Learning While Staying Curious: Entropy-Preserving Supervised Fine-Tuning via Adaptive Self-Distillation for Large Reasoning Models (2026.acl-long)
Copied to clipboard
Hao Wang, Hao Gu, Hongming Piao, Kaixiong Gong, Yuxiao Ye, Xiangyu Yue, Sirui Han, Yike Guo, Dapeng Wu
| Challenge: | Recent advances establish "SFT-then-RL" as the defacto paradigm for enhancing large reasoning mod- els on automatically verifiable tasks. |
| Approach: | They propose an entropy-preserving SFT method to enhance exploration capabilities through intrinsic curiosity. |
| Outcome: | The proposed method outperforms the vanilla method on reasoning tasks by 2.5 points . it also outperformed the vanilla SFT by 2.9 points on out-of-distribution tasks . |
Similar Papers
SED-SFT: Selectively Encouraging Diversity in Supervised Fine-Tuning (2026.acl-short)
Copied to clipboard
| Challenge: | Existing studies have proposed a new approach to optimize for SFT followed by RL . existing studies have suggested a method to optimize SFT for large language models . |
| Approach: | They propose a framework that encourages diversity based on token exploration space. |
| Outcome: | Experiments show that SED-SFT significantly improves generation diversity with a negligible computational overhead increase over CE loss. |
Complexity-aware fine-tuning (2026.findings-eacl)
Copied to clipboard
| Challenge: | General-purpose Large Language Models (LLMs) are often fine-tuned through supervised fine- tuning (SFT) to enhance performance in specific domains. |
| Approach: | They propose a novel approach that uses reasoning only for complex data identified by entropy to refine large language models. |
| Outcome: | The proposed model outperforms the standard SFT approach while using 81% less data. |
Skill-Aware Data Selection and Fine-Tuning for Data-Efficient Reasoning Distillation (2026.acl-short)
Copied to clipboard
| Challenge: | Large reasoning models such as DeepSeek-R1 and their distilled variants achieve impressive performance on complex reasoning tasks, yet their costs remain substantial. |
| Approach: | They propose a skill-centric distillation framework that efficiently transfers reasoning ability to weaker models with two components: (1) Skill-based data selection, which prioritizes examples targeting the student model’s weaker skills, and (2) Skillaware fine-tuning, which encourages explicit skill decomposition during problem solving. |
| Outcome: | The proposed framework surpasses baselines on Qwen3-4B and Qwend3-8B and focuses on skills emphasized during training. |
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). |
ReFT: Reasoning with Reinforced Fine-Tuning (2024.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to improve the generalization of large language models are using Supervised Fine-Tuning (SFT) this approach does not show sufficient generalization ability because it only relies on the given CoT data. |
| Approach: | They propose to use Chain-of-Thought annotations to train Large Language Models using supervised fine-tuning to improve generalization. |
| Outcome: | The proposed approach outperforms SFT on GSM8K, MathQA, and SVAMP datasets and shows a superior generalization ability. |
Beyond High-Entropy Exploration: Correctness-Aware Low-Entropy Segment-Based Advantage Shaping for Reasoning LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent work studies RLVR through token entropy, arguing that high-entropies drive exploration and should receive stronger updates. |
| Approach: | They propose a correctness-aware reinforcement framework that performs fine-grained advantage modulation over low-entropy segments. |
| Outcome: | The proposed framework improves accuracy over strong RL baselines across three backbones and six math benchmarks while maintaining high-entropy exploration. |
Bridging SFT and RL: Dynamic Policy Optimization for Robust Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing unified optimization strategies overlook the statistical conflict between these distinct gradient signals. |
| Approach: | They propose a framework to reduce bias-variance trade-offs in Large Language Models . they propose DYPO, which leverages intrinsic group dynamics to significantly reduce RL gradient variance . |
| Outcome: | The proposed framework outperforms traditional pipelines on reasoning benchmarks and out-of-distribution tasks. |
GSM-Noise: Exploring and Enhancing Large Language Models’ Reasoning under Noisy Inputs (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models struggle when dealing with complex, ill-formed, or noisy inputs . open-source models are less robust, while closed-source ones are more robust . |
| Approach: | They propose to use GSM-Noise to refine inputs before engaging in in-depth analysis to improve LLM robustness under noisy conditions. |
| Outcome: | The proposed model can achieve consistent performance gains under noisy conditions with prompt engineering, supervised finetuning, and reinforcement learning. |
Enhancing Efficiency and Exploration in Reinforcement Learning for LLMs (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches allocate an equal number of rollouts to all questions during the RL process, which is inefficient. |
| Approach: | They propose a mechanism for dynamically allocating rollout budgets based on the difficulty of the problems, enabling more efficient RL training. |
| Outcome: | The proposed model improves response precision while preserving exploratory ability to uncover potential correct pathways. |
Warm Up Before You Train: Unlocking General Reasoning in Resource-Constrained Settings (2025.emnlp-main)
Copied to clipboard
| Challenge: | Reasoning-capable large language models (LLMs) have driven a major shift in artificial intelligence . these models generate long CoTs, capturing reasoning behaviors such as self-reflection, self-correction, and hypothesis testing. |
| Approach: | They propose a sample-efficient, two-stage training strategy to build reasoning LLMs . they "warm up" a model by distilling Long CoTs from a toy domain to acquire general reasoning skills . |
| Outcome: | The proposed training strategy outperforms existing models on a range of tasks. |