What Happened in LLMs Layers when Trained for Fast vs. Slow Thinking: A Gradient Perspective (2025.acl-long)
Copied to clipboard
| Challenge: | Xu et al., 2024) study shows that slow thinking can distinguish correct and irrelevant reasoning paths. |
| Approach: | They investigate how fast vs. slow thinking affects layer-wise gradients in large language models . they find that slow thinking can distinguish correct and irrelevant reasoning paths . |
| Outcome: | The results show that slow thinking can distinguish correct and irrelevant reasoning paths. |
Similar Papers
Imbalanced Gradients in RL Post-Training of Multi-Task LLMs (2026.findings-eacl)
Copied to clipboard
Runzhe Wu, Ankur Samanta, Ayush Jain, Scott Fujimoto, Jeongyeol Kwon, Ben Kretzu, Youliang Yu, Kaveh Hassani, Boris Vidolov, Yonathan Efroni
| Challenge: | Large-gradient tasks can achieve similar or even much lower learning gains than small-grading ones. |
| Approach: | They show that large-gradient tasks can achieve lower learning gains than small-grading ones . large-grade tasks can accomplish similar or even lower learning gain than small grade ones if they are large . |
| Outcome: | The proposed approach fails when certain tasks produce larger gradients . Large-gradient tasks can achieve lower learning gains than small-gradent ones . |
How Instruction and Reasoning Data shape Post-Training: Data Quality through the Lens of Layer-wise Gradients (2026.acl-long)
Copied to clipboard
| Challenge: | Spectral properties of low/high-quality instruction and reasoning data are used to explain finetuning dynamics in large language models. |
| Approach: | They propose to analyze layer-wise gradients induced by low/high-quality instruction and reasoning data for LLM post-training. |
| Outcome: | The results show that higher-quality data are associated with lower nuclear norms and higher effective ranks. |
Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens (2026.findings-acl)
Copied to clipboard
Chengshuai Zhao, Zhen Tan, Pingchuan Ma, Dawei Li, Bohan Jiang, Yancheng Wang, Yingzhen Yang, Huan Liu
| Challenge: | Chain-of-Thought (CoT) prompting has been shown to be effective in eliciting structured reasoning from large language models (LLMs). |
| Approach: | They propose a data distribution lens to understand when and why CoT reasoning fails . they propose 'data-based' training that trains LLMs from scratch . |
| Outcome: | The proposed model enables models to generate reasoning trajectories that approximate those observed during training. |
How Likely Do LLMs with CoT Mimic Human Reasoning? (2025.coling-main)
Copied to clipboard
| Challenge: | Using chain-of-thought to elicit reasoning capabilities is not always effective and accurate. |
| Approach: | They compare the reasoning process of LLMs with humans to understand the causal chain . they find that LLM deviates from the ideal causal chain, resulting in spurious correlations . |
| Outcome: | The proposed method does not improve performance or accurately represent reasoning processes in LLMs. |
Mapping the Minds of LLMs: A Graph-Based Analysis of Reasoning LLMs (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Reasoning Models (LRMs) often display unstable behaviors, e.g., hallucinating unsupported premises, overthinking simple tasks, and displaying higher sensitivity to prompt variations. |
| Approach: | They propose a graph-based analytical framework that clusters long, verbose CoT outputs into semantically coherent reasoning steps, then constructs directed reasoning graphs to capture contextual and logical dependencies among these steps. |
| Outcome: | The proposed framework enables quantitative evaluation of internal reasoning structure and quality beyond conventional metrics and provides practical insights for prompt engineering and cognitive analysis of LLMs. |
Do LLMs Really Need 10+ Thoughts for “Find the Time 1000 Days Later”? Towards Structural Understanding of LLM Overthinking (2026.acl-long)
Copied to clipboard
Xinliang Frederick Zhang, Anhad Mohananey, Alexandra Chronopoulou, Pinelopi Papalampidi, Somit Gupta, Tsendsuren Munkhdalai, Lu Wang, Shyam Upadhyay
| Challenge: | Existing studies on LLMs' thought processes are limited to superficial, profiling-based observations, failing to delve into their inner workings. |
| Approach: | They propose a utility-based definition of overthinking that moves beyond length-based metrics and provides a more insightful understanding of LLMs' thought progression. |
| Outcome: | The proposed model decomposes the LLM thought process into minimally complete sub-thoughts and identifies common thinking patterns for topically similar queries. |
Can LLMs Learn From Mistakes? An Empirical Study on Reasoning Tasks (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing work has shown that simple learning can enhance the chain-of-thought (CoT) reasoning of large language models. |
| Approach: | They construct mistake-correction datasets to identify and correct mistakes in CoTs . they conclude that LLMs can learn from mistakes to enhance their CoT reasoning . |
| Outcome: | The proposed datasets show that LLMs can learn from mistakes to enhance their CoT reasoning performance. |
Through the Valley: Path to Effective Long CoT Training for Small Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Long chain-of-thought (CoT) supervision is effective for large language models . but small models trained on limited long CoT data experience performance degradation . |
| Approach: | They identify a phenomenon called Long CoT Degradation in small language models . long CoT data can be used to generate long chain-of-thought (CoT) responses . |
| Outcome: | The results show that models trained on 8k long CoT examples lose up to 75% of their original performance before fine-tuning. |
Think Faster Than Words: Efficient LLM Chain-of-Thought Reasoning via Dynamic Shortcut Decoding (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods that prune or employ early stopping to reduce latency often compromise reasoning reliability. |
| Approach: | They propose a shortcut decoding framework that integrates probes over internal hidden states with step-level entropy to detect convergence of reasoning during generation and adaptively selects between a fast-exit path and a stability-verified path to remove redundant steps while preserving answer correctness. |
| Outcome: | The proposed framework reduces token usage by approximately 35% and maintains accuracy comparable to full CoT decoding. |
ReEfBench: Quantifying the Reasoning Efficiency of LLMs (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for Chain-of-Thought evaluations do not distinguish between genuine reasoning and mere verbosity. |
| Approach: | They propose a framework for the non-intrusive, comprehensive process-centric evaluation of reasoning grounded in First-Order Logic. |
| Outcome: | The proposed framework identifies four distinct behavioral prototypes and diagnoses the failure modes. |