Papers by Feifan Song
Two Pathways to Truthfulness: On the Intrinsic Encoding of LLM Hallucinations (2026.acl-long)
Copied to clipboard
Wen Luo, Guangyue Peng, Wei Li, Shaohang Wei, Feifan Song, Liang Wang, Nan Yang, Xingxing Zhang, Jing Jin, Furu Wei, Houfeng Wang
| Challenge: | Previous work shows that large language models generate hallucinations, yet the origins and mechanisms of these signals remain unclear. |
| Approach: | They propose to validate and disentangle two different pathways for truthfulness cues . they also propose to use the same mechanism to derive self-contained evidence from the generated answer . |
| Outcome: | The proposed applications improve hallucination detection performance by integrating two different inputs. |
ATLANTIS: Weak-to-Strong Learning via Importance Sampling (2025.acl-long)
Copied to clipboard
| Challenge: | ATLANTIS is a new technique to improve the performance of large language models. |
| Approach: | They propose a new technique to bridge the gap between the distribution of current datasets and the real-world data distribution by using importance sampling. |
| Outcome: | The proposed technique can bring consistent and significant improvements to models’ performance and can be flexibly transferred among models with different structures. |
Odysseus Navigates the Sirens’ Song: Dynamic Focus Decoding for Factual and Diverse Open-Ended Text Generation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing decoding methods struggle to balance factuality and diversity . Deterministic decoding approaches suffer from degeneration and lack of diversity - a problem that is not addressed by the current literature. |
| Approach: | They propose a plug-and-play stochastic approach that adjusts decoding focus based on distributional differences across layers, leveraging the modular nature of factual knowledge within LLMs. |
| Outcome: | Extensive experiments on seven datasets show that DFD significantly improves performance. |
Scaling Data Diversity for Fine-Tuning Language Models in Human Alignment (2024.lrec-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) can reveal toxic or offensive content inadvertently or intentionally. |
| Approach: | They propose to control the diversity of both sides according to the number of samples for fine-tuning, which can directly reflect their impact. |
| Outcome: | The proposed approach improves the performance of large language models after fine-tuning. |
MPO: Boosting LLM Agents with Meta Plan Optimization (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for interactive planning tasks suffer from planning hallucinations and require retraining for each new agent. |
| Approach: | They propose a framework that leverages explicit guidance through meta plans to assist agent planning and enables continuous optimization based on feedback from the agent’s task execution. |
| Outcome: | The proposed framework outperforms existing baselines on two representative tasks and significantly improves task completion efficiency and generalization capabilities. |
Well Begun is Half Done: Low-resource Preference Alignment by Weak-to-Strong Decoding (2025.findings-acl)
Copied to clipboard
| Challenge: | Low-resource methods for LLM alignment have been popular, but still face challenges in obtaining high-quality and aligned content. |
| Approach: | They propose a framework to enhance alignment ability of base models by the guidance of a small aligned model. |
| Outcome: | The proposed framework outperforms baseline methods while avoiding degradation on downstream tasks. |
Instantly Learning Preference Alignment via In-context DPO (2025.naacl-long)
Copied to clipboard
| Challenge: | Presently, mainstream approaches to HPA heavily depend on fine-tuning . however, the huge computational and annotation costs of fine-timing are hard to ignore . |
| Approach: | They propose a tuning-free approach to HPA using LLMs' decoding . they first rethink the derivation procedures of DPO and build an instant scorer . |
| Outcome: | The proposed approach outperforms existing methods even with tuning-free baselines and an upgraded scorer. |
Towards A Better Initial Policy Model For Scalable Long-CoT Reinforcement Learning (2025.findings-acl)
Copied to clipboard
| Challenge: | Long-CoT reasoning and reinforcement learning are demonstrating remarkable performance and scalability, however, there is a lack of systematic guidelines for obtaining a better initial policy model. |
| Approach: | They propose a systematic guideline and a novel Re-RFT method to obtain more efficient reasoning patterns from different initial models. |
| Outcome: | The proposed method surpasses DeepSeek-R1-Distill-Qwen-14B model by 4.6%, demonstrating its effectiveness and superiority. |
API-Bank: A Comprehensive Benchmark for Tool-Augmented LLMs (2023.emnlp-main)
Copied to clipboard
Minghao Li, Yingxiu Zhao, Bowen Yu, Feifan Song, Hangyu Li, Haiyang Yu, Zhoujun Li, Fei Huang, Yongbin Li
| Challenge: | Recent research has demonstrated that Large Language Models (LLMs) can enhance their capabilities by utilizing external tools. |
| Approach: | They propose a runnable evaluation system consisting of 73 API tools and an annotation system for 314 tool-use dialogues with 753 API calls. |
| Outcome: | The proposed benchmark assesses the effectiveness of existing LLMs by analyzing 314 tool-use dialogues with 753 API calls. |