📰 2026-03-14 · 每日技术速递
精选来自 arXiv cs.LG / cs.CL 的深度技术论文,聚焦大模型训练、推理、对齐与系统优化。
1. Matching Features, Not Tokens: Energy-Based Fine-Tuning of Language Models
- 来源:arXiv cs.LG
- 链接:https://arxiv.org/abs/2603.12248
- 摘要:arXiv:2603.12248
2. A Quantitative Characterization of Forgetting in Post-Training
- 来源:arXiv cs.LG
- 链接:https://arxiv.org/abs/2603.12163
- 摘要:arXiv:2603.12163
3. IsoCompute Playbook: Optimally Scaling Sampling Compute for LLM RL
- 来源:arXiv cs.LG
- 链接:https://arxiv.org/abs/2603.12151
- 摘要:arXiv:2603.12151
4. SciMDR: Benchmarking and Advancing Scientific Multimodal Document Reasoning
- 来源:arXiv cs.CL
- 链接:https://arxiv.org/abs/2603.12249
- 摘要:arXiv:2603.12249
5. Sparking Scientific Creativity via LLM-Driven Interdisciplinary Inspiration
- 来源:arXiv cs.CL
- 链接:https://arxiv.org/abs/2603.12226
- 摘要:arXiv:2603.12226
6. CLASP: Defending Hybrid Large Language Models Against Hidden State Poisoning Attacks
- 来源:arXiv cs.CL
- 链接:https://arxiv.org/abs/2603.12206
- 摘要:arXiv:2603.12206
更新时间:2026-03-14 · 由 AI Reading List Bot 自动生成
评论