研究探讨大模型编辑权重能力：前沿模型表现差异明显

在LLM中。但众所周知，真实场景下的LLM训练十分复杂，并且高度依赖预训练和后训练两个阶段。正如@AlexKrentsel在X上所指出的，似乎只有前沿模型才能完成这项任务。我们可以合理推测，这可能是OpenAI或Anthropic的工程师在做RSI，或者通过在这些任务上微调模型来开展其他工作。虽然这既不充分也不必要，但我们也可以合理地说，如果某个模型纯粹通过预训练获得了这些技能，那么智力水平相近的模型应该具备相似的原始权重操控能力。我以Artificial Analysis衡量的一般智力为 기준，对SOTA模型做了一个快速测试，前提是假设该评测是公平的。有趣的是，在前沿模型中，Kimi K3.1 Pro是一个异常值，它完成该任务的水平只是中等，而其AA分数却与Fable 5.1大致相同。这要么说明该能力是涌现出来的，而Kimi只是一个糟糕的benchmaxxing模型；要么说明Google、OpenAI和Anthropic在后训练阶段专门做了微调并试验了相关想法，而作为中国模型公司的Moonshot则有着不同的侧重点，没有针对这项任务做适配。免责声明：关于先前文献的大量研究工作由LLM协助完成。参考文献 N. Elhage et al., “A Mathematical Framework for Transformer Circuits,” Transformer Circuits, 2021. [Online]. Available: https://transformer-circuits.pub/2021/framework/index.htmlC. Olsson et al., “In-context Learning and Induction Heads,” 2022, arXiv:2209.11895. doi: 10.48550/arXiv.2209.11895.R. Maclin and J. W. Shavlik, “Incorporating Advice into Agents that Learn from Reinforcements,” in Proc. 12th Nat. Conf. Artificial Intelligence (AAAI), 1994, pp. 694–699. [Online]. Available: https://cdn.aaai.org/AAAI/1994/AAAI94-106.pdfR. Maclin and J. W. Shavlik, “Creating advice-taking reinforcement learners,” Machine Learning, vol. 22, pp. 251–281, 1996, doi: 10.1007/BF00114730.A. L. Zobrist, “A model of visual organization for the game of GO,” in Proc. AFIPS Spring Joint Computer Conference, vol. 34, 1969, pp. 103–112, doi: 10.1145/1476793.1476819.D. Lindner, J. Kramár, S. Farquhar, M. Rahtz, T. McGrath, and V. Mikulik, “Tracr: Compiled Transformers as a Laboratory for Interpretability,” in Adv. Neural Inf. Process. Syst., vol. 36, 2023. [Online]. Available: https://arxiv.org/abs/2301.05062R. A. Genadi, M. S. Nwadike, N. Mukhituly, T. Hiraoka, H. AlQuabeh, and K. Inui, “Sycophancy Hides Linearly in the Attention Heads,” in Proc. 19th Conf. European Chapter Assoc. Comput. Linguistics (EACL), vol. 1, 2026, pp. 6896–6912, doi: 10.18653/v1/2026.eacl-long.324.S. Chen, Z. J. Ying, L. R. Smith, J. Wertheimer, and N. Shapira, “Tracing mechanisms of sycophantic agreement in language models,” 2026, arXiv:2609.35822. doi: 10.48550/arXiv.2609.35822.W. Chen et al., “From Yes-Men to Truth-Tellers: Addressing Sycophancy in Large Language Models with Pinpoint Tuning,” in Proc. 41st Int. Conf. Machine Learning (ICML), PMLR, vol. 235, 2024, pp. 6950–6972. [Online]. Available: https://proceedings.mlr.press/v235/chen24u.htmlA. Turner, L. Thiergart, D. Udell, G. Leech, U. Mini, and M. MacDiarmid, “Activation Addition: Steering Language Models Without Optimization,” 2023, arXiv:2308.10248, ver. 1. [Online]. Available: https://arxiv.org/abs/2308.10248v1N. Rimsky, N. Gabrieli, J. Schulz, M. Tong, E. Hubinger, and A. M. Turner, “Steering Llama 2 via Contrastive Activation Addition,” in Proc. 62nd Annu. Meeting Assoc. Comput. Linguistics (ACL), vol. 1, 2024, pp. 15504–15522, doi: 10.18653/v1/2024.acl-long.828.A. Zou et al., “Representation Engineering: A Top-Down Approach to AI Transparency,” 2023, arXiv:2310.01405. doi: 10.48550/arXiv.2310.01405.A. Arditi, O. Obeso, A. Syed, D. Paleka, N. Panickssery, W. Gurnee, and N. Nanda, “Refusal in Language Models Is Mediated by a Single Direction,” 2024, arXiv:2406.11717. doi: 10.48550/arXiv.2406.11717.<a href=