Tag
The paper proposes ChainDoRA, a parameter-efficient fine-tuning method for large language models that combines Tensor-Train factorization with weight-decomposed low-rank adaptation, achieving a 90.62% reduction in trainable parameters and improved accuracy over LoRA and DoRA.
CARD proposes a hierarchical framework for personalized text generation that clusters users and uses reward-guided decoding, demonstrating improved quality and efficiency on LaMP benchmarks.
The paper diagnoses failures in test-time reinforcement learning for medical QA due to answer-space structure and introduces PROSE, which rewards reasoning quality to improve model performance without labeled data.
Presents TEXAS, a method for downstream adaptation of Mixture-of-Experts LLMs that discovers task-relevant experts via correctness-conditioned activations and applies token-level supervision allocation, improving performance across multiple benchmarks.
RimRule proposes a neuro-symbolic method that distills compact, interpretable rules from failure traces using the Minimum Description Length principle, improving LLM tool-use performance without modifying weights, and demonstrating rule portability across models.
This paper investigates parameter-efficient strategies for adapting large language models to 3D CT report generation, introducing RAD3D-Prefix, a lightweight diagnostic-prior conditioning framework that keeps the LLM frozen and requires minimal trainable parameters. It shows that freezing larger LLMs (~1B+) and training only lightweight projection layers provides a superior trade-off between performance, generalization, and computational efficiency.
Hybrid-LoRA proposes a framework that selectively applies full fine-tuning to a small subset of modules while using LoRA for the rest, achieving performance near full fine-tuning with significantly lower computational cost. Experiments show improvements of up to 5.65% over existing parameter-efficient baselines.