Li Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Li Li has a total of 155 co-authors. Here are the top co-authors by number of shared publications:
Xiaofei Xu
Murdoch University
19 shared publicationsFei Hu
Southwest University
11 shared publicationsJingyuan Wang
Southwest University
11 shared publications- Jun Shen
University of Wollongong
10 shared publications Jinjing Zhang
School of Computer and Information Science, Southwest University, Chongqing, China. s
7 shared publicationsSong Yu
6 shared publications
Latest Publications
- 26 September 2026
- Published by Springer Nature
- p. 167-182
A Chain-of-Thought Guided Framework for Identity Preservation Video Generation in Education
- 26 September 2026
- Published by Springer Nature
- p. 167-182
A Chain-of-Thought Guided Framework for Identity Preservation Video Generation in Education
- Open Access
- 17 September 2026
- Published by Elsevier in SSRN Electronic Journal
EP-GRPO: Entropy-Progress Aligned Group Relative Policy Optimization with Outcome-Anchored Policy-Deviation Guidance
Reinforcement learning with verifiable rewards (RLVR), particularly Group Relative Policy Optimization (GRPO), has advanced large language model reasoning. However, sequence-level GRPO does not explicitly distinguish token informativeness or token-level polarity, and its outcome advantage vanishes when all rewards in a group are identical. We quantify these behaviors through token perturbation, step-level annotation, and training-log analyses.To address them, we propose Entropy-Progress Aligned GRPO (EP-GRPO). EP-GRPO combines an entropy gate for token-wise outcome weighting, an outcome-anchored policy-deviation signal for dense shaping without an external reward model, and cumulative-entropy buckets for normalization at comparable information-progress stages. All non-smooth anchors and rollout statistics are detached from gradient computation, so optimization retains the standard clipped policy-gradient form.Across five mathematical reasoning benchmarks, EP-GRPO raises average accuracy from 18.14\% to 22.93\% for Qwen2.5-3B and from 27.11\% to 30.34\% for Qwen2.5-7B relative to matched GRPO runs. The code will be made available.
- 17 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessEP-GRPO: Entropy-Progress Aligned Group Relative Policy Optimization with Outcome-Anchored Policy-Deviation Guidance
Reinforcement learning with verifiable rewards (RLVR), particularly Group Relative Policy Optimization (GRPO), has advanced large language model reasoning. However, sequence-level GRPO does not explicitly distinguish token informativeness or token-level polarity, and its outcome advantage vanishes when all rewards in a group are identical. We quantify these behaviors through token perturbation, step-level annotation, and training-log analyses.To address them, we propose Entropy-Progress Aligned GRPO (EP-GRPO). EP-GRPO combines an entropy gate for token-wise outcome weighting, an outcome-anchored policy-deviation signal for dense shaping without an external reward model, and cumulative-entropy buckets for normalization at comparable information-progress stages. All non-smooth anchors and rollout statistics are detached from gradient computation, so optimization retains the standard clipped policy-gradient form.Across five mathematical reasoning benchmarks, EP-GRPO raises average accuracy from 18.14\% to 22.93\% for Qwen2.5-3B and from 27.11\% to 30.34\% for Qwen2.5-7B relative to matched GRPO runs. The code will be made available.
- 21 August 2026
- Published by Elsevier in Information Sciences
- Vol. 759
Multi-level knowledge fusion and uncertainty-aware distillation for zero-shot cross-lingual named entity recognition
- 21 August 2026
- Published by Elsevier in Information Sciences
- Vol. 759
Multi-level knowledge fusion and uncertainty-aware distillation for zero-shot cross-lingual named entity recognition
- 21 July 2026
- Published by Elsevier in Information Processing & Management
- Vol. 64 (1)
CoTMind: A closed-loop framework for optimizing Chain-of-Thought reasoning in complex question answering
- 21 July 2026
- Published by Elsevier in Information Processing & Management
- Vol. 64 (1)
CoTMind: A closed-loop framework for optimizing Chain-of-Thought reasoning in complex question answering
- 2 July 2026
- Published by Springer Nature
- p. 432-444
Defense Against Prompt Injection Attacks via Intention Emphasis
- 2 July 2026
- Published by Springer Nature
- p. 432-444
Defense Against Prompt Injection Attacks via Intention Emphasis
- 7 April 2026
- Published by Elsevier in Neural Networks
- Vol. 201, 108954
Retrieval-augmented and feedback-optimized large language model for recommendation
- 7 April 2026
- Published by Elsevier in Neural Networks
- Vol. 201, 108954
Retrieval-augmented and feedback-optimized large language model for recommendation
- 15 March 2026
- Published by Springer Nature
- p. 476-491
TDCRec: Time-Varying Demand Causal Modeling for Recommendation Debiasing
- 15 March 2026
- Published by Springer Nature
- p. 476-491
TDCRec: Time-Varying Demand Causal Modeling for Recommendation Debiasing
- Open Access
- 21 February 2026
- Published by Association for Computing Machinery (ACM)
- p. 1007-1016
XLingLearn: Balancing Alignment and Robustness for Zero-Shot Cross-Lingual Transfer
- 21 February 2026
- Published by Association for Computing Machinery (ACM)
- p. 1007-1016
Open AccessXLingLearn: Balancing Alignment and Robustness for Zero-Shot Cross-Lingual Transfer
- 21 February 2026
- Published by Springer Nature in Neural Computing and Applications
- Vol. 38 (4) , 73
Enhancing Chinese legal judgment prediction in LLMs via legal norms integration
- 21 February 2026
- Published by Springer Nature in Neural Computing and Applications
- Vol. 38 (4) , 73
Enhancing Chinese legal judgment prediction in LLMs via legal norms integration