Calibrating Small Language Models for Claim Check-Worthiness Detection
为解决小模型在检测声明可信度时准确性低的问题,提出NN-PPI方法,在不重新训练模型的情况下校准预测,提高准确性并降低成本。
为解决小模型在检测声明可信度时准确性低的问题,提出NN-PPI方法,在不重新训练模型的情况下校准预测,提高准确性并降低成本。
为解决LLM模拟人类行为时缺乏真实认知限制的问题,通过引入具有认知边界约束的用户模拟器框架CBUS,并采用两种测试策略来更准确地模仿年轻读者的行为。
This work proposes a privacy-preserving, instruction-free approach to personalized text rewriting that achieves style transfer using only a user’s own documents, eliminating the need for in-context prompting during inference. The method leverages multiple local auxiliary large language models to construct training pairs, fine-tunes base models ranging from 0.5B to 7B parameters via LoRA, and incorporates automatic text chunking with length-aware generation. Evaluation on 219 scientific paper rewriting pairs reveals a performance plateau at an automatic composite score of 0.69, indicating that even small models are sufficiently effective. Human evaluations by LLM judges show an average reduction of over 20% in perceived “AI-ness” of outputs, which further decreases with larger model scales. This study is the first to demonstrate a scaling saturation effect in this task, validating the efficacy and practicality of lightweight, single-user adaptation frameworks.
This work addresses the deployment challenges of existing parameter-efficient fine-tuning methods—such as LoRA and Soft Prompting—which require modifications to the model’s computation graph and are thus incompatible with high-throughput inference engines like vLLM. The authors propose ART, a novel approach that treats visual inputs as learnable “computational art.” By optimizing only pixel-level visual prompts while keeping all parameters of the multimodal large language model frozen, ART injects task-specific information without altering the model architecture or computation graph. Consequently, it seamlessly integrates with precompiled inference systems and supports arbitrary fine-tuning objectives. Experimental results demonstrate that ART achieves performance on par with LoRA on Qwen-series models across multiple textual benchmarks, particularly excelling in mathematical reasoning and structured tool-use tasks.
This work addresses the challenge of high-throughput, low-latency fact-checking in multilingual settings by proposing a modular system comprising three stages: claim detection, evidence retrieval and re-ranking, and veracity prediction. Instead of relying on general-purpose large language models, the system employs task-specific fine-tuned models—XLM-RoBERTa-Large for claim detection, mmBERT-base for stance classification, and a SetFit-based multilingual re-ranker—achieving robust and efficient performance across 114 languages for claim detection and 28 languages for veracity prediction. Experimental results demonstrate that this approach significantly outperforms mainstream large language models such as GPT-5.2, Claude Opus 4.6, and Qwen3-8B in terms of inference efficiency, data privacy, and resource consumption, while maintaining comparable or superior accuracy.
为解决小模型在检测声明可信度时准确性低的问题,提出NN-PPI方法,在不重新训练模型的情况下校准预测,提高准确性并降低成本。
为解决LLM模拟人类行为时缺乏真实认知限制的问题,通过引入具有认知边界约束的用户模拟器框架CBUS,并采用两种测试策略来更准确地模仿年轻读者的行为。
This work proposes a privacy-preserving, instruction-free approach to personalized text rewriting that achieves style transfer using only a user’s own documents, eliminating the need for in-context prompting during inference. The method leverages multiple local auxiliary large language models to construct training pairs, fine-tunes base models ranging from 0.5B to 7B parameters via LoRA, and incorporates automatic text chunking with length-aware generation. Evaluation on 219 scientific paper rewriting pairs reveals a performance plateau at an automatic composite score of 0.69, indicating that even small models are sufficiently effective. Human evaluations by LLM judges show an average reduction of over 20% in perceived “AI-ness” of outputs, which further decreases with larger model scales. This study is the first to demonstrate a scaling saturation effect in this task, validating the efficacy and practicality of lightweight, single-user adaptation frameworks.
This work addresses the deployment challenges of existing parameter-efficient fine-tuning methods—such as LoRA and Soft Prompting—which require modifications to the model’s computation graph and are thus incompatible with high-throughput inference engines like vLLM. The authors propose ART, a novel approach that treats visual inputs as learnable “computational art.” By optimizing only pixel-level visual prompts while keeping all parameters of the multimodal large language model frozen, ART injects task-specific information without altering the model architecture or computation graph. Consequently, it seamlessly integrates with precompiled inference systems and supports arbitrary fine-tuning objectives. Experimental results demonstrate that ART achieves performance on par with LoRA on Qwen-series models across multiple textual benchmarks, particularly excelling in mathematical reasoning and structured tool-use tasks.
This work addresses the challenge of high-throughput, low-latency fact-checking in multilingual settings by proposing a modular system comprising three stages: claim detection, evidence retrieval and re-ranking, and veracity prediction. Instead of relying on general-purpose large language models, the system employs task-specific fine-tuned models—XLM-RoBERTa-Large for claim detection, mmBERT-base for stance classification, and a SetFit-based multilingual re-ranker—achieving robust and efficient performance across 114 languages for claim detection and 28 languages for veracity prediction. Experimental results demonstrate that this approach significantly outperforms mainstream large language models such as GPT-5.2, Claude Opus 4.6, and Qwen3-8B in terms of inference efficiency, data privacy, and resource consumption, while maintaining comparable or superior accuracy.