ACL Accepts DataCanvas Research Paper Tackling the Challenge of Analogy Reasoning in Language Models

Recently, the DataCanvas research team’s paper “Can Language Models Serve as Analogy Annotators?” was accepted as a Findings paper by the Association for Computational Linguistics (ACL), one of the most influential international conferences in computational linguistics. This work is the first to systematically uncover key limitations of large language models (LLMs) in analogy reasoning tasks and proposes an innovative solution, offering important theoretical support for overcoming barriers to machine analogy reasoning.

Top Conference Recognition: The Rise of DataCanvas in AI Research

Founded in 1962, ACL is the longest-running and most influential international academic conference in the field of natural language processing (NLP), recognized as a Class A conference by the China Computer Federation (CCF). Each year, ACL showcases the most cutting-edge research and technological breakthroughs in the NLP community. Its rigorous peer review process and extremely low acceptance rate make it a gold standard for measuring the innovation and applicability of academic work. In 2025, ACL reached a record-breaking 8,000+ submissions.

The acceptance of DataCanvas’s paper into ACL 2025 underscores the growing strength of Chinese research in AI. Not long ago, the DataCanvas research team had two original works — “A Solvable Attention for Neural Scaling Laws” and “DyCAST: Learning Dynamic Causal Structure from Time Series” — accepted by the International Conference on Learning Representations (ICLR), one of AI’s top 3 conferences. DataCanvas’s track record of top-tier conference recognition also includes:

2022 (ICLR): Implicit Bias of Adversarial Training for Deep Neural Networks;

2023 (NeurIPS): Implicit Bias of (Stochastic) Gradient Descent for Rank-1 Linear Neural Network;

2024 (AAAI): Effects of Momentum in Implicit Bias of Gradient Flow for Diagonal Linear Networks.

Technical Impact: Advancing AI from Data Fitting to Logical Abstraction

Although LLMs excel in tasks such as text generation and question answering, their abilities in analogy reasoning have long been underexplored. To address this gap, the DataCanvas team proposed a multi-stage progressive analogy reasoning framework called A3E. This framework guides models step-by-step to decompose analogy tasks, integrating contextual semantics with logical constraints, thereby significantly improving their ability to understand analogy relationships.

Experiments show that A3E enables LLMs to achieve human-expert-level quality in analogy annotation for the first time, offering a scalable technical pathway to overcome the cognitive bottleneck in machine analogy reasoning.

The innovation lies not only in exposing the inherent shortcomings of LLMs in analogy tasks but also in proving the feasibility of enabling machines to perform high-order reasoning through methodological advances. The general design of A3E can be extended to domains such as scientific discovery, intelligent education, and business decision-making — for example, by enabling automated analogy mining to aid interdisciplinary research, or by generating educational assessment content based on logical relationships. In essence, this breakthrough represents a key step in moving language models from data-driven shallow semantic understanding toward logic-driven deep cognition.

The inclusion of this work in ACL reflects the international academic community’s recognition of DataCanvas’s technological innovation and provides valuable insights into exploring the cognitive capabilities of language models. Looking ahead, DataCanvas will continue to explore the integration of language models with causal reasoning and cross-modal cognition, driving machine intelligence to simulate human high-order thinking in complex scenarios, and empowering intelligent transformation in education, research, healthcare, and beyond.