arXiv cs.CL· Tianda Sun, Dimitar Kazakov·· 12 小时前AI 评分31
从目标选择到条件复用：审计工具调用模型中的残差干预
From Target Selection to Condition Reuse: Auditing Residual Interventions in Tool-Use Models
AI 导读
研究在 Qwen3.5-27B 的工具调用任务上审计残差写入干预：两种针对具体案例的变体将目标选择正确率从 35/128 提升至 119/128、从 33/128 提升至 114/128，但当熟悉事实与标识符在同一选择任务中重组后增益消失，对候选是否合格或应被选择的独立判断也几乎无改善。
正文 · 原文
Abstract:Activation interventions can make a language model produce a target answer, but can the model use the intended condition in other computations? We study this question in Qwen3.5-27B on tool-use tasks where time or permissions determine eligibility and selection. Across 32 fresh worlds, two case-specific residual-writing variants raise familiar target-selection correctness from 35/128 to 119/128 and from 33/128 to 114/128. Yet these gains disappear when familiar facts and identifiers are recombined within the same selection task, while selection of previously rewarded identifiers increases. The gains persist across output formats but yield little improvement in independent judgments about whether a candidate is eligible or should be selected. Internal edits can nevertheless change how a source fact affects a judgment: in a prespecified secondary comparison, an early-layer source patch changes 50/64 judgments when that source is used, with one adverse change among 64 cases where it is not. This selective effect does not ensure correct complete actions. Together, these results separate target-selection control, condition-sensitive judgments, and complete-action correctness. Evaluating how an edited state is used, rather than only its trained answer, exposes these differences.
| Comments: | 21 pages, 5 figures. Code: this https URL |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.31686 [cs.CL] |
| (or arXiv:2610.31686v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.31686 arXiv-issued DOI via DataCite |
Submission history
From: Tianda Sun [view email]
[v1]
Mon, 5 Oct 2026 11:46:10 UTC (48 KB)
来源：arXiv cs.CL · arxiv.org