is not static: Evolving parameter isolation for supervised fine-tuning.arXiv preprint arXiv:2604.14010, 2026

    Zekai Lin, Chao Xue, Di Liang, Xingsheng Han, Peiyang Liu, Xianjie Wu, Lei Jiang, Yu Lu, Haibo Shi, Shuang Liang, et al. Parameter importance is not static: Evolving parameter isolation for supervised fine-tuning.arXiv preprint arXiv:2604.14010, 2026

    Pith/arXiv arXiv 2026

  22. [22]

    Learning from contrasts: Synthesizing reasoning paths from diverse search trajectories.arXiv preprint arXiv:2604.11365, 2026

    Peiyang Liu, Zhirui Chen, Xi Wang, Di Liang, Youru Li, Zhi Cai, and Wei Ye. Learning from contrasts: Synthesizing reasoning paths from diverse search trajectories.arXiv preprint arXiv:2604.11365, 2026

    Pith/arXiv arXiv 2026

  23. [23]

    Artificial Intelligence and Law: Risks, Research, and Limits,

    A.-L. Collomb, A. Conesa-Bes, and L. Lautman, “Artificial Intelligence and Law: Risks, Research, and Limits,”Artificial Intelligence and Law, vol. 28, 2020

    2020

  24. [24]

    Connecting the dots: Document-level relation extraction with edge-oriented graphs,

    F. Christopoulou, M. Miwa, and S. Ananiadou, “Connecting the dots: Document-level relation extraction with edge-oriented graphs,” inPro- ceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020

    2020

  25. [25]

    Enhanced Attentive Convolutional Neural Networks for Sentence Pair Modeling,

    S. Xu, S. E, and Y . Xiang, “Enhanced Attentive Convolutional Neural Networks for Sentence Pair Modeling,”Expert Systems with Applica- tions, 2020

    2020

  26. [26]

    Using Prior Knowledge to Guide BERT’s Attention in Semantic Textual Matching Tasks,

    T. Xia, Y . Wang, Y . Tian, and Y . Chang, “Using Prior Knowledge to Guide BERT’s Attention in Semantic Textual Matching Tasks,” in Proceedings of the Web Conference (WWW), pp. 2466–2475, 2021

    2021

  27. [27]

    TextFlint: Unified Multilingual Robustness Evaluation Toolkit for Natural Language Processing,

    T. Gui, X. Wang, Q. Zhang, et al., “TextFlint: Unified Multilingual Robustness Evaluation Toolkit for Natural Language Processing,” in Proceedings of the 59th Annual Meeting of the Association for Com- putational Linguistics (ACL), pp. 347–355, 2021

    2021

  28. [28]

    Explainaboard: An Explainable Leaderboard for NLP,

    P. Liu, J. Fu, Y . Xiao, W. Yuan, J. Chang, and Z. Geng, “Explainaboard: An Explainable Leaderboard for NLP,” inProceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), pp. 280–289, 2021

    2021

  29. [29]

    Hope: Hyperbolic rotary positional encoding for stable long-range dependency modeling in large language models.arXiv preprint arXiv:2509.05218, 2025

    Chang Dai, Hongyu Shan, Mingyang Song, and Di Liang. Hope: Hyperbolic rotary positional encoding for stable long-range dependency modeling in large language models.arXiv preprint arXiv:2509.05218, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    Cqg: A simple and effective controlled generation framework for multi-hop question generation

    Zichu Fei, Qi Zhang, Tao Gui, Di Liang, Sirui Wang, Wei Wu, and Xuan-Jing Huang. Cqg: A simple and effective controlled generation framework for multi-hop question generation. InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6896–6906, 2022

    2022

  31. [31]

    Decorl: Decoupling reasoning chains via parallel sub-step generation and cascaded reinforcement for interpretable and scalable rlhf

    Ziyuan Gao, Di Liang, Xianjie Wu, Philippe Morel, and Minlong Peng. Decorl: Decoupling reasoning chains via parallel sub-step generation and cascaded reinforcement for interpretable and scalable rlhf. InPro- ceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 30789–30797, 2026

    2026

  32. [32]

    Transferring from formal newswire domain with hypernet for twitter pos tagging

    Tao Gui, Qi Zhang, Jingjing Gong, Minlong Peng, Di Liang, Keyu Ding, and Xuan-Jing Huang. Transferring from formal newswire domain with hypernet for twitter pos tagging. InProceedings of the 2018 conference on empirical methods in natural language processing, pages 2540–2549, 2018

    2018

  33. [33]

    Comateformer: Combined at- tention transformer for semantic sentence matching.arXiv preprint arXiv:2412.07220, 2024

    Bo Li, Di Liang, and Zixin Zhang. Comateformer: Combined at- tention transformer for semantic sentence matching.arXiv preprint arXiv:2412.07220, 2024

    Pith/arXiv arXiv 2024

  34. [34]

    When safety becomes a vulnerability: Exploiting llm alignment homogeneity for transferable blocking in rag.arXiv preprint arXiv:2603.03919, 2026

    Junchen Li, Chao Qi, Rongzheng Wang, Qizhi Chen, Liang Xu, Di Liang, Bob Simons, and Shuang Liang. When safety becomes a vulnerability: Exploiting llm alignment homogeneity for transferable blocking in rag.arXiv preprint arXiv:2603.03919, 2026

    arXiv 2026

  35. [35]

    Local and global: Text matching via syntax graph calibration

    Liang Li, Qisheng Liao, Meiting Lai, Di Liang, and Shangsong Liang. Local and global: Text matching via syntax graph calibration. InICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 11571–11575. IEEE, 2024

    2024

  36. [36]

    TRACE: Discovering task-specific parameter via adaptation-aware prob- ing for continual fine-tuning.arXiv preprint arXiv:2605.31025, 2026

    Xiaosong Han, Ke Chen, Xindi Dai, Di Liang, Minlong Peng, Wei Pang, Fausto Giunchiglia, Xiaoyue Feng, Yonghao Liu, and Renchu Guan. TRACE: Discovering task-specific parameter via adaptation-aware prob- ing for continual fine-tuning.arXiv preprint arXiv:2605.31025, 2026

    Pith/arXiv arXiv 2026

  37. [37]

    Adaptive multi-attention network incorporating answer information for duplicate question detection

    Di Liang, Fubao Zhang, Weidong Zhang, Qi Zhang, Jinlan Fu, Minlong Peng, Tao Gui, and Xuanjing Huang. Adaptive multi-attention network incorporating answer information for duplicate question detection. In Proceedings of the 42nd international ACM SIGIR conference on research and development in information retrieval, pages 95–104, 2019

    2019

  38. [38]

    Who stole your data? a method for detecting unauthorized rag theft.arXiv preprint arXiv:2510.07728, 2025

    Peiyang Liu, Ziqiang Cui, Di Liang, and Wei Ye. Who stole your data? a method for detecting unauthorized rag theft.arXiv preprint arXiv:2510.07728, 2025

    arXiv 2025

  39. [39]

    Dpi: Exploiting parameter heterogeneity for interference-free fine-tuning.arXiv preprint arXiv:2601.17777, 2026

    Xiaoyu Liu, Xiaoyu Guan, Di Liang, and Xianjie Wu. Dpi: Exploiting parameter heterogeneity for interference-free fine-tuning.arXiv preprint arXiv:2601.17777, 2026

    arXiv 2026

  40. [40]

    Structural reward model: Enhancing interpretability, efficiency, and scalability in reward modeling

    Xiaoyu Liu, Di Liang, Hongyu Shan, Peiyang Liu, Yonghao Liu, Muling Wu, Yuntao Li, Xianjie Wu, Li Miao, Jiangrong Shen, et al. Structural reward model: Enhancing interpretability, efficiency, and scalability in reward modeling. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 672– 685, 2025

    2025

  41. [41]

    Resolving word vagueness with scenario-guided adapter for natural language inference.arXiv preprint arXiv:2405.12434, 2024

    Yonghao Liu, Mengyu Li, Di Liang, Ximing Li, Fausto Giunchiglia, Lan Huang, Xiaoyue Feng, and Renchu Guan. Resolving word vagueness with scenario-guided adapter for natural language inference.arXiv preprint arXiv:2405.12434, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    Time-aware multiway adaptive fusion network for temporal knowledge graph question answering

    Yonghao Liu, Di Liang, Fang Fang, Sirui Wang, Wei Wu, and Rui Jiang. Time-aware multiway adaptive fusion network for temporal knowledge graph question answering. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2023

    2023

  43. [43]

    Local and global: Temporal question answering via information fusion

    Yonghao Liu, Di Liang, Mengyu Li, Fausto Giunchiglia, Ximing Li, Sirui Wang, Wei Wu, Lan Huang, Xiaoyue Feng, and Renchu Guan. Local and global: Temporal question answering via information fusion. InIJCAI, pages 5141–5149, 2023

    2023

  44. [44]

    Searching for optimal subword tokenization in cross-domain ner.arXiv preprint arXiv:2206.03352, 2022

    Ruotian Ma, Yiding Tan, Xin Zhou, Xuanting Chen, Di Liang, Sirui Wang, Wei Wu, Tao Gui, and Qi Zhang. Searching for optimal subword tokenization in cross-domain ner.arXiv preprint arXiv:2206.03352, 2022

    Pith/arXiv arXiv 2022

  45. [45]

    Adaptive curriculum strategies: Stabilizing reinforcement learning for large language models

    Qi Qian, Muling Wu, Zisu Huang, Wenhao Liu, Changze Lv, Xiaohua Wang, Zhenghua Wang, Zhengkang Guo, Zhibo Xu, Lina Chen, et al. Adaptive curriculum strategies: Stabilizing reinforcement learning for large language models

  46. [46]

    Improving semantic matching through dependency-enhanced pre-trained model with adaptive fusion.arXiv preprint arXiv:2210.08471, 2022

    Jian Song, Di Liang, Rumei Li, Yuntao Li, Sirui Wang, Minlong Peng, Wei Wu, and Yongxin Yu. Improving semantic matching through dependency-enhanced pre-trained model with adaptive fusion.arXiv preprint arXiv:2210.08471, 2022

    Pith/arXiv arXiv 2022

  47. [47]

    Beyond semantic relevance: Counterfactual risk minimization for robust retrieval-augmented generation.arXiv preprint arXiv:2605.01302, 2026

    Peiyang Liu, Qiang Yan, Ziqiang Cui, Di Liang, Xi Wang, and Wei Ye. Beyond semantic relevance: Counterfactual risk minimization for robust retrieval-augmented generation.arXiv preprint arXiv:2605.01302, 2026

    Pith/arXiv arXiv 2026

  48. [48]

    Rethinking llm- driven heuristic design: Generating efficient and specialized solvers via dynamics-aware optimization.arXiv preprint arXiv:2601.20868, 2026

    Rongzheng Wang, Yihong Huang, Muquan Li, Jiakai Li, Di Liang, Bob Simons, Pei Ke, Shuang Liang, and Ke Qin. Rethinking llm- driven heuristic design: Generating efficient and specialized solvers via dynamics-aware optimization.arXiv preprint arXiv:2601.20868, 2026

    Pith/arXiv arXiv 2026

  49. [49]

    Dabert: Dual attention enhanced bert for semantic matching.arXiv preprint arXiv:2210.03454, 2022

    Sirui Wang, Di Liang, Jian Song, Yuntao Li, and Wei Wu. Dabert: Dual attention enhanced bert for semantic matching.arXiv preprint arXiv:2210.03454, 2022

    Pith/arXiv arXiv 2022

  50. [50]

    Not all parameters are created equal: Smart isolation boosts fine-tuning performance.arXiv preprint arXiv:2508.21741, 2025

    Yao Wang, Di Liang, and Minlong Peng. Not all parameters are created equal: Smart isolation boosts fine-tuning performance.arXiv preprint arXiv:2508.21741, 2025

    arXiv 2025

  51. [51]

    Pro- gressive mastery: Customized curriculum learning with guided prompt- ing for mathematical reasoning.arXiv preprint arXiv:2506.04065, 2025

    Muling Wu, Qi Qian, Wenhao Liu, Xiaohua Wang, Zisu Huang, Di Liang, LI Miao, Shihan Dou, Changze Lv, Zhenghua Wang, et al. Pro- gressive mastery: Customized curriculum learning with guided prompt- ing for mathematical reasoning.arXiv preprint arXiv:2506.04065, 2025

    Pith/arXiv arXiv 2025

  52. [52]

    Breaking size barrier: Enhancing reasoning for large-size table question answering

    Xianjie Wu, Di Liang, Jian Yang, Xianfu Cheng, LinZheng Chai, Tongliang Li, Liqun Yang, and Zhoujun Li. Breaking size barrier: Enhancing reasoning for large-size table question answering. InInter- national Conference on Database Systems for Advanced Applications, pages 241–256. Springer, 2025

    2025

  53. [53]

    Mmtablebench: A multi-level multimodal benchmark for reasoning and layout complexity in table qa

    Xianjie Wu, Xiaohang Xu, Tingyu Jiang, Jian Yang, Di Liang, Xianfu Cheng, Zhenhe Wu, Linzheng Chai, Wei Zhang, Jiaheng Liu, et al. Mmtablebench: A multi-level multimodal benchmark for reasoning and layout complexity in table qa. InProceedings of the ACM Web Conference 2026, pages 3881–3892, 2026

    2026

  54. [54]

    Tablebench: A comprehensive and complex benchmark for table ques- tion answering

    Xianjie Wu, Jian Yang, Linzheng Chai, Ge Zhang, Jiaheng Liu, Xe- ron Du, Di Liang, Daixin Shu, Xianfu Cheng, Tianzhen Sun, et al. Tablebench: A comprehensive and complex benchmark for table ques- tion answering. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 25497–25506, 2025

    2025

  55. [55]

    Unleashing potential of evidence in knowledge-intensive dialogue generation

    Xianjie Wu, Jian Yang, Tongliang Li, Shiwei Zhang, Yiyang Du, LinZheng Chai, Di Liang, and Zhoujun Li. Unleashing potential of evidence in knowledge-intensive dialogue generation. InICASSP 2025- 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2025

    2025

  56. [56]

    Question calibration and multi-hop modeling for temporal question answering

    Chao Xue, Di Liang, Pengfei Wang, and Jing Zhang. Question calibration and multi-hop modeling for temporal question answering. InProceedings of the AAAI Conference on Artificial Intelligence, vol- ume 38, pages 19332–19340, 2024

    2024

  57. [57]

    Dual path modeling for semantic matching by perceiving subtle conflicts

    Chao Xue, Di Liang, Sirui Wang, Jing Zhang, and Wei Wu. Dual path modeling for semantic matching by perceiving subtle conflicts. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2023

    2023

  58. [58]

    Robust lottery tickets for pre-trained language models.arXiv preprint arXiv:2211.03013, 2022

    Rui Zheng, Rong Bao, Yuhao Zhou, Di Liang, Sirui Wang, Wei Wu, Tao Gui, Qi Zhang, and Xuanjing Huang. Robust lottery tickets for pre-trained language models.arXiv preprint arXiv:2211.03013, 2022

    Pith/arXiv arXiv 2022

  59. [59]

    Parameter importance is not static: Evolving parameter isolation for supervised fine-tuning, 2026

    Zekai Lin, Chao Xue, Di Liang, Xingsheng Han, Peiyang Liu, Xianjie Wu, Lei Jiang, Yu Lu, Haibo Shi, Shuang Liang, and Minlong Peng. Parameter importance is not static: Evolving parameter isolation for supervised fine-tuning, 2026

    2026

  60. [60]

    Reason only when needed: Efficient generative reward modeling via model-internal uncertainty.arXiv preprint arXiv:2604.10072, 2026

    Chao Xue, Yao Wang, Mengqiao Liu, Di Liang, Xingsheng Han, Peiyang Liu, Xianjie Wu, Chenyao Lu, Lei Jiang, Yu Lu, et al. Reason only when needed: Efficient generative reward modeling via model-internal uncertainty.arXiv preprint arXiv:2604.10072, 2026

    Pith/arXiv arXiv 2026

  61. [61]

    Why supervised fine-tuning fails to learn: A systematic study of incomplete learning in large language models.arXiv preprint arXiv:2604.10079, 2026

    Chao Xue, Yao Wang, Mengqiao Liu, Di Liang, Xingsheng Han, Peiyang Liu, Xianjie Wu, Chenyao Lu, Lei Jiang, Yu Lu, et al. Why supervised fine-tuning fails to learn: A systematic study of incomplete learning in large language models.arXiv preprint arXiv:2604.10079, 2026

    Pith/arXiv arXiv 2026