Recursa lifts DeepSeek performance via self-improvement

The important idea is practical system-level self-improvement rather than weight updates, but the benchmark results still need stronger validation before they prove general RSI.

• –Stores reusable skills, instructions, memories, and reflections as editable system state
• –Parallel learners create local improvements while validation controls what enters the shared repository
• –Particularly promising for specialized company domains with abundant documents but scarce expert labels
• –Reported gains come from limited runs and model-based judging, so independent evaluators, repeated trials, and ablations are essential
• –The clearest near-term path is guarded deployment with approval gates, rollback, and repository versioning