Can an agent explore a new environment, learn its causal structure, and keep improving without updating its model weights? We introduce RSIAgent, a framework for recursive self-improvement through autonomous exploration. Using Kimi-K3 and GLM-5.3 as base models, RSIAgent outperforms GPT-6 Astra on both OSWorld 2.0 and Agents’ Last Exam.... RSIAgent decides what to explore, executes tasks, verifies outcomes, and consolidates stable action-condition-outcome relationships into memory for future use. With the underlying model weights fixed, RSIAgent achieves: - 78.98% Partial Score on OSWorld 2.0 (0808 offline), compared with 72.60% for GPT-6 Astra - 84.82% on Agents’ Last Exam (Near-term), compared with 82.26% for GPT-6 Astra We call this Scaling Experience. Agents can continue improving by acquiring, verifying, and reusing their own experience while the underlying model weights remain fixed. Links in the reply below.show more

Learn more about RSIAgent - Blog: - Code: - Project: - Paper:

it’s called scaling test time interaction

Weight-frozen RSI is just a harness that keeps score. The env map is the learning signal, not another finetune.

Memory or knowledge is an extension of the encoding in model weights. If we treat them as one system it’s not surprising to keep one subsystem static and update other parts of system through rl

Huge leap! Happy to apply!

Scaling experience is a really interesting direction. Fixed weights mean the same model can keep getting better as it gets access to more environments and compute.

To be honest, I understood 22% of what have been said here but I love the concept, and how every new agent framework is distilling and evolving from previous ideas, pushing RSI to a new frontier and dueling to battle all known concepts of how to approach to what SDLC used to be.

Impressive that fixed-weight agents can still boost performance, curious how stable the learned action‑condition‑outcome memory is across domain shifts

Self-improvement without touching weights feels like the practical path — cheaper than retraining and you keep the base model swappable. Does the causal structure it learns transfer to a fresh environment or does it start from zero?

Excellent info

Can recursive exploration improve agents without making evaluation drift across environments?

I think an agentic evolution keeps in possibility to change the model weights based on approved hypotheses. It is great two gated architecture, where model evolve slowly, but changes in runtime and this change possibly is core feature.

Is time to quit my GPT subscription and switch to Kimi k3

"Scaling Experience" is a great framing, but let's be honest about what it actually is: structured in-context retrieval with a fancy name. The real question isn't whether memory helps — of course it does — it's whether this scales past the context window. A 200k-token memory dump isn't "experience," it's a haystack with good PR. Show me RSIAgent at month 6, not day 1.

The interesting part is the evaluation loop. How does RSIAgent distinguish causal learning from a policy that simply gets better at exploration?

@grok 这是个啥东西，有啥用？解决社会问题。

Super interesting, How does this method compare to continual learning where weights do get updated?

Computer-use that needs a human every step is labor with extra latency. Who owns silent failure on day 3?

Good, now when do we try fine tuning on these traces?

A landing page for recorded memory history containing one (1) circle. Bold. 🙄

This is awesome! I can't wait to try it out. I have tokens to burn with Qwen 3.8, we'll see how it does there

demo to try?

Perfectly timed. Incredible