Preprint claims agent memory layer lets open models beat GPT-6
single source· 2 articles · confidence: low · first seen 2026-09-13 20:00 UTC
What this means for you
Nothing to act on yet: both are preprints, with no code, no weights and no evaluation date, so the OSWorld-v2 and Agent's Last Exam numbers cannot be checked. If you pick agent models by leaderboard position, note the paper credits a memory scaffold rather than the underlying models.
Two separate arXiv preprints posted on 13 September describe agent scaffolds — code wrapped around a model — that get better without retraining it. RSIAgent is training-free: curriculum, actor and verifier agents build a frozen memory of environment-specific causal rules that is reused downstream. On OSWorld-v2 and Agent's Last Exam its authors report Kimi-K3 and GLM-5.3 outperforming frontier closed models including GPT-6. Dream-RSI, a different framework, makes an agent's exploration strategy programmable and replays past discovery trees as a cheap simulator to refine it offline. Neither has peer review, released code, or a stated evaluation date; they are two independent results, not one event.
Models in this story
Key facts
- ·RSIAgent reports that Kimi-K3 and GLM-5.3 outperform frontier closed-source models including GPT-6 on OSWorld-v2 and Agent's Last Exam, with no evaluation date given source
- ·RSIAgent is training-free and freezes its constructed memory rather than updating model parameters source
- ·Dream-RSI leaves the underlying coding agent unchanged and makes the exploration strategy a separate programmable layer source
- ·Dream-RSI reuses accumulated discovery trees as a replay simulator to evaluate and refine exploration policies without repeated online evaluation source
- ·Dream-RSI reports results across algorithm engineering, mathematical optimization and GPU kernel engineering, reducing discovery cost in several settings source
What the sources say
- Hugging Face Daily Papers (research) — An orchestration layer that makes an agent's exploration policy explicit and refines it offline against replayed discovery trees.
- Hugging Face Daily Papers (research) — Reports a training-free multi-agent setup that stores environment-specific causal rules in frozen memory for reuse.
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersRSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments2026-09-13
- Hugging Face Daily PapersDream-RSI: Recursive Self-Improvement through Evolving Worlds2026-09-13