Today's edition highlights diverse approaches to self-improvement, from agents managing their own context windows to reflecting on internal representations for skill evolution.
The loop: The Rep2Skill agent models its own internal representation trajectories to localize execution errors, generating textual feedback to evolve its external skills.
A representation-guided framework allows LLM agents to improve their external textual skills by reflecting on their own internal execution state representations.
The loop: A coding agent is trained to decide when and how to compact its own context during long-horizon tasks, using task-success rewards to improve its policy.
A framework trains long-horizon coding agents to manage their own context by deciding when to compact and what working state to preserve, optimizing these decisions through reinforcement learning.
The loop: LiteEvo meta-agents mine agent trajectories for reusable components to evolve a harness library, which improves the agent's performance on unseen tasks.
A lightweight harness-evolution algorithm uses tool-free meta-agents to mine trajectories and build a reusable component library, improving agent performance on unseen tasks at lower cost.
The loop: An advisor model uses reflection to propose corrections to its own decisions, then uses self-distillation from a feedback-conditioned copy to improve its future advice.
A method for steering frozen language models pairs outcome-based reinforcement learning with selective self-distillation to improve a trainable advisor's natural-language advice.
The loop: DSX MaxLPS optimizes power usage in AI factories, which improves throughput for the models running in that factory.
Every unused watt is capacity left on the table. AI factories are typically provisioned for the unlikely moment when every GPU reaches peak power, creating a...
The loop: Torch Spyre uses the CRCR relay to automate CI testing, which improves the stability of PyTorch for its own development.
TL;DR PyTorch’s Cross-Repository CI Relay (CRCR) gives out-of-tree accelerators a clean, scalable way to plug into upstream CI - and it leaves each backend free to decide which of PyTorch’s...
The loop: A safety case framework investigates misalignment incidents, which improves technical safeguards for the training of successor frontier models.
Our early guidelines for safety cases in frontier AI training cover technical safeguards, operational practices, and investigating misalignment incidents
