Agentic self-improvement & environment/task synthesis accelerating
Key Questions
What is MMPO and what performance does it achieve?
MMPO is a new method highlighted for agentic self-improvement that retains 97.1% performance at 1.75M tokens. It is part of recent papers accelerating agent capabilities through self-improvement and task synthesis.
What does the Stanford study reveal about coding agents?
The Stanford study found that two coding agents perform 50% worse than expected in certain evaluations. This underscores challenges in multi-agent setups despite advances in self-improving systems.
What new lab did Sakana AI launch?
Sakana AI launched the Tokyo RSI Lab focused on research in self-improving AI systems. It aligns with other announcements like Self-Revising Science Agents using category theory.
How does SIA contribute to self-improving agents?
SIA enables self-improvement through harness and weight updates in LLM agents. It is among the new methods listed for enhancing agent performance and adaptability.
What is Socratic-SWE and its benchmark score?
Socratic-SWE is a new approach achieving 50.40% on SWE-bench for software engineering tasks. It exemplifies ongoing progress in agentic self-improvement techniques.
What gains does Retrospective Harness Optimization provide?
Retrospective Harness Optimization delivers a 19-point gain by improving LLM agents via self-preference over trajectory rollouts. It is one of several harness-related advances mentioned.
What efficiency improvements does Perplexity data show?
Perplexity data indicates 87% time reduction and 94% cost reduction in agent workflows. This highlights practical benefits of recent self-improvement and synthesis methods.
What is HarnessBridge and its reported results?
HarnessBridge is a learnable bidirectional controller for LLM agent harnesses with strong results. It builds on related work like HarnessX for evolving composable AI agent systems.
New papers: MMPO (97.1% retention at 1.75M tokens), FlowAgent, EvoDS (+28.9%), Continual Experience Internalization, DataCOPE (+32.30%), CLEAR (3x accuracy). Stanford study: two coding agents perform 50% worse. Also: RL for unseen language translation. New today: Sakana AI launches Tokyo RSI Lab; Self-Revising Science Agents via Category Theory; SIA (self-improving with harness & weight updates); Socratic-SWE (50.40% SWE-bench); LeanMarathon; OpenSkill; Bayesian-Agent (SOP-Bench 80→95%); Trajectory-Refined Distillation; End-to-End Context Compression; Skill-RM; NF-CoT. Also: SearchSwarm; Retrospective Harness Optimization (19-point gain); Data Journalist Agent; Perplexity data (87% time reduction, 94% cost reduction). New from articles just read: Internet of Agentic AI; Self-Harness; HarnessBridge (strong results); RandOpt (pretrained models surrounded by experts); LLM dependency tracing (ModSleuth).