Research Papers 论文研究 13h ago Updated 8h ago 更新于 8小时前 45

K/V-Cache Interventions Dissociate Representation Alignment from Persona Expression in Decoder-Only Language Models K/V缓存干预将表示对齐与角色表达解耦——面向解码器-only语言模型

K/V-cache interventions transplant target-conditioned key/value trajectories into source-persona generation, enabling structured persona control in decoder-only language models Strong representation-level alignment (V-gap 0.84-0.94) does not guarantee behavioral expression; only mid-layer replacement (layers 9-20) achieves both target-marker expression and preserved lexical diversity Position perturbations (lag and shuffle) uniformly suppress target-persona expression, revealing a common behavio 研究将K/V-cache干预(将目标条件化的K/V轨迹移植到源人格生成中)作为解码器-only语言模型中人格控制的结构性表面 发现表征级对齐与行为表达之间存在一致的解离:所有层带K/V替换都实现强局部V空间对齐,但只有中期层(层9-20)能同时实现目标标记表达和词汇多样性保留 完整替换与中期层替换产生可比的对齐度(V-gap 0.94 vs. 0.89)但词汇多样性差异显著(TTR 0.65 vs. 0.77) 位置扰动(滞后和洗牌)应用不同操作却一致抑制目标人格表达,表明表征相似度指标不足以预测下游人格表达 移植轨迹携带目标自身生成的token历史,因此该干预被定义为轨迹级移植而非孤立的人格

58
Hot 热度
72
Quality 质量
63
Impact 影响力

Analysis 深度分析

TL;DR

  • K/V-cache interventions transplant target-conditioned key/value trajectories into source-persona generation, enabling structured persona control in decoder-only language models
  • Strong representation-level alignment (V-gap 0.84-0.94) does not guarantee behavioral expression; only mid-layer replacement (layers 9-20) achieves both target-marker expression and preserved lexical diversity
  • Position perturbations (lag and shuffle) uniformly suppress target-persona expression, revealing a common behavioral failure mode rather than a strict dissociation
  • Representation similarity metrics alone are insufficient predictors of downstream persona expression; the K/V cache is a controllable but structurally constrained intervention surface
  • The intervention is characterized as trajectory-level transplantation rather than isolated persona-representation injection, with controls confirming representational shifts are not solely explained by imported token history

Why It Matters

This research directly addresses a critical challenge in AI safety and controllability: how to precisely manipulate model behavior without degrading output quality. For practitioners building persona-driven or role-playing systems, understanding that representation alignment and behavioral expression can dissociate means that standard interpretability metrics may be misleading when evaluating intervention strategies. The findings also have implications for alignment research, as they demonstrate that intermediate layers (particularly mid-layers) play a disproportionately important role in persona expression.

Technical Details

  • Intervention method: K/V-cache transplantation where target-conditioned key/value trajectories are inserted into source-persona generation across 13 configurations on Llama-3.1-8B
  • Layer-band analysis: Early, mid (layers 9-20), and late layer replacements all achieved strong V-space alignment (V-gap: 0.91, 0.89, 0.84 respectively), but only mid-layer replacement combined substantial target-marker expression with preserved lexical diversity (TTR 0.77 vs. 0.65 for full replacement)
  • Position perturbation experiments: Both lag and shuffle perturbations uniformly suppressed target-persona expression, indicating structural constraints in the intervention surface
  • Control experiment: Same-token-sequence decoding under source vs. target conditioning reproduced the sign and layer localization of the L28 representational shift, confirming the shift is not explained solely by imported token history
  • Key metrics: V-gap (representation alignment), TTR (type-token ratio for lexical diversity), L28 representational shift analysis

Industry Insight

  • Practitioners should not rely solely on representation-level similarity metrics when designing persona control or alignment interventions; behavioral outcomes may diverge significantly from representational alignment scores
  • Mid-layer K/V cache interventions appear to be the optimal sweet spot for persona expression, suggesting that interpretability efforts targeting these layers could yield the highest return for behavior manipulation
  • The structural constraints revealed by position perturbations indicate that trajectory coherence matters as much as layer selection, pointing toward the need for temporally-aware intervention strategies rather than static cache replacement approaches

TL;DR

  • 研究将K/V-cache干预(将目标条件化的K/V轨迹移植到源人格生成中)作为解码器-only语言模型中人格控制的结构性表面
  • 发现表征级对齐与行为表达之间存在一致的解离:所有层带K/V替换都实现强局部V空间对齐,但只有中期层(层9-20)能同时实现目标标记表达和词汇多样性保留
  • 完整替换与中期层替换产生可比的对齐度(V-gap 0.94 vs. 0.89)但词汇多样性差异显著(TTR 0.65 vs. 0.77)
  • 位置扰动(滞后和洗牌)应用不同操作却一致抑制目标人格表达,表明表征相似度指标不足以预测下游人格表达
  • 移植轨迹携带目标自身生成的token历史,因此该干预被定义为轨迹级移植而非孤立的人格表征注入

为什么值得看

这篇论文揭示了语言模型内部表征与行为表达之间的关键解离现象,为理解K/V-cache在人格控制中的作用机制提供了实证依据。研究结果对开发更精细的人格控制方法和评估指标具有重要参考价值。

技术解析

  • 在Llama-3.1-8B模型上测试了13种干预配置,针对固定源-目标人格对,使用V-gap(V空间对齐度)和TTR(类型-令牌比)作为评估指标
  • 早期、中期、晚期层带K/V替换均实现强局部V空间对齐(V-gap分别为0.91、0.89、0.84),但只有中期层(层9-20)能同时实现目标标记表达和词汇多样性保留
  • 位置扰动实验(滞后和洗牌)表明不同操作类型均会抑制目标人格表达,揭示了K/V-cache作为可控但结构受限的干预表面
  • 通过相同token序列控制实验(在源和目标条件下解码相同token序列)验证了L28表征偏移不能仅由导入的token历史解释

行业启示

  • 当前基于表征相似度的评估指标存在局限,不能单独作为人格表达行为的预测依据,需要结合行为层面的评估方法
  • K/V-cache干预为人格控制提供了结构化途径,但存在层定位敏感性和位置依赖等结构性约束,需精细调优
  • 研究结果强调了在解码器-only架构中理解表征-行为解离的重要性,为开发更可靠的人格控制机制指明了方向

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 LLaMA LLaMA Alignment 对齐 Research 科学研究 Inference 推理