Introduces TrACE, a training-free controller that allocates inference-time compute adaptively by measuring inter-rollout action agreement. High agreement signals low difficulty, allowing for immediate commitment, while low agreement triggers additional rollouts.
RSI Bench Relevance: Efficient "Adaptive Compute" for agents. Directly impacts the efficiency of RSI loops by optimizing the compute-to-improvement ratio.
Proposed at ICLR 2026 RSI Workshop. Uses real-data "goalpost" questions to ground asymmetric self-play. The teacher generates curriculum variants to close the gap to hard exploration challenges.
RSI Bench Relevance: Solve the "Grounding" problem in open-ended data generation. Curriculum-based RSI is the next step for coding models.
Self-Improving 4D Perception via Self-Distillation
Introduces SelfEvo, a framework for continuous improvement of 4D perception models using unlabeled videos via spatiotemporal context asymmetry and self-distillation.
RSI Bench Relevance: Extends RSI to vision/perception. Proves that the "Self-Evolution" paradigm is cross-modal.
Self-Debias: Self-correcting for Debiasing Large Language Models
A progressive framework for intrinsic self-correction. Treats debiasing as resource redistribution in the output probability space, using online self-improvement with consistency filtering.
RSI Bench Relevance: "Alignment via RSI." Shows that safety and debiasing can be achieved through internal self-correcting loops rather than just external oversight.
X/Twitter Signal Monitoring (Morning Update)
Real-time Signals | 2026-04-12
Workshop Launch: Official Call for Papers for the 1st Workshop on Recursive Self-Improvement (RSI) at ICLR 2026. Topics include RSI loops, test-time adaptation, and model editing.
New RSI Lab: Azalia Mirhoseini and Anna Goldie launch @RicursiveAI, focusing on the "AI-Chip RSI Loop"—AIs designing their own chips to accelerate self-improvement.
Singularity Sentiment: X sentiment shift: "2026: the singularity. The year recursive self-improvement becomes impossible to ignore."
Signal Relevance: Transition from "AI research" to "AI-led infrastructure" (Ricursive Intelligence). The RSI workshop at ICLR 2026 marks the academic formalization of the field.