Self‑evolving agents continuously rewrite a persistent state from feedback, yet it remains unclear whether they can acquire a rule that was never explicitly given. We introduce CLR‑Bench, a Curriculum of Latent Rules benchmark. Each chain adds one hidden rule on top of the previous levels; a rule operates on an opaque label that denotes a named grouping published on the Web (e.g., the founding members of an airline alliance). The task supplies one example value, practice feedback reveals one additional member and one non‑member, and the test evaluates unseen members and similar non‑members. Probes verify that the base model already knows the members of each grouping and does not guess without feedback, and that the evidence identifies the grouping once labels are read as named groups. An automated pipeline constructs chains from new data domains; we release 20 chains across 10 domains. Six recent self‑evolution methods are evaluated. With the certifying model they barely learn; the best gain is at most 5 points over a frozen agent and they exclude only 6‑12% of unseen members, far less than the 20% achieved by simply keeping the raw practice history. Adding a fixed lesson that states the named‑grouping convention lifts the same methods by 14.5‑16.5 points. Their states retain the revealed member in 75‑98% of cases and almost never the class name, and when practice exposes a missing member they add it to the list. Using the stronger GPT‑6‑Sol, ACE and Dynamic Cheatsheet learn from feedback alone, reaching 78% accuracy and solving 59‑60% of single cells, compared with 53% and 8% with the certifying model, and they surpass raw history. CLR‑Bench therefore measures how well an agent learns, and this ability varies markedly across models.
Review