arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2602.04903cs.LG
收藏

注意性能差距:特征引导中的能力-行为权衡

Mind the Performance Gap: Capability-Behavior Trade-offs in Feature Steering

BAISH, Universidad de Buenos Aires, Argentina(BAISH,布宜诺斯艾利斯大学,阿根廷);AISAR, AI Safety Argentina(AISAR,人工智能安全阿根廷);Instituto de Investigación en Informática LIDI, Universidad Nacional de La Plata, Argentina(信息研究所LIDI,拉普拉塔国立大学,阿根廷);FAIR, IALAB UBA, Universidad de Buenos Aires, Argentina(FAIR,IALAB UBA,布宜诺斯艾利斯大学,阿根廷);Università degli Studi di Genova, Italy(热那亚大学,意大利)

Eitan Sprejer, Oscar Agustín Stanchi, María Victoria Carro, Denise Alejandra Mester, Iván Arcuschin

AI总结:

研究发现特征引导方法在控制行为时显著降低模型性能,揭示了能力与行为之间的权衡问题。

摘要:

特征引导已成为通过直接操控内部表示来控制大语言模型(LLM)行为的一种有前景的方法,相较于提示工程具有优势。然而,其在现实世界应用中的实际有效性仍不明确,尤其是在与输出质量可能存在的权衡方面。我们证明,特征引导方法即使在成功控制目标行为的情况下,也会显著降低模型性能,这是一个关键的权衡。具体而言,我们评估了Goodfire的Auto Steer在14个引导查询(涵盖无害和安全相关行为)上,使用Llama-8B和Llama-70B对171个大规模多任务语言理解(MMLU)问题进行测试,测量准确率、连贯性和行为控制。我们的发现表明,Auto Steer成功地修改了目标行为(在Llama-8B上达到3.33 vs. 2.98,在Llama-70B上达到3.57 vs. 3.10),但导致了显著的性能下降:在MMLU问题上的准确率从Llama-8B的66%降至46%,从Llama-70B的87%降至73%,连贯性从4.62降至2.24,从4.94降至3.89。简单的提示方法实现了最佳的整体平衡。这些发现突显了当前特征引导方法在实际部署中的局限性,其中任务性能不能被牺牲。更广泛地说,我们的工作表明,机械控制方法面临根本的能力-行为权衡,必须在部署前通过实证来表征。

英文摘要:

Feature steering has emerged as a promising approach for controlling LLM behavior through direct manipulation of internal representations, offering advantages over prompt engineering. However, its practical effectiveness in real-world applications remains poorly understood, particularly regarding potential trade-offs with output quality. We show that feature steering methods substantially degrade model performance even when successfully controlling target behaviors, a critical trade-off. Specifically, we evaluate Goodfire's Auto Steer against prompt engineering baselines across 14 steering queries (covering innocuous and safety-relevant behaviors) on 171 Massive Multitask Language Understanding (MMLU) questions using Llama-8B and Llama-70B, measuring accuracy, coherence, and behavioral control. Our findings show that Auto Steer successfully modifies target behaviors (achieving scores of 3.33 vs. 2.98 for prompting on Llama-8B and 3.57 vs. 3.10 on Llama-70B), but causes dramatic performance degradation: accuracy on the MMLU questions drops from 66% to 46% on Llama-8B and 87% to 73% on Llama-70B, with coherence falling from 4.62 to 2.24 and 4.94 to 3.89 respectively. Simple prompting achieves the best overall balance. These findings highlight limitations of current feature steering methods for practical deployment where task performance cannot be sacrificed. More broadly, our work demonstrates that mechanistic control methods face fundamental capability-behavior trade-offs that must be empirically characterized before deployment.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑