From AI Assistant to AI Scientist: Autonomous Discovery of LLM-RL Algorithms with LLM Agents

arXiv 2026 62.8 method

TLDR

POISE is a closed-loop framework using LLM agents to autonomously discover improved policy optimization algorithms for language models.

Reasoning

The paper presents a novel automated discovery framework (POISE) with strong empirical results on mathematical reasoning tasks, demonstrating significant improvements over GRPO. However, the abstract lacks details on the generalizability to other domains and the computational cost of the iterative process.

Read-first score

Read-first score 62.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 72.

Recency 8%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,evaluation,experiment,validation

Topical relevance 42%

Uses existing LLM keyword relevance scores normalized to 0-100. AI scientist,automated scientific discovery,autonomous research agent,automated research,literature review agent,survey generation,automated experimentation,experiment design agent,AI for scientific research,paper writing agent,research automation,scientific discovery agent

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 25.

Keyword Scores

automated scientific discovery

automated experimentation

AI scientist

automated research

AI for scientific research

research automation

scientific discovery agent

autonomous research agent

experiment design agent

literature review agent

survey generation

paper writing agent

Deep Analysis

Innovations

POISE: a closed-loop framework for autonomous discovery of policy optimization algorithms for language models
Structured, genealogically linked archive linking proposals, implementations, evaluations, and reflections to support evidence-driven iteration
Discovery of analytic-variance scaling and validity masking mechanisms that improve over GRPO

Methodology

POISE iteratively generates, implements, evaluates, and reflects on candidate policy optimization algorithms, using a structured archive to guide evidence-driven search over algorithmic mechanisms coupled with training dynamics. Starting from GRPO, it evaluates 64 candidate algorithms on mathematical reasoning tasks.

Key Results

The best discovered variant improves weighted Overall from 47.8 to 52.5 (+4.6) and increases AIME25 pass@32 from 26.7% to 43.3% over the GRPO baseline.

Resources

Paper