SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2602.00722 · ICML · 2026

Spectral Imbalance Causes Forgetting in Low-Rank Continual Adaptation

Min-Ling Zhang, Tong Wei, Zi-Hao Zhou, Mao-Lin Luo, Hao Gu, Han-Chen Zhang, Lora-Ft O-Lora Cl-Moe, Sefe Keeplora

arXiv · PDF · Open in the Atlas

Code that ran

7 of 8 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (1 of them constructed an object rather than computing a result; 2 honoured the contract we drafted, 4 violated it and 1 has no contract verdict; 0 failures of our instrument; 1 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/2602.00722.json · badge: /api/ran/2602.00722.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
haodotgu/EBLoRA community
Papers with Code archive
7 ran · 1 no recorded run
FunctionStatusWhere it lives
RiemannianAdamW Ran
ran on a synthesized input · constructed an object; no result was computed
haodotgu/EBLoRA/llava/train/riemannian_adam.py
code served (permissive licence) · get_code("f2240d43dba2b3d0")
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 an instance of its own class, RiemannianAdamW: the run constructed the object and did not call it, so no result was computed. 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
from torch.optim import AdamW

torch.manual_seed(0)

# Create a simple model with parameters
model = nn.Sequential(
    nn.Linear(4, 3),
    nn.ReLU(),
    nn.Linear(3, 2)
)

# Get parameters from the model
params = list(model.parameters())

# Create parameter groups with lora_tag attributes
param_groups = [
    {
        'params': [params[0]],
        'lr': 0.001,
        'lora_tag': 'U'
    },
    {
        'params': [params[1]],
        'lr': 0.001,
        'lora_tag': 'Vt'
    },
    {
        'params': [params[2]],
        'lr': 0.001,
        'lora_tag': 'other'
    }
]

# Create gradient_prev_task dictionary
# The keys should be torch.nn.Parameter objects
# The values should be tensors representing F matrices
gradient_prev_task = {
    params[0]: torch.randn(4, 8),  # F matrix for first param
}

# Create grad_scaling function (optional)
def grad_scaling_fn(grad, param):
    return grad * 0.1

ARGS = {
    'params': param_groups,
    'lr': 0.001,
    'betas': (0.9, 0.999),
    'eps': 1e-08,
    'weight_decay': 0.0,
    'gradient_prev_task': gradient_prev_task,
    'grad_scaling': grad_scaling_fn
}
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 RiemannianAdamW — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a RiemannianAdamW output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not 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
euclidean2riemannian_U Ran
ran, violated the contract we drafted
haodotgu/EBLoRA/llava/train/riemannian_adam.py
code served (permissive licence) · get_code("68db4636d010c513")
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 Tensor, ndim 3, dtype family float, shape (2, 4, 3) 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)

# The function takes two parameters: x and grad
# Both are Tensors used in matrix operations
# x @ (grad.transpose(-2, -1) @ x) suggests:
# - x should be a 2D or higher tensor (matrix-like)
# - grad should be a 2D or higher tensor (matrix-like)
# - Operations use @ (matrix multiplication) and transpose(-2, -1)

# Create x: shape (2, 4, 3) - batch of 2, 4x3 matrices
x = torch.randn(2, 4, 3)

# Create grad: shape (2, 4, 3) - same shape as x for compatibility
# grad.transpose(-2, -1) will give (2, 3, 4)
# (grad.transpose(-2, -1) @ x) will give (2, 3, 3)
# x @ (2, 3, 3) will give (2, 4, 3)
grad = torch.randn(2, 4, 3)

ARGS = {
    "x": x,
    "grad": grad,
}
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 Tensor, ndim 3, dtype family float, shape (2, 4, 3) — shape (2, 4, 3) 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
euclidean2riemannian_Vt Ran
ran, violated the contract we drafted
haodotgu/EBLoRA/llava/train/riemannian_adam.py
code served (permissive licence) · get_code("55350aadc942d740")
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 Tensor, ndim 3, dtype family float, shape (2, 4, 3) 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)

# Create small tensors for the function parameters
# x: Tensor - used in matrix operations @ and transpose
# grad: Tensor - used in matrix operations @ and transpose
# F: Optional[Tensor] - used in matrix operations @ and transpose, can be None

# These tensors need compatible shapes for matrix multiplication
# Shape: (2, 4, 3) - batch size 2, with 4x3 matrices
x = torch.randn(2, 4, 3)
grad = torch.randn(2, 4, 3)
F = torch.randn(2, 4, 4)  # Different size to distinguish from x and grad

ARGS = {
    "x": x,
    "grad": grad,
    "F": F
}
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 Tensor, ndim 3, dtype family float, shape (2, 4, 3) — shape (2, 4, 3) 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
polar_retraction_U Ran
ran, honoured the contract we drafted
haodotgu/EBLoRA/llava/train/riemannian_adam.py
code served (permissive licence) · get_code("2a0933175e363abe")
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 Tensor, ndim 2, dtype family float, shape (4, 3) 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)

# Create small tensors for the function parameters
X = torch.randn(4, 3)
grad = torch.randn(4, 3)

ARGS = {
    "X": X,
    "grad": grad
}
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 Tensor, ndim 2, dtype family float, shape (4, 3) — shape (4, 3) 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
polar_retraction_Vt Ran
ran, violated the contract we drafted
haodotgu/EBLoRA/llava/train/riemannian_adam.py
code served (permissive licence) · get_code("0921dc421f2d3073")
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 Tensor, ndim 3, dtype family float, shape (2, 4, 3) 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)

# Construct arguments for polar_retraction_Vt
# The function signature is:
# polar_retraction_Vt(X: Tensor, grad: Tensor, F: Optional[Tensor]) -> Tensor

# X: a tensor representing a point on the manifold
# grad: a tangent vector (gradient)
# F: an optional tensor used in projection, or None

# Looking at the function body:
# - X and grad are added together
# - new_V is projected: new_V - F @ (F.transpose(-2, -1) @ new_V)
# - This means F should have compatible dimensions for matrix multiplication
# - F.transpose(-2, -1) @ new_V means F is (..., m, n) and new_V is (..., n, k)
# - Then F @ result means F is (..., m, n) and result is (..., n, k)

# Let's use small tensors with distinct dimensions
# X: (2, 4, 3) - batch of 2, with 4x3 matrices
# grad: (2, 4, 3) - same shape as X
# F: (2, 4, 2) - batch of 2, with 4x2 matrices (for projection)

X = torch.randn(2, 4, 3)
grad = torch.randn(2, 4, 3)
F = torch.randn(2, 4, 2)

ARGS = {
    "X": X,
    "grad": grad,
    "F": F,
}
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 Tensor, ndim 3, dtype family float, shape (2, 4, 3) — shape (2, 4, 3) 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
polar_uf Ran
ran, honoured the contract we drafted
haodotgu/EBLoRA/llava/train/riemannian_adam.py
code served (permissive licence) · get_code("68db4da3e9514261")
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 Tensor, ndim 2, dtype family float, shape (4, 3) 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
from torch import Tensor

torch.manual_seed(0)

# Create a small matrix for SVD decomposition
# Using distinct dimensions: 4 rows, 3 columns
m = torch.randn(4, 3)

ARGS = {
    "m": m
}
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 Tensor, ndim 2, dtype family float, shape (4, 3) — shape (4, 3) 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
tangent_project_U Ran
ran, violated the contract we drafted
haodotgu/EBLoRA/llava/train/riemannian_adam.py
code served (permissive licence) · get_code("88022d42405baaea")
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 Tensor, ndim 3, dtype family float, shape (2, 4, 3) 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)

# Analyze the function:
# - x: Tensor, used in matrix multiplication and transpose operations
# - grad: Tensor, used in matrix multiplication and transpose operations
# The operation: grad - x @ (grad.transpose(-2, -1) @ x)
# This suggests x and grad should be compatible for these operations
# For batched matrix operations, shapes should be (batch, n, m) or similar

# Create small tensors with distinct dimensions
batch_size = 2
n_dim = 4
m_dim = 3

# x should be shape (batch, n_dim, m_dim) 
# grad should be shape (batch, n_dim, m_dim)
# So that grad.transpose(-2, -1) gives (batch, m_dim, n_dim)
# And grad.transpose(-2, -1) @ x gives (batch, m_dim, m_dim)
# And x @ (...) gives (batch, n_dim, m_dim)
# And grad - (...) is valid

x = torch.randn(batch_size, n_dim, m_dim)
grad = torch.randn(batch_size, n_dim, m_dim)

ARGS = {
    "x": x,
    "grad": grad
}
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 Tensor, ndim 3, dtype family float, shape (2, 4, 3) — shape (2, 4, 3) 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
tangent_project_Vt Not yet run
unverified (harvested, no recorded run)
haodotgu/EBLoRA/llava/train/riemannian_adam.py
code served (permissive licence) · get_code("5b4bb8ce66457224")
How it ran

no invocation recorded.

Also present in other papers' code

Of 8 samples lifted from this paper's linked repositories, 8 carry a function identity and 0 do not. Of the 8 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 8 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.

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

Parameter-efficient continual learning aims to adapt pre-trained models to sequential tasks without forgetting previously acquired knowledge. Most existing approaches treat continual learning as avoiding interference with past updates, rather than considering what properties make the current task-specific update naturally preserve previously acquired knowledge. From a knowledgedecomposition perspective, we observe that lowrank adaptations exhibit highly imbalanced singular value spectra: a few dominant components absorb most of the adaptation energy, thereby (i) more likely to disrupt previously acquired knowledge and (ii) making the update more vulnerable to interference from subsequent tasks. To enable explicit balance among components, we decouple the magnitude of the task update from its directional structure and formulate it as a constrained optimization problem on a restricted Stiefel manifold. We address this problem using a projected first-order method compatible with standard deep-learning optimizers used in vision-language models. Our method mitigates both backward and forward forgetting, consistently outperforming continual learning baselines. The implementation code is available at https: //github.com/haodotgu/EBLoRA.

For agents

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

get_harvested_code_for_paper("2602.00722")
get_code_for_paper("2602.00722")
have("2602.00722")

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

curl https://syntology.ai/api/ran/2602.00722.json

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

[![Syntology run record](https://syntology.ai/api/ran/2602.00722.svg)](https://syntology.ai/paper/2602.00722)

Connect an agent — have() is free.