SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2602.01212 · ICML · 2026

SimpleGPT: Improving GPT via A Simple Normalization Strategy

Xianbiao Qi, Rong Xiao, Yelin He, Jiaquan Ye, Marco Chen

arXiv · PDF · Open in the Atlas

Code that ran

5 of 7 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; 1 honoured the contract we drafted, 1 violated it and 0 have no contract verdict; 3 failures of our instrument; 2 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/2602.01212.json · badge: /api/ran/2602.01212.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
Ocram7/SimpleGPT canonical (classifier label)
classifier basis: census
5 ran · 2 no recorded run
FunctionStatusWhere it lives
apply_rotary_emb Ran
ran; our contract draft was wrong, not the code
Ocram7/SimpleGPT/src/torchtitan/models/llama/simplegpt_model.py
code served (permissive licence) · get_code("3f4b3af840a38f53")
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, ndim 5, 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 torch

torch.manual_seed(0)

# Analyzing the function:
# - apply_rotary_emb takes xq, xk, freqs_cis (all torch.Tensor)
# - xq and xk are reshaped to (*xq.shape[:-1], -1, 2) for complex view
# - freqs_cis is sliced by seqlen (x.shape[1]) and must have shape (seqlen, x.shape[-1])
# - reshape_for_broadcast expects ndim >= 2 and uses x.shape[1] as seqlen

# Let's construct valid tensors:
# xq shape: (batch=2, seqlen=3, dim=4)
# xk shape: (batch=2, seqlen=3, dim=4)
# freqs_cis shape: (max_seqlen=5, dim=2) where dim=2 because after reshape xq becomes (..., 2, 2) so last dim is 2

batch_size = 2
seqlen = 3
head_dim = 4

# xq and xk: shape (batch, seqlen, head_dim)
xq = torch.randn(batch_size, seqlen, head_dim)
xk = torch.randn(batch_size, seqlen, head_dim)

# freqs_cis: shape (max_seqlen, head_dim // 2)
# After reshape in apply_rotary_emb, xq becomes (batch, seqlen, head_dim//2, 2)
# viewed as complex: (batch, seqlen, head_dim//2)
# So freqs_cis needs shape (seqlen, head_dim//2)
freqs_cis = torch.randn(5, head_dim // 2, dtype=torch.complex64)

ARGS = {
    "xq": xq,
    "xk": xk,
    "freqs_cis": freqs_cis,
}
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, ndim 5, dtype family float — a tuple output has no array shape of its own; none is stated
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
linear_warmup_linear_decay Ran
ran, honoured the contract we drafted
Ocram7/SimpleGPT/src/torchtitan/lr_scheduling.py
code served (permissive licence) · get_code("9a8cdcab05d572c1")
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)

# The function linear_warmup_linear_decay takes a single parameter: current_step
# It uses this integer to compute a learning rate multiplicative factor
# No complex objects needed - just an integer step count

ARGS = {
    "current_step": 50
}
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
precompute_freqs_cis Ran
ran, violated the contract we drafted
Ocram7/SimpleGPT/src/torchtitan/models/llama/simplegpt_model.py
code served (permissive licence) · get_code("6d2474603ba25d5b")
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 (2, 2) 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 = {
    "dim": 4,
    "end": 2,
    "theta": 10000.0,
}
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 (2, 2) — shape (2, 2) 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
reshape_for_broadcast Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
Ocram7/SimpleGPT/src/torchtitan/models/llama/simplegpt_model.py
code served (permissive licence) · get_code("bba3feaecfc713d6")
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 (1, 3, 4) 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 freqs_cis tensor
# The function expects freqs_cis to have shape (seqlen, x.shape[-1])
# Let's use seqlen=3 and x.shape[-1]=4
seqlen = 3
head_dim = 4
freqs_cis = torch.randn(seqlen, head_dim)

# Construct x tensor
# x needs ndim >= 2 (due to assert 0 <= 1 < ndim)
# x.shape[1] should match seqlen for the slicing
# x.shape[-1] should match freqs_cis.shape[-1]
# Let's make x have shape (batch=2, seq=3, head_dim=4) for 3D tensor
# This satisfies: ndim=3, x.shape[1]=seqlen, x.shape[-1]=head_dim
x = torch.randn(2, seqlen, head_dim)

ARGS = {
    "freqs_cis": freqs_cis,
    "x": x,
}
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 (1, 3, 4) — shape (1, 3, 4) 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
string_list Ran
ran; our contract draft was wrong, not the code
Ocram7/SimpleGPT/src/torchtitan/config_manager.py
code served (permissive licence) · get_code("8524ea85f104747e")
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 list, ndim 1, 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 torch

ARGS = {
    "raw_arg": "hello,world,test"
}
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 list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
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
dist_max Not yet run
unverified (harvested, no recorded run)
Ocram7/SimpleGPT/src/torchtitan/utils.py
code served (permissive licence) · get_code("6f1cb76a7fc611d6")
How it ran

no invocation recorded.

dist_mean Not yet run
unverified (harvested, no recorded run)
Ocram7/SimpleGPT/src/torchtitan/utils.py
code served (permissive licence) · get_code("a4431453cab716c0")
How it ran

no invocation recorded.

Also present in other papers' code

Of 7 samples lifted from this paper's linked repositories, 7 carry a function identity and 0 do not. Of the 7 distinct functions with an identity, 7 are also present in other papers whose code we hold (3 with at least one paper dated earlier, 0 the same month, 5 later) and 0 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
apply_rotary_emb 55 other papers
51 dated earlier, 4 later · 1 only as an assembly brick
2023.emnlp-main.534 (2023 (year only))
Disentangling Transformer Language Models as Superposed Topic Models
2303.09736 (2023-03)
Dynamic Structure Pruning for Compressing CNNs
2307.09288 (2023-07)
Llama 2: Open Foundation and Fine-Tuned Chat Models
and 52 more not named here
precompute_freqs_cis 2 other papers
2 dated earlier
2406.16793 (2024-06)
Adam-mini: Use Fewer Learning Rates To Gain More
2410.06511 (2024-10)
TorchTitan: One-stop PyTorch native solution for production ready LLM pre-training
dist_max 1 other paper
1 later
2606.06470 (2026-06)
PC Layer: Polynomial Weight Preconditioning for Improving LLM Pre-Training
dist_mean 1 other paper
1 later
2606.06470 (2026-06)
PC Layer: Polynomial Weight Preconditioning for Improving LLM Pre-Training
linear_warmup_linear_decay 1 other paper
1 later
2606.06470 (2026-06)
PC Layer: Polynomial Weight Preconditioning for Improving LLM Pre-Training
reshape_for_broadcast 1 other paper
1 dated earlier
2406.16793 (2024-06)
Adam-mini: Use Fewer Learning Rates To Gain More
string_list 1 other paper
1 later
2606.06470 (2026-06)
PC Layer: Polynomial Weight Preconditioning for Improving LLM Pre-Training

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

In this work, we revisit Transformer optimization through the lens of second-order geometry and establish a direct connection between architectural design, activation scale, the Hessian matrix, and the maximum tolerable learning rate. We introduce a simple normalization strategy, termed SimpleNorm, which stabilizes intermediate activation scales by construction. Then, by analyzing the Hessian of the loss with respect to network activations, we theoretically show that SimpleNorm significantly reduces the spectral norm of the Hessian, thereby permitting larger stable learning rates. We validate our theoretical findings through extensive experiments on large GPT models at parameter scales 1B, 1.4B, 7B and 8B. Empirically, SimpleGPT, our SimpleNorm-based network, tolerates learning rates 3×-10× larger than standard convention, consistently demonstrates strong optimization stability, and achieves substantially better performance than well-established baselines. Specifically, when training 7B-scale models for 60K steps, SimpleGPT achieves a training loss that is 0.08 lower than that of LLaMA2 with QKNorm, reducing the loss from 2.290 to 2.208. Our source code will be released at https: //github.com/Ocram7/SimpleGPT.

For agents

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

get_harvested_code_for_paper("2602.01212")
get_code_for_paper("2602.01212")
have("2602.01212")

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

curl https://syntology.ai/api/ran/2602.01212.json

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

[![Syntology run record](https://syntology.ai/api/ran/2602.01212.svg)](https://syntology.ai/paper/2602.01212)

Connect an agent — have() is free.