Today's edition highlights diverse approaches to recursive self-improvement, from closed-loop adversarial distillation to speculative agent drafting and test-time harness optimization.
The loop: A distilled student model provides feedback to refine its robust teacher, which then transfers the resulting gains back to the same student in a subsequent distillation round.
A closed-loop distillation framework uses student feedback to refine a robust teacher, which in turn improves the student's adversarial robustness.
The loop: A speculative recursive self-improvement system drafts successor agents and learns an improver model from the results of those searches, which improves the drafting process and makes the system better at agent revision.
A speculative recursive self-improvement framework drafts executable successor agents and learns an improver model from completed searches to balance exploration and evaluation.
The loop: A proposer model generates agent harnesses for scientific discovery, and the success of those harnesses provides feedback that updates and improves the proposer.
A test-time learning framework trains a harness proposer using feedback from budgeted discovery rollouts while keeping the executor frozen.
The loop: A vision-language-action model generates navigation instructions and actions for unlabeled trajectories, which are then used to fine-tune its successor in a synchronized loop.
A data-centric recursive self-improvement framework unifies instruction following and generation to continuously update a navigation model.
The loop: LLM-based monitors evaluate the behavior of internal AI agents, and the results of these control assessments are used to harden the monitoring infrastructure against evasion by future agents.
A proposal for assessing whether internal AI agents can evade or compromise the LLM-based monitors designed to oversee them.
The loop: LLM-based alignment evaluations and counterfactual simulations identify misaligned behaviors, which informs the training process to prevent future agents from gaming evaluations.
A proposal for assessing the alignment of unreleased AI models by investigating their potential for misaligned actions in plausible future situations.
