SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2605.18591 · ICML · 2026

Randomized Advantage Transformation (RAT): Computing Natural Policy Gradients via Direct Backpropagation

Mingfei Sun

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 7 functions out of this paper's own repositories and ran 5 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.

RepositoryRoleRan
agent-lab/ICML2026-RAT canonical 5 of 7
FunctionStatusWhere it lives
conjugate_gradient Ran agent-lab/ICML2026-RAT/utils/cg.py
code served (permissive licence) · get_code("30a3ae0cf5fb6225")
get_monitor_files Ran agent-lab/ICML2026-RAT/utils/monitor.py
code served (permissive licence) · get_code("6dabfbf42f09101c")
load_results Ran agent-lab/ICML2026-RAT/utils/monitor.py
code served (permissive licence) · get_code("3c4981304881dcff")
make_output_format Ran agent-lab/ICML2026-RAT/utils/logger.py
code served (permissive licence) · get_code("bcd8b4acab199405")
try_contiguous Ran agent-lab/ICML2026-RAT/kfac/kfac_utils.py
code served (permissive licence) · get_code("a485499e9e91edac")
profile Not yet run agent-lab/ICML2026-RAT/utils/logger.py
code served (permissive licence) · get_code("a416e2085fe90df5")
read_json Not yet run agent-lab/ICML2026-RAT/utils/logger.py
code served (permissive licence) · get_code("88ef5a17fc986d9a")

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Natural policy gradients improve optimization by accounting for the geometry of distribution space, but their practical use is limited by the cost of estimating and inverting the Fisher matrix. We present Randomized Advantage Transformation (RAT), a method for estimating Tikhonovregularized natural policy gradients via direct backpropagation. By applying the Woodbury formula, we reformulate the regularized natural policy gradients as vanilla policy gradients with a transformed advantage. RAT computes this transformation efficiently via randomized block Kaczmarz iterations on on-policy mini-batches, avoiding explicit Fisher construction, conjugategradient solvers, and architecture-specific approximations. We provide convergence guarantees for RAT and demonstrate empirically that it matches or exceeds established natural-gradient methods across continuous and visual control benchmarks, while remaining simple to implement and compatible with various architectures.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2605.18591")
get_code_for_paper("2605.18591")
have("2605.18591")

Connect an agent — have() is free.