Self Evolving Dual AI Agent System = AutoResearch 2.0 (LSE)

Description:
The most elegant thing about LSE (Learning to Self-Evolve) is how it tackles the "credit assignment" nightmare of self-correction by essentially collapsing a messy, long-horizon trajectory into a single-step RL objective.

Instead of just hoping for "emergent" reasoning to fix a bad prompt, they’ve explicitly trained a tiny 4B-parameter "meta-policy" model to act as a specialized optimizer for the action model.

By using the performance delta as the reward, they’ve basically baked a control variate directly into the objective. This is a total "cheat code" because it kills the noise from varying prompt difficulty and forces the model to learn the marginal utility of its edits rather than just preserving already-good instructions. However, this induces limitations, see my last video on "VISTA".

Then, at test time, they wrap this RL-honed instinct in a UCB tree search so the system can actually backtrack when it hits a "hallucination wall."

It’s basically teaching an AI the scientific method: it observes its own failures, deduces the structural invariants (!) of the domain (like the weird topological quirks of a SQL schema) and rewrites its own operating manual until it out-navigates a frozen GPT-5 or a CLAUDE Sonnet 4.5, which is impressive.

It proves that "task execution" and "meta-optimization" are two completely different skills in latent space, and the latter is a learnable, transferable feature of intelligence.

All rights w/ authors:
LEARNING TO SELF-EVOLVE
Xiaoyin Chen 1, 2∗ Canwen Xu 3 Yite Wang 3 Boyi Liu 3 Zhewei Yao 3 Yuxiong He 3
from
1 Mila – Quebec AI Institute
2 University of Montreal
3 Snowflake

@Mila-Quebec-AI-Institute @umontreal @SnowflakeInc