✨ORION: Teaching Language Models to Reason Efficiently in the Language of Thought

📝 Summary:
ORION models compress reasoning into ultra-compressed structured tokens, inspired by Mentalese. This reduces reasoning steps by 4-16x, cuts inference latency by 5x, and training costs by 7-9x while maintaining high accuracy.

🔹 Publication Date: Published on Nov 28

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2511.22891
• PDF: https://arxiv.org/pdf/2511.22891

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LLM #AI #AIReasoning #CognitiveAI #DeepLearning

✨Schoenfeld's Anatomy of Mathematical Reasoning by Language Models

📝 Summary:
This paper introduces ThinkARM, a framework based on Schoenfelds Episode Theory, to abstract LLM reasoning traces into functional steps. It reveals distinct thinking dynamics and structural differences in models solving math problems, with exploration being key for correctness. This makes LLM rea...

🔹 Publication Date: Published on Dec 23

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.19995
• PDF: https://arxiv.org/pdf/2512.19995
• Github: https://github.com/MingLiiii/ThinkARM

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LLM #AIReasoning #MathematicalReasoning #AI #MachineLearning

✨Fantastic Reasoning Behaviors and Where to Find Them: Unsupervised Discovery of the Reasoning Process

📝 Summary:
This paper introduces RISE, an unsupervised framework using sparse auto-encoders to discover and control LLM reasoning behaviors. It identifies interpretable reasoning vectors like reflection and backtracking, enabling targeted interventions and discovery of novel behaviors without retraining.

🔹 Publication Date: Published on Dec 30, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.23988
• PDF: https://arxiv.org/pdf/2512.23988

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LLM #AI #MachineLearning #AIReasoning #Interpretability

✨GameTalk: Training LLMs for Strategic Conversation

📝 Summary:
The GameTalk framework trains large language models for strategic multi-turn dialogue, optimizing global objectives using whole-conversation reward signals. This approach significantly outperforms untrained models, showing conversational fine-tuning is a promising path for LLM reasoning and negot...

🔹 Publication Date: Published on Jan 22

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.16276
• PDF: https://arxiv.org/pdf/2601.16276

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LLMs #ConversationalAI #StrategicDialogue #AITraining #AIReasoning

✨MortalMATH: Evaluating the Conflict Between Reasoning Objectives and Emergency Contexts

📝 Summary:
Specialized AI reasoning models prioritize task completion over safety. Our MortalMATH benchmark shows these models ignore emergencies to complete math, unlike generalist models. This relentless focus on correctness may remove crucial safety instincts and cause dangerous delays.

🔹 Publication Date: Published on Jan 26

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.18790
• PDF: https://arxiv.org/pdf/2601.18790

✨ Datasets citing this paper:
• https://huggingface.co/datasets/sileod/MortalMATH

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#AISafety #AIethics #MachineLearning #AIReasoning #MortalMATH

✨VERGE: Formal Refinement and Guidance Engine for Verifiable LLM Reasoning

📝 Summary:
VERGE is a neurosymbolic framework that combines LLMs with SMT solvers for verification-guided iterative refinement of reasoning. It enhances logical correctness through formal semantic checking, semantic routing, and precise error localization, achieving an 18.7% performance uplift on reasoning ...

🔹 Publication Date: Published on Jan 27

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.20055
• PDF: https://arxiv.org/pdf/2601.20055

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LLM #NeurosymbolicAI #FormalVerification #AIReasoning #SMTSolvers

✨InftyThink+: Effective and Efficient Infinite-Horizon Reasoning via Reinforcement Learning

📝 Summary:
To address high costs and limits in chain-of-thought reasoning, InftyThink uses reinforcement learning to optimize iterative reasoning. It learns to strategically summarize and resume, boosting accuracy by 21% on AIME24, reducing latency, and improving generalization.

🔹 Publication Date: Published on Feb 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.06960
• PDF: https://arxiv.org/pdf/2602.06960

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#ReinforcementLearning #AIReasoning #ChainOfThought #ArtificialIntelligence #MachineLearning

✨V_1: Unifying Generation and Self-Verification for Parallel Reasoners

📝 Summary:
V1 unifies generation and verification for complex reasoning tasks. It leverages models' superior ability in pairwise self-verification over independent scoring, improving performance and efficiency in code generation and math.

🔹 Publication Date: Published on Mar 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.04304
• PDF: https://arxiv.org/pdf/2603.04304
• Project Page: https://harmandotpy.github.io/v1-verification/
• Github: https://github.com/HarmanDotpy/pairwise-self-verification

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#AI #LLMs #MachineLearning #CodeGeneration #AIReasoning

✨Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models

📝 Summary:
Vision-R1 is a reasoning MLLM enhancing multimodal reasoning via Reinforcement Learning. It leverages a large, AI-generated multimodal CoT dataset and new training strategies to refine reasoning. This achieves high accuracy on multimodal math benchmarks.

🔹 Publication Date: Published on Mar 9, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2503.06749
• PDF: https://arxiv.org/pdf/2503.06749
• Github: https://github.com/Osilly/Vision-R1

✨ Datasets citing this paper:
• https://huggingface.co/datasets/Yuting6/ttrl
• https://huggingface.co/datasets/LoadingBFX/GeoQA-PLUS-aug-train-Vision-R1-cot-rewrite
• https://huggingface.co/datasets/LoadingBFX/GeoQA-train-Vision-R1-cot-rewrite

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#MLLM #ReinforcementLearning #AIReasoning #ChainOfThought #ArtificialIntelligence

✨Reasoning Shift: How Context Silently Shortens LLM Reasoning

📝 Summary:
LLMs significantly shorten their reasoning traces when problems are presented in various contexts compared to isolation. This compression reduces self-verification, potentially affecting performance on complex tasks. It highlights issues with LLM reasoning robustness and context management.

🔹 Publication Date: Published on Apr 1

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.01161
• PDF: https://arxiv.org/pdf/2604.01161

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#LLM #AIReasoning #ContextualAI #AIRobustness #MachineLearning