Thank you for reading. For more details, please check out our paper: https://t.co/yzbBa0HobA (5/5)

This work highlights a critical gap in the spatial reasoning capabilities of current models, identifying key failure modes like repetition. We provide a rigorous new framework for future research. (4/5)

Our benchmark uses two formal regex tasks: Minimization & Equivalence. The empirical results are striking—even LLMs and Reasoning Models (LRMs) show a significant performance drop, especially on minimization, where most fail to produce even an equivalent expression.(3/5)

How do LLMs perform when faced with problems requiring massive search space exploration under memory constraints? We introduce RegexPSPACE, the first benchmark designed to test this, moving beyond NP-hard tasks to the more challenging PSPACE-complete class. (2/5)

📣 Excited to share our new paper "RegexPSPACE: A Benchmark for Evaluating LLM Reasoning on PSPACE-complete Regex Problems"! (1/5) Paper: https://t.co/A3asGnRYzU Code: https://t.co/jTbNyqw7BN #AIResearch #Regex #PSPACE-complete #Benchmark #ReasoningTasks

🚧 What’s next for #TCProF? Future directions include extending to dynamic complexity calculation, space-complexity prediction, generalizing across more programming languages, and adapting to zero-shot scenarios. Excited for more advances ahead! 🌐📌

📈 #TCProF achieves groundbreaking performance, improving accuracy by over 60% compared to standard self-training and outperforming state-of-the-art methods like JointMatch by up to 131%! Reliable even with minimal labeled data. 💪

🧩 At the heart of #TCProF is our symbolic module—leveraging regex matching and AST parsing to systematically identify loops, recursions, and complexity-critical patterns, enabling precise pseudo-labeling without neural biases! 🔍🔗

📌Introducing #TCProF, our innovative SSL framework tailored for code complexity prediction! By effectively combining data augmentation, symbolic analysis, and co-training, TCProF robustly addresses data scarcity, greatly boosting prediction accuracy. 📊✨

🚀Just released! We will introduce "TCProF: Time-Complexity Prediction SSL Framework" at #NAACL2025. Ideal for programming competitions and code education! 🎉💻 Paper: https://t.co/fiOhQ44e8q Code: https://t.co/DWJIqJwhhc #AIResearch #CodeTimeComplexity #SemiSupervisedLearning

🔍 Curious about improving implicit hate speech detection just through data selection? Our paper introduces CONELA, showing how selecting training data based on human agreement patterns & model dynamics can significantly boost performance - without changing model architecture!

@coling2025 @dk_search_ai

@coling2025

📣 Excited to share our new paper "Analyzing Offensive Language Dataset Insights from Training Dynamics and Human Agreement Level" at #COLING2025! 🎉 Paper: https://t.co/y0kog2RdSa CONELA Code: https://t.co/mEWHMGBKeq #NLP2025 #dk_search_ai #finetuning #hatespeechdetection

Paper: https://t.co/HfTKYpl3P5 Github:

That’s the dataset in a nutshell! This work created by me (Seung-Yeop Baik) with the help of Joonghyuk Hahn, Jungin Kim, Mingi Jeon, Aditi, Yo-Sub Han, and Sang-Ki Ko.If you want to access the dataset or get into further details, follow the links below to get the full picture!

To avoid misleading, the GPT in the figure is an accusation module for the human annotators to think over their reasoning.

https://t.co/Bhi6oD8T53

https://t.co/QofdeRQOM0

We share the annotation process of creating the dataset. CodeComplex leverages competitive programming submissions from Codeforces and builds on the CodeContests dataset developed by DeepMind.