SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2602.17155 · 2026

Powering Up Zeroth-Order Training via Subspace Gradient Orthogonalization

Sijia Liu, Wotao Yin, Mingyi Hong, Zheng Zhang, Yihua Zhang, Changsheng Wang, Yicheng Lang

arXiv · PDF · Open in the Atlas

Code that ran

8 of 15 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 5 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 3 failures of our instrument; 7 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2602.17155.json · badge: /api/ran/2602.17155.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
OPTML-Group/ZO-Muon canonical (classifier label)
classifier basis: census
8 ran · 7 no recorded run
FunctionStatusWhere it lives
f1 Ran
ran, honoured the contract we drafted
OPTML-Group/ZO-Muon/llm/metrics.py
code served (permissive licence) · get_code("b989d4bce26f77ce")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float64, ndim 0, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
import re
import string
from collections import Counter

# Define ARGS for the f1 function
ARGS = {
    "pred": "The answer is Paris",
    "gold": ["Paris", "the city of Paris"]
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float64, ndim 0, dtype family float — no output shape is stated: observe_signature.py recorded float ndim 0, which disagrees with the fixture run
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
fast_svd_method_v2 Ran
ran; our contract draft was wrong, not the code
OPTML-Group/ZO-Muon/llm/SubZeroTrainer.py
code served (permissive licence) · get_code("bd2c31732da89465")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

ARGS = {
    "w_shape": (4, 3),
    "device": "cpu",
    "dtype": torch.float32,
    "rank": 2,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
find_module Ran
ran; our contract draft was wrong, not the code
OPTML-Group/ZO-Muon/llm/lora.py
pointer only (licence: MIT); the stored copy get_code serves, from JingXuTHU/Random-Masking-Finds-Winning-Tickets-for-Parameter-Efficient-Fine-tuning, is pointer only (licence: NONE) · get_code("b5adfd6ac99a1b18")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from JingXuTHU/Random-Masking-Finds-Winning-Tickets-for-Parameter-Efficient-Fine-tuning, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_constant_schedule_with_warmup Ran
ran, honoured the contract we drafted
OPTML-Group/ZO-Muon/llm/Hessian_smooth_scheduler.py
code served (permissive licence) · get_code("4a22da68f93d77ed")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

ARGS = {
    "learning_rate": 0.001,
    "name": "constant_schedule_with_warmup",
    "num_warmup_steps": 1000,
    "num_decay_steps": 5000,
    "current_step": 500,
    "num_training_steps": 10000,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_inverse_sqrt_schedule Ran
ran, honoured the contract we drafted
OPTML-Group/ZO-Muon/llm/Hessian_smooth_scheduler.py
code served (permissive licence) · get_code("efcead5088fac5fa")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

ARGS = {
    "learning_rate": 0.001,
    "name": "inverse_sqrt_schedule",
    "num_warmup_steps": 1000,
    "num_decay_steps": 5000,
    "current_step": 500,
    "num_training_steps": 10000
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_reduce_on_plateau_schedule Ran
ran, honoured the contract we drafted
OPTML-Group/ZO-Muon/llm/Hessian_smooth_scheduler.py
code served (permissive licence) · get_code("c04f796282ba4842")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

ARGS = {
    "learning_rate": 0.001,
    "name": "reduce_on_plateau",
    "num_warmup_steps": 100,
    "num_decay_steps": 500,
    "current_step": 50,
    "num_training_steps": 1000,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
normalize_answer Ran
ran; our contract draft was wrong, not the code
OPTML-Group/ZO-Muon/llm/metrics.py
pointer only (licence: MIT); the stored copy get_code serves, from borealisai/radar-decoding, is pointer only (licence: LGPL-3.0) · get_code("e7e75981cb464788")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from borealisai/radar-decoding, and LGPL-3.0, as GitHub detects it on that repository, is not in the permissive set Syntology serves inline. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

zeropower_via_svd Ran
ran, honoured the contract we drafted
OPTML-Group/ZO-Muon/llm/LowDimTrainer.py
pointer only (licence: MIT); the stored copy get_code serves, from MOFA-LAB/ZO-MOPI, is pointer only (licence: NONE) · get_code("26e94eb74c93a705")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from MOFA-LAB/ZO-MOPI, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

attn_forward_hook Not yet run
unverified (harvested, no recorded run)
OPTML-Group/ZO-Muon/llm/prefix.py
code served (permissive licence) · get_code("bdb460cb247501dc")
How it ran

no invocation recorded.

calculate_metric Not yet run
unverified (harvested, no recorded run)
OPTML-Group/ZO-Muon/llm/metrics.py
code served (permissive licence) · get_code("84c658f2c9252878")
How it ran

no invocation recorded.

get_constant_schedule Not yet run
unverified (harvested, no recorded run)
OPTML-Group/ZO-Muon/llm/lr_scheduler.py
code served (permissive licence) · get_code("66ebbc67206f9124")
How it ran

no invocation recorded.

get_task Not yet run
unverified (harvested, no recorded run)
OPTML-Group/ZO-Muon/llm/tasks.py
code served (permissive licence) · get_code("f7c384f1e681ea7a")
How it ran

no invocation recorded.

prepare_inputs_for_generation Not yet run
unverified (harvested, no recorded run)
OPTML-Group/ZO-Muon/llm/prefix.py
code served (permissive licence) · get_code("84920a8095dca4fa")
How it ran

no invocation recorded.

zeropower_via_newtonschulz5 Not yet run
unverified (harvested, no recorded run)
OPTML-Group/ZO-Muon/llm/LowDimTrainer.py
code served (permissive licence) · get_code("635c93a80a199840")
How it ran

no invocation recorded.

zeropower_via_newtonschulz5_float Not yet run
unverified (harvested, no recorded run)
OPTML-Group/ZO-Muon/llm/SubZeroTrainer.py
code served (permissive licence) · get_code("d3a6a969126c5294")
How it ran

no invocation recorded.

Also present in other papers' code

Of 15 samples lifted from this paper's linked repositories, 15 carry a function identity and 0 do not. Of the 15 distinct functions with an identity, 9 are also present in other papers whose code we hold (8 with at least one paper dated earlier, 1 the same month, 2 later) and 6 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
normalize_answer 133 other papers
120 dated earlier, 4 the same month, 8 later, 1 undated · 1 only as an assembly brick
1508.06615 (2015-08)
Character-Aware Neural Language Models
1606.05250 (2016-06)
SQuAD: 100,000+ Questions for Machine Comprehension of Text
1608.07905 (2016-08)
Machine Comprehension Using Match-LSTM and Answer Pointer
and 130 more not named here
find_module 8 other papers
8 dated earlier
2305.17333 (2023-05)
Fine-Tuning Language Models with Just Forward Passes
2310.09639 (2023-10)
DPZero: Private Fine-Tuning of Language Models without Backpropagation
2312.15184 (2023-12)
ZO-AdaMU Optimizer: Adapting Perturbation by the Momentum and Uncertainty in Zeroth-order Optimization
and 5 more not named here
calculate_metric 7 other papers
7 dated earlier
2305.17333 (2023-05)
Fine-Tuning Language Models with Just Forward Passes
2310.09639 (2023-10)
DPZero: Private Fine-Tuning of Language Models without Backpropagation
2312.15184 (2023-12)
ZO-AdaMU Optimizer: Adapting Perturbation by the Momentum and Uncertainty in Zeroth-order Optimization
and 4 more not named here
f1 7 other papers
7 dated earlier
2305.17333 (2023-05)
Fine-Tuning Language Models with Just Forward Passes
2310.09639 (2023-10)
DPZero: Private Fine-Tuning of Language Models without Backpropagation
2312.15184 (2023-12)
ZO-AdaMU Optimizer: Adapting Perturbation by the Momentum and Uncertainty in Zeroth-order Optimization
and 4 more not named here
get_task 7 other papers
7 dated earlier
2305.17333 (2023-05)
Fine-Tuning Language Models with Just Forward Passes
2310.09639 (2023-10)
DPZero: Private Fine-Tuning of Language Models without Backpropagation
2312.15184 (2023-12)
ZO-AdaMU Optimizer: Adapting Perturbation by the Momentum and Uncertainty in Zeroth-order Optimization
and 4 more not named here
attn_forward_hook 6 other papers
6 dated earlier
2305.17333 (2023-05)
Fine-Tuning Language Models with Just Forward Passes
2310.09639 (2023-10)
DPZero: Private Fine-Tuning of Language Models without Backpropagation
2312.15184 (2023-12)
ZO-AdaMU Optimizer: Adapting Perturbation by the Momentum and Uncertainty in Zeroth-order Optimization
and 3 more not named here
prepare_inputs_for_generation 6 other papers
6 dated earlier
2305.17333 (2023-05)
Fine-Tuning Language Models with Just Forward Passes
2310.09639 (2023-10)
DPZero: Private Fine-Tuning of Language Models without Backpropagation
2312.15184 (2023-12)
ZO-AdaMU Optimizer: Adapting Perturbation by the Momentum and Uncertainty in Zeroth-order Optimization
and 3 more not named here
fast_svd_method_v2 1 other paper
1 dated earlier
2410.08989 (2024-10)
Zeroth-Order Fine-Tuning of LLMs in Random Subspaces
zeropower_via_svd 1 other paper
1 later
2605.09034 (2026-05)
Accelerating Zeroth-Order Spectral Optimization with Partial Orthogonalization from Power Iteration

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

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

Abstract

Zeroth-order (ZO) optimization provides a gradient-free alternative to first-order (FO) methods by estimating gradients via finite differences of function evaluations, and has recently emerged as a memory-efficient paradigm for fine-tuning large-scale models by avoiding backpropagation. However, ZO optimization has a fundamental tension between accuracy and query efficiency. In this work, we show that ZO optimization can be substantially improved by unifying two complementary principles: (i) a projection-based subspace view that reduces gradient estimation variance by exploiting the intrinsic low-rank structure of model updates, and (ii) Muon-style spectral optimization that applies gradient orthogonalization to extract informative spectral structure from noisy ZO gradients. These findings form a unified framework of subspace gradient orthogonalization, which we instantiate in a new method, ZO-Muon, admitting a natural interpretation as a low-rank Muon optimizer in the ZO setting. Extensive experiments on large language models (LLMs) and vision transformers (ViTs) demonstrate that ZO-Muon significantly accelerates convergence and achieves a win-win improvement in accuracy and query/runtime efficiency. Notably, compared to the popular MeZO baseline, ZO-Muon requires only 24.7% of the queries to reach the same SST-2 performance for LLM fine-tuning, and improves accuracy by 25.1% on ViT-B fine-tuning on CIFAR-100. Code is available at https://github.com/OPTML-Group/ZO-Muon.

For agents

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

get_harvested_code_for_paper("2602.17155")
get_code_for_paper("2602.17155")
have("2602.17155")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2602.17155.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2602.17155.svg)](https://syntology.ai/paper/2602.17155)

Connect an agent — have() is free.