LM-Oriented Web Observations

Convert pages and search hits into tokenizer-stable text with numbered links, readability extracts, and media placeholders shaped for LM agents—not browsers.

Browser ACIs for LMs should emit text observations, not raw HTML. WebGPT’s environment: search via an API to a simplified results page; fetches cleaned with Readability; remaining HTML → text (html2text); PDFs via a text extractor; unsupported types return an error. Links become a stable special form such as 【<id>†<text>†<domain>】 so the model can cite and re-click by ID without burning fragile URL tokens. Images collapse to [Image: <alt>]; titles carry page title and domain.

This is the web sibling of bounded search observations and collapsed observations: informative, capped, consistent columns under ACI design principles. Pair with token-efficient quote observations when the agent must extract evidence. Filter source domains that invite answer- copying when the task is open-book QA, and apply eval observation anti-leakage in benchmark environments.

---

Sources

• WebGPT (Nakano et al.), Appendix A