Papers with AIME24
Light-R1: Curriculum SFT, DPO and RL for Long COT from Scratch and Beyond (2025.acl-industry)
Copied to clipboard
Liang Wen, Yunke Cai, Fenrui Xiao, Xin He, Qi An, Zhenyu Duan, Yimin Du, Junchen Liu, Tanglifu Tanglifu, Xiaowei Lv, Haosheng Zou, Yongchao Deng, Shousheng Jia, Xiangzheng Zhang
| Challenge: | Experimental results show that opensource curriculum training is more effective when distinct datasets are available for different training stages. |
| Approach: | They propose an opensource suite for training long reasoning models using publicdata and models. |
| Outcome: | The proposed model outperforms DeepSeek-R1-DistillQwen-32B models in math reasoning. |
Beyond Token Length: Step Pruner for Efficient and Accurate Reasoning in Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing reinforcement learning methods for large reasoning models suffer from excessive verbosity, known as "overthinking." Existing models penalize generated tokens to promote conciseness, but these methods encounter two challenges: they may develop hacking behavior in later stages of training by discarding reasoning steps. |
| Approach: | They propose a framework that steers large reasoning models toward more efficient reasoning . they prioritize correctness while imposing penalties for redundant steps . |
| Outcome: | The proposed framework reduces token usage by 69.7% on AIME24. |
SLIM: Subtrajectory-Level Elimination for More Effective Reasoning (2025.findings-emnlp)
Copied to clipboard
Xifeng Yao, Chengyuan Ma, Dongyu Lang, Yinhao Ni, Zhiwei Xu, Huarui Xie, Zihao Chen, Guang Shen, Dandan Tu, Yi Bai, Changzheng Zhang
| Challenge: | Notable examples include OpenAI’s o1/o3/o4 series and DeepSeek-R1 . |
| Approach: | They develop a framework to identify suboptimal subtrajectories based on human-established criteria . they also use a sampling algorithm to select data whose reasoning process is free from suboptimally subtravertories to the highest degree . |
| Outcome: | The proposed method reduces the number of suboptimal subtrajectories by 25.9% during the inference process. |
Think Hard Only When Needed: A Hybrid Best-of-N and Beam Search for Efficient Test-Time Compute (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) exhibit remarkable reasoning and planning capabilities, yet their substantial inference-time cost significantly impedes deployment in resourceconstrained applications. |
| Approach: | They propose a hybrid inference pipeline that combines beam search and Best-of-N . THROW generates shorter initial trajectories and evaluates them using PRMs . |
| Outcome: | THROW achieves 1.54 and 14.38 latency speedups and 35.7% and 80.4% token reductions on average compared to Best-of-N and beam search . |
TrigReason: Trigger-Based Collaboration between Small and Large Reasoning Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Reasoning Models suffer from high inference latency due to autoregressive reasoning . SpecReason adopts a polling-based design that repeatedly invokes the LRM for verification at every step . |
| Approach: | They propose a trigger-based collaborative reasoning framework that delegates most reasoning to the SRM and activates LRM intervention only when necessary. |
| Outcome: | The proposed framework reduces latency and API cost by 73.3% under edge–cloud conditions. |
LLaMA-Berry: Pairwise Optimization for Olympiad-level Mathematical Reasoning via O1-like Monte Carlo Tree Search (2025.naacl-long)
Copied to clipboard
Di Zhang, Jianbo Wu, Jingdi Lei, Tong Che, Jiatong Li, Tong Xie, Xiaoshui Huang, Shufei Zhang, Marco Pavone, Yuqiang Li, Wanli Ouyang, Dongzhan Zhou
| Challenge: | LLaMA-Berry is an advanced mathematical reasoning framework to enhance the problem-solving ability of large language models (LLMs). |
| Approach: | They propose a Monte Carlo Tree Search and Self-Refine framework to optimize reasoning paths and a pairwise reward model to evaluate different paths globally. |
| Outcome: | The proposed framework overcomes inefficiencies and limitations of step-wise and greedy search algorithms, enabling more efficient exploration of solution spaces. |
START: Self-taught Reasoner with Tools (2025.emnlp-main)
Copied to clipboard
Chengpeng Li, Mingfeng Xue, Zhenru Zhang, Jiaxi Yang, Beichen Zhang, Bowen Yu, Binyuan Hui, Junyang Lin, Xiang Wang, Dayiheng Liu
| Challenge: | Large Reasoning Models (LRMs) have demonstrated remarkable capabilities in complex reasoning through long chain-of-thought, yet they struggle with precise computations and algorithmic operations. |
| Approach: | They propose a training-free approach that activates LRMs’ latent tool-use capabilities through artificial hints and a framework that enables models to learn effective tool utilization through diverse hint patterns and rejection-based data synthesis. |
| Outcome: | Experiments show that START significantly improves state-of-the-art LRMs across challenging benchmarks, including competition-level mathematics (AMC23: 95.0%, AIME24: 75.6%) and graduate-level science questions (GPQA: 64.6%). |
PRIME: A Process-Outcome Alignment Benchmark for Verifiable Reasoning in Mathematics and Engineering (2026.acl-long)
Copied to clipboard
Xiangfeng Wang, Hangyu Guo, Yanlin Lai, Mitt Huang, Liang Zhao, Chengyuan Yao, Yinmin Zhang, Qi Han, null Xiaoxiaoren, Chun Yuan, Tong Xu, Zheng Ge, Xiangyu Zhang, Daxin Jiang
| Challenge: | Current outcome-centric verification paradigms neglect potential errors in the derivation process. |
| Approach: | They propose a process-aware RLVR training paradigm utilizing verifiers selected via **PRIME**. |
| Outcome: | The proposed approach outperforms the baseline verification paradigm on AIME24, AIME25, and Beyond-AIME models. |
Distributional Clarity: The Hidden Driver of RL-Friendliness in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | RL-friendly models exhibit intra-class compactness and inter-class separation in probability assignments . under identical training, Qwen models achieve substantial gains, while others like Llama yield limited improvements. |
| Approach: | They propose a method to quantify distributional clarity in probability space . they show distributional clearness is a trainable property underlying RL-Friendliness . |
| Outcome: | The proposed model families achieve substantial gains under identical training, while others like Llama yield limited improvements. |
s1: Simple test-time scaling (2025.emnlp-main)
Copied to clipboard
Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candes, Tatsunori Hashimoto
| Challenge: | OpenAI’s o1 model showed this capability but did not publicly share its methodology, leading to many replication efforts. |
| Approach: | They curate a small dataset s1K with 1,000 reasoning questions based on three criteria we validate through ablations: difficulty, diversity, and quality. |
| Outcome: | The proposed model exceeds o1-preview on competition math questions by up to 27% (MATH and AIME24). |
MUR: Momentum Uncertainty guided Reasoning for Large Language Models (2026.acl-long)
Copied to clipboard
Hang Yan, Fangzhi Xu, Rongman Xu, Yifei Li, Jian Zhang, Haoran Luo, Xiaobao Wu, Anh Tuan Luu, Haiteng Zhao, Qika Lin, Jun Liu
| Challenge: | Existing methods for optimizing reasoning quality are limited by overthinking. |
| Approach: | They propose a method that allocates thinking budgets to critical reasoning steps by tracking and aggregating step-wise uncertainty over time. |
| Outcome: | The proposed method reduces computation by over 45% on average while improving accuracy by 0.33–3.46%. |
Dynamic Sampling that Adapts: Self-Aware Iterative Data Persistent Optimization for Mathematical Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Current data selection paradigms rely on static, externally defined metrics, which fail to adapt to the evolving capabilities of models during training. |
| Approach: | They propose a dynamic sampling framework that aligns training data with the model's intrinsic competence by iterating on real-time feedback. |
| Outcome: | Extensive experiments on eight benchmarks show that SAI-DPO outperforms static baselines at most nearly 6 points, achieving state-of-the-art efficiency with significantly less data. |
Long Chain-of-Thought Fine-tuning via Understanding-to-Reasoning Transition (2025.emnlp-main)
Copied to clipboard
Chenxin An, Zhihui Xie, Xiaonan Li, Ming Zhong, Shansan Gong, Lei Li, Jun Zhang, Jingjing Xu, Lingpeng Kong
| Challenge: | Existing research on long-context scaling in language models has focused on managing lengthy input prompts instead of producing long outputs. |
| Approach: | They propose a sequence-level curriculum learning framework that shifts a model’s focus from interpreting long chain-of-thoughts to generating them. |
| Outcome: | Experiments on rigorous reasoning benchmarks, including AIME24 and GPQA Diamond, show that the proposed approach surpasses standard fine-tuning by over 10% while maintaining robust performance on understanding tasks. |
Thinking Traps in Long Chain-of-Thought: A Measurable Study and Trap-Aware Adaptive Restart (2026.findings-acl)
Copied to clipboard
null Chenkang, Fan Yu, Junjie Nian, Sihan Zhao, Zhuoka Feng, Zijun Yao, Wang Heng, Yu Minshen, Yixin Cao
| Challenge: | Experiments show that extended generation does not guarantee correctness . a recurring pattern in Long-CoT failures is a problem for large reasoning models . |
| Approach: | They propose a test-time control framework that truncates the trajectory before the trap segment and adaptively restarts decoding. |
| Outcome: | Experiments show that TAAR improves reasoning performance without fine-tuning model parameters. |