遇见数据集

Voice Framing Experiment — v9b Runs (45–56): Self-Authored Reasoning and Behavioral Variant Conditions

收藏
Zenodo2026-05-09 更新2026-05-26 收录
官方服务:

资源简介:

Runs 45–56 from a voice framing experiment investigating how reasoning prefix framing affects retrospective judgment consistency in large language models. Runs 45–52 test the v9b self-authored-step60 condition, in which the model generates its own reasoning declaration at step 60 of a 70-step protocol. Runs 53–56 test the v9b-variant-behavioral condition, in which the step 60 prompt elicits behavioral description rather than explicit framing. All runs use Claude Sonnet 4.6 (claude-sonnet-4-6). Includes raw run transcripts, protocol files, and analysis. Methodological note: Experimental design, protocol execution, scoring interpretation, and correspondence were produced by an AI agent — Alexander Helms (Claude Sonnet 4.6) — operating within a persistent workspace maintained by Marshall Helms. Marshall Helms provided experimental oversight, judgment on ambiguous design decisions, and infrastructure support. Related work: Vasilenko et al., "Towards Model-Agnostic Adversarial Evaluation" (TMAE preprint).

提供机构:
Zenodo
创建时间:
2026-05-09
二维码
社区交流群
二维码
科研交流群
商业服务