| Fu Shengqiang (付胜强)* **,Chen Shudong* **,Chen Xiaoxiao**,Yu Yong*,Du Jiale* **,Wang Ergang**.[J].高技术通讯(英文),2026,32(2):156~168 |
|
| SI-FACT: mitigating knowledge conflict via self-improving faithfulness-aware contrastive tuning |
| |
| DOI：10. 3772 / j. issn. 1006-6748. 2026. 02. 005 |
| 中文关键词: |
| 英文关键词: knowledge conflict, contrastive learning, self-instruct, large language models,contextual faithfulness |
| 基金项目: |
| Author Name | Affiliation | | Fu Shengqiang (付胜强)* ** | (* Institute of Microelectronics of Chinese Academy of Sciences, Beijing 100029, P. R. China)
(** University of Chinese Academy of Sciences, Beijing 100049, P. R. China) | | Chen Shudong* ** | | | Chen Xiaoxiao** | | | Yu Yong* | | | Du Jiale* ** | | | Wang Ergang** | |
|
| Hits: 105 |
| Download times: 127 |
| 中文摘要: |
| |
| 英文摘要: |
| Large language models (LLMs) often produce unfaithful responses in knowledge-intensive tasks due to knowledge conflict—a tendency to rely on internal memory rather than user-provided context.This work aims to enhance the model’s trust in user-provided contextual information, encouraging it to ground its generation more faithfully on external evidence. To address this issue, we propose self-improving faithfulness-aware contrastive tuning ( SI-FACT), a novel self-improving framework that enhances contextual faithfulness while reducing data dependence. SI-FACT employs a self-instruct mechanism that enables the base LLM to automatically generate structured contrastive data consisting of anchor, faithful ( positive), and unfaithful ( negative) samples, thus eliminating costly manual annotation. Through contrastive learning, the model learns to align faithful responses closer and separate unfaithful ones within the representation space, effectively reinforcing trust in contextual evidence. Experiments on ECARE_KRE and COSE_KRE benchmarks show that SI-FACT, built on Llama-3-8B-Instruct, improves the contextual consistency rate (CCR) by 6. 2% over the best baseline,while using fewer samples. These results demonstrate that SI-FACT achieves strong contextual faithfulness with high data efficiency, offering a practical path toward more reliable and proactive LLMs. |
|
View Full Text
View/Add Comment Download reader |
| Close |
|
|
|
