Rethinking Expert Trajectory Utilization in LLM Post-training for Mathematical Reasoning (2026.acl-long)
Copied to clipboard
Bowen Ding, Yuhan Chen, Jiayang Lyu, Jiyao Yuan, Qi Zhu, Shuangshuang Tian, Dantong Zhu, Futing Wang, Heyuan Deng, Fei Mi, Lifeng Shang, Tao Lin
| Challenge: | evolving generic Large Language Models into specialized Large Reasoning Models requires effective post-training. |
| Approach: | They propose a plasticity-ceiling framework to harness expert trajectories . they establish the Sequential SFT-then-RL pipeline as the superior standard . |
| Outcome: | The proposed framework overcomes stability and premature convergence deficits in synchronized approaches. |
Similar Papers
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. |
Why Does Reinforcement Learning Generalize? A Feature-Level Mechanistic Study of Post-Training in Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Reinforcement learning (RL)-based post-training often improves the reasoning performance of large language models beyond the training domain, while supervised fine-tuning (SFT) frequently leads to general capabilities forgetting. |
| Approach: | They propose a feature-level mechanistic analysis methodology to probe RL generalization using a controlled experimental setup. |
| Outcome: | The proposed method identifies a compact, task-agnostic set of features that directly mediate generalization across diverse tasks. |
A Survey of Post-Training Scaling in Large Language Models (2025.acl-long)
Copied to clipboard
Hanyu Lai, Xiao Liu, Junjie Gao, Jiale Cheng, Zehan Qi, Yifan Xu, Shuntian Yao, Dan Zhang, Jinhua Du, Zhenyu Hou, Xin Lv, Minlie Huang, Yuxiao Dong, Jie Tang
| Challenge: | Large language models (LLMs) have demonstrated proficiency in understanding and generating human natural languages. |
| Approach: | They propose a framework for scaling large language models using supervised fine-tuning, RLxF and test-time compute methodologies. |
| Outcome: | The proposed model can be used to understand and generate human natural languages. |
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. |
DecIF: Improving Instruction-Following through Decomposition (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to obtain high-quality instruction-following data rely heavily on existing documents and existing methods. |
| Approach: | They propose a data synthesis framework, DecIF, which automatically generates accurate and diverse instruction-following data from scratch for supervised fine-tuning and reinforcement learning. |
| Outcome: | Extensive experiments show that the proposed framework can synthesize accurate instruction-following data for both SFT and RL paradigms compared to baselines. |
Knowledge is Not Enough: Injecting RL Skills for Continual Adaptation (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models face the "knowledge cutoff" problem because their parametric memory remains frozen after pretraining, preventing them from natively internalizing new information or tools on the fly. |
| Approach: | They propose a framework that supports modular skill transfer for efficient and effective knowledge adaptation by extracting a domain-agnostic **Skill Vector from a source domain. |
| Outcome: | Experiments on knowledge-incorporation QA (SQuAD, LooGLE) and agentic tool-use benchmarks show that the proposed framework outperforms state-of-the-art self-editing SFT by 9.9 points. |
Beyond Reasoning Gains: Mitigating General-Capability Forgetting in Large Reasoning Models (2026.findings-acl)
Copied to clipboard
Hoang Phan, Xianjun Yang, Yuanshun Yao, Jingyu Zhang, Shengjie Bi, Xiaocheng Tang, Madian Khabsa, Lijuan Liu, Deren Lei
| Challenge: | Reinforcement learning with verifiable rewards (RLVR) has delivered impressive gains in mathematical and multimodal reasoning . however, the recipe introduces a significant risk of capability regression, where models forget foundational skills after prolonged training without employing regularization strategies. |
| Approach: | They propose a replay strategy with dynamic objective reweighting for general knowledge preservation using short-horizon signals of convergence and instability. |
| Outcome: | The proposed method preserves general capabilities and improves reasoning . it can be applied to existing RLVR pipelines without training additional models or tuning . |
Balancing Classification and Calibration Performance in Decision-Making LLMs via Calibration Aware Reinforcement Learning (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are increasingly deployed in decision-making tasks where accuracy and reliable confidence estimates are essential. |
| Approach: | They propose a calibration-aware reinforcement learning formulation that directly adjusts decision-token probabilities. |
| Outcome: | The proposed model preserves RLVR’s accuracy level while mitigating overconfidence, reducing ECE scores up to 9 points. |
A Layer-wise Analysis of Supervised Fine-Tuning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for fine-tuning ignore depth-dependent heterogeneity of instruction-following . a critical gap remains in understanding where these changes occur across the model's depth and which layers are essential for instruction- following. |
| Approach: | They propose a method which selectively updates critical intermediate layers . they show that effective alignment is architecturally localized rather than distributed . |
| Outcome: | The proposed method outperforms standard LoRA up to 10.2% on GSM8K with reduced parameter overhead. |
Don’t Tell the Answer, Truly Guide the Reasoning During RL Rollouts (2026.findings-acl)
Copied to clipboard
Xinyi Wang, Jinyi Han, Zishang Jiang, Tingyun li, Jiaqing Liang, Sihang Jiang, Zhaoqian Dai, Ma Shuguang, Fei Yu, Yanghua Xiao
| Challenge: | Existing methods such as GRPO often break down when task difficulty exceeds the model’s capacity, resulting in sparse rewards and inefficient training. |
| Approach: | They propose to measure the compatibility between external guidance and a model's intrinsic policy by introducing an adaptive framework to enhance reasoning performance while explicitly preserving high Affinity. |
| Outcome: | The proposed framework outperforms baseline models while maintaining high Affinity. |