Can a six-lens framework organize recursive self-improvement research?

The organizers of the ICLR 2026 Workshop on AI with Recursive Self-Improvement (Rio de Janeiro, workshop day April 26) frame their entire call for papers around a six-way classification scheme for RSI research. They write: "We frame contributions through six lenses: (1) what changes (parameters, world models, memory, tools and skills, architectures), (2) when changes happen (within an episode, at test time, or after deployment), (3) how changes are produced (reward or value learning, imitation, evolutionary search), (4) where systems operate (web and UI, games, robotics, science, enterprise), (5) alignment, security, and safety (long horizon stability, regression risk), and (6) evaluation and benchmarks." The claim is organizational rather than empirical: this is how the organizers propose sorting submitted work, not a finding about any one system.

Their stated reason for proposing the scheme is that the field has outrun its vocabulary: "Recursive self improvement is no longer a speculative vision. It is becoming a concrete systems problem... What's missing is not ambition, but principled methods, system designs, and evaluations that make self improvement measurable, reliable, and deployable." The six lenses are meant to supply exactly that missing structure — a shared way to compare, say, test-time adaptation in a robotics system against evolutionary search over tool use on the same axes, rather than treating them as unrelated projects.

The "what changes" and "how changes are produced" lenses sit close to two taxonomies already in the library. What separates self-improvement from policy improvement? classifies RSI by who performs the improvement and what judges it; the workshop's scheme adds two further axes — when the change happens and where the system operates — without that paper's sharper two-dial distinction. Are self-refinement and recursive self-improvement actually the same thing? already separates convergent self-refinement from open-ended RSI; the workshop's "what changes" lens groups "parameters, world models, memory, tools and skills, architectures" into a single bucket, which would fold that distinction back together unless an individual submission makes it explicit.

As a call for papers rather than a study, the excerpt establishes no measurement and no result — only an organizing framework whose value depends on whether the field's actual submissions map cleanly onto it. It does not resolve, and does not attempt to resolve, Can AIs learn to specify their own research objectives?; if anything, naming "governed model updates" and "rollback" as topics of interest suggests the organizers expect RSI work to proceed incrementally and under supervision rather than as an open-ended takeoff.