推理坡度作为低成本路径假说: 关于 Llm 内容采纳机制的一份第一人称观察报告
Year: 2026
Type: preprint
Abstract: Existing research on Generative Engine Optimization (GEO) defines the core problem as improving the *visibility* of content in generative engine outputs (Aggarwal et al., 2024, KDD 2024). This paper identifies a gap downstream of visibility: when a document has already entered an LLM's candidate context, what determines whether it is structurally adopted into the generated response? Through a controlled experiment across three LLM platforms (Claude Sonnet 4.6, Qwen 3.6 Plus, DeepSeek public web interface) with two structurally different documents × 18 trials, we observe a consistent directional asymmetry in M2 adoption depth (+1.67 to +2.34 in favor of reasoning-slope documents), while M1 visibility differences remain negligible (-0.22). Using Simon's (1955, 1957) bounded rationality as an external theoretical anchor, we propose that this asymmetry can be understood as a constrained computational system's systematic preference for low-cost continuation paths — and that *reasoning slope* (Meng, 2026, Zenodo DOI 10.5281/zenodo.19027371), with its three structural elements (question frame, causal paving, perspective lock), may constitute one concrete content-layer implementation of this preference, each element reducing a specific type of marginal generation cost. This paper does not claim that reasoning slope is the only implementation of this preference, does not claim cross-platform or cross-temporal stability of the observations, and provides three falsifiable paths for independent researchers. This paper also distinguishes reasoning slope from prompt engineering: the former operates on the internal structure of uploaded documents, while the latter operates on the dialogue input layer — the two target different cost items and can vary independently within the same LLM session.
Author: meng, junwoo
Language: Chinese
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7153762924
