From Reasoning to Answer: Empirical, Attention-Based and Mechanistic Insights into Distilled DeepSeek R1 Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Reasoning Models generate explicit reasoning traces alongside final answers . the extent to which these traces influence answer generation remains unclear . |
| Approach: | They conduct empirical evaluation of Large Reasoning Models that include explicit reasoning . they also show that answer tokens attend substantially to reasoning tokens . |
| Outcome: | The results show that including explicit reasoning improves answer quality across domains . they also show that answer tokens attend substantially to reasoning tokens - the authors . |
Similar Papers
How Should We Enhance the Safety of Large Reasoning Models: An Empirical Study (2026.acl-long)
Copied to clipboard
Zhexin Zhang, Xian Qi Loye, Victor Shea-Jay Huang, Junxiao Yang, Qi Zhu, Shiyao Cui, Fei Mi, Lifeng Shang, Yingkang Wang, Hongning Wang, Minlie Huang
| Challenge: | Large Reasoning Models have achieved remarkable success on reasoning-intensive tasks, but their enhanced reasoning capabilities do not translate to improved safety performance. |
| Approach: | They propose to use supervised fine tuning to enhance the safety of Large Reasoning Models. |
| Outcome: | The proposed method improves the safety of large reasoning models on reasoning-intensive tasks. |
Reasoning Traces Shape Outputs but Models Won’t Say So (2026.acl-long)
Copied to clipboard
| Challenge: | Large reasoning models (LRMs) generate explicit reasoning traces before producing answers, offering a window into their decisionmaking. |
| Approach: | They propose a method that injects synthetic reasoning snippets into a model’s reasoning trace and measures whether the model follows the injected reasoning and acknowledges doing so. |
| Outcome: | The proposed method reveals that models refuse to disclose their influence when asked to explain their changed answers. |
The Illusion of Insight in Reasoning Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing evidence suggests that reasoning models exhibit "Aha!" moments . however, it remains unclear whether such intrinsic shifts in reasoning strategy improve performance . |
| Approach: | They analyze reasoning traces, training checks, three reasoning domains to detect mid-reasoning shifts . they find that reasoning shifts are rare, do not become more frequent with training . |
| Outcome: | The proposed method detects mid-reasoning shifts and instrument training runs. |
Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failures (2026.acl-long)
Copied to clipboard
Yi Hu, Jiaqi Gu, Ruxin Wang, Zijun Yao, Hao Peng, Xiaobao Wu, Jianhui Chen, Muhan Zhang, Liangming Pan
| Challenge: | Recent research has shown that reinforcement learning can elicit intriguing emergent reasoning behaviors. |
| Approach: | They propose a comprehensive survey of the mechanistic understanding of large reasoning models . they organize findings into three core dimensions: 1) training dynamics, 2) reasoning mechanisms, and 3) unintended behaviors. |
| Outcome: | This paper synthesizes the mechanistic understanding of large reasoning models into three dimensions . authors outline a roadmap for future studies including improved interpretability and methodologies . |
Marco-o1 v2: Towards Widening The Distillation Bottleneck for Reasoning Models (2025.acl-long)
Copied to clipboard
Huifeng Yin, Yu Zhao, Minghao Wu, Xuanfan Ni, Bo Zeng, Huaiyu.wh Huaiyu.wh, Tianqi Shi, Liangying Shao, Chenyang Lyu, Longyue Wang, Weihua Luo, Kaifu Zhang
| Challenge: | Recent efforts to distill large reasoning models into smaller lightweight models have shown competitive performances. |
| Approach: | They propose to distill long Chain-of-Thought data to improve SFT and RL methods by constructing data from scratch using Monte Carlo Tree Search. |
| Outcome: | The proposed method significantly improves reasoning performance on various benchmarks such as math (GSM8K, MATH, AIME). |
Thinking Out Loud: Do Reasoning Models Know When They’re Right? (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large reasoning models (LRMs) have recently demonstrated impressive capabilities in complex reasoning tasks by leveraging increased test-time computation and exhibiting behaviors reminiscent of human-like self-reflection. |
| Approach: | They analyze verbalized confidence, how models articulate their certainty, as a lens into the nature of self-reflection in large reasoning models. |
| Outcome: | The proposed model exhibits human-like self-reflection in reasoning tasks, but how this ability interacts with other model behaviors remains underexplored. |
How Do Answer Tokens Read Reasoning Traces? Self-Reading Patterns in Thinking LLMs for Quantitative Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Prior work on activation steering has focused on shaping reasoning traces, but it remains unclear how answer tokens actually read and integrate the reasoning to produce reliable outcomes. |
| Approach: | They propose a training-free steering method that uses self-reading quality scores to guide inference toward benign self-readiness and away from uncertain and disorganized reading. |
| Outcome: | The proposed method yields consistent accuracy gains in the reasoning traces generated by thinking LLMs. |
Interpretable Traces, Unexpected Outcomes: Investigating the Disconnect in Trace-Based Knowledge Distillation (2026.acl-long)
Copied to clipboard
| Challenge: | Recent advances in reasoning-oriented Large Language Models have been driven by the introduction of Chain-of-Thought (CoT) traces. |
| Approach: | They propose to use CoT traces to guide model inference and serve as supervision signals for Knowledge Distillation to improve smaller models. |
| Outcome: | The proposed model is based on a rule-based problem decomposition method and is valid for both semantic correctness and interpretability to the end user. |
Beyond ’Aha!’: Toward Systematic Meta-Abilities Alignment in Large Reasoning Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Prior work has shown that outcome-based reinforcement learning (RL) can incidentally elicit advanced reasoning behaviors such as self-correction, backtracking, and verification. |
| Approach: | They explicitly align large reasoning models with three meta-abilities: deduction, induction, and abduction, using automatically generated, self-verifiable tasks. |
| Outcome: | The proposed model aligns models with deduction, induction, and abduction meta-abilities using automatically generated, self-verifiable tasks. |
Why Can Distillation Work with Limited Resources? A Systematic Study (2026.findings-acl)
Copied to clipboard
Xiao Hu, Xingyu Lu, Liyuan Mao, YiFan Zhang, Tianke Zhang, Bin Wen, Fan Yang, Tingting Gao, Guorui Zhou
| Challenge: | Recent advances in large language models have driven reasoning performance . low-resource distillation can boost models' performance, but a framework is missing . |
| Approach: | They conduct a controlled experiment to find out why low-resource distillation can boost model performance . they find that distillation enhances the presence of advanced cognitive behaviors . |
| Outcome: | The proposed model shows more flexible reasoning, the authors show . they show that distillation enhances the presence of advanced cognitive behaviors . |