SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2607.03451 · 2026

SkillOpt-Lite: Better and Faster Agent Self-evolution via One Line of Vibe

Bo Li, Yifei Shen, Xinjie Zhang

arXiv · PDF · Open in the Atlas

Code that ran

8 of 9 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; 0 honoured the contract we drafted, 2 violated it and 0 have no contract verdict; 6 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/2607.03451.json · badge: /api/ran/2607.03451.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
EvolvingLMMs-Lab/SkillOpt-Lite canonical (classifier label)
classifier basis: census
8 ran · 1 no recorded run
FunctionStatusWhere it lives
default_model_for_backend Ran
ran; our contract draft was wrong, not the code
EvolvingLMMs-Lab/SkillOpt-Lite/skillopt/model/common.py
code served (permissive licence) · get_code("d4e412018e3e2c89")
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 str. 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.

# Setup code for default_model_for_backend function

ARGS = {
    'backend': 'azure_openai'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str 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
format_codex_trace_steps Ran
ran; our contract draft was wrong, not the code
EvolvingLMMs-Lab/SkillOpt-Lite/skillopt/model/codex_harness.py
code served (permissive licence) · get_code("da869f4bb7938850")
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 str. 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

# Construct a valid raw string containing Codex CLI output with proper markers
raw_input = """user
This is a user prompt asking to solve something
codex
def solve():
    return 42
exec
python solve.py
succeeded in 0.5s
user
Another follow-up question
codex
def improved():
    return 43
exec
python improved.py
succeeded in 0.3s
"""

ARGS = {
    'raw': raw_input,
    'max_chars': 4000,
}
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 str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str 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
is_structured Ran
ran, violated the contract we drafted
EvolvingLMMs-Lab/SkillOpt-Lite/skillopt/config.py
code served (permissive licence) · get_code("2a69caa819e48fbc")
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 bool, 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)

# Test case 1: cfg with structured section 'model'
cfg_with_model = {
    'model': {
        'type': 'transformer',
        'layers': 2
    },
    'other_key': 'value'
}

# Test case 2: cfg with structured section 'train'
cfg_with_train = {
    'train': {
        'batch_size': 32,
        'epochs': 10
    }
}

# Test case 3: cfg with structured section 'optimizer'
cfg_with_optimizer = {
    'optimizer': {
        'lr': 0.001,
        'weight_decay': 0.0001
    },
    'some_other': 'data'
}

# Test case 4: cfg with no structured sections
cfg_without_structured = {
    'learning_rate': 0.001,
    'batch_size': 32,
    'random_key': 'random_value'
}

# Test case 5: cfg with structured section but value is not a dict (should return False)
cfg_with_non_dict_structured = {
    'model': 'some_string_value',
    'train': 123
}

# The function only needs cfg parameter
ARGS = {
    'cfg': cfg_with_model
}
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 bool, 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
load_config Ran
ran; our contract draft was wrong, not the code
EvolvingLMMs-Lab/SkillOpt-Lite/skillopt/config.py
code served (permissive licence) · get_code("3a44fa3f8d7e6ead")
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 dict. 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 tempfile
import os
import yaml

# Create temporary directory for test YAML files
tmpdir = tempfile.mkdtemp()

# Create a base config file
base_config = {
    'model': {
        'type': 'resnet',
        'depth': 50
    },
    'training': {
        'lr': 0.001,
        'epochs': 100
    }
}
base_path = os.path.join(tmpdir, 'base.yaml')
with open(base_path, 'w') as f:
    yaml.dump(base_config, f)

# Create a main config file that inherits from base
main_config = {
    '_base_': 'base.yaml',
    'training': {
        'epochs': 200
    }
}
main_path = os.path.join(tmpdir, 'config.yaml')
with open(main_path, 'w') as f:
    yaml.dump(main_config, f)

# Define ARGS for load_config
ARGS = {
    'path': main_path,
    'overrides': ['model.depth=101', 'training.lr=0.01', 'debug=true']
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
yaml — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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
needs_responses_api Ran
ran, violated the contract we drafted
EvolvingLMMs-Lab/SkillOpt-Lite/skillopt/model/common.py
code served (permissive licence) · get_code("498cb843c85d2f97")
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 bool, 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

# Analyze the function: needs_responses_api(model: str) -> bool
# The function:
# 1. Takes a model parameter (string)
# 2. Normalizes it to lowercase and strips whitespace
# 3. Checks if it matches or starts with prefixes in _RESPONSES_API_MODELS
# 4. Returns a boolean

# Create test arguments - the function only takes one parameter: model (str)
ARGS = {
    'model': 'gpt-5-codex'
}
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 bool, 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_backend_name Ran
ran; our contract draft was wrong, not the code
EvolvingLMMs-Lab/SkillOpt-Lite/skillopt/model/common.py
code served (permissive licence) · get_code("e9012e57939b6cec")
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 str. 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.

# Setup code for normalize_backend_name function

ARGS = {
    "name": "azure"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str 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
parse_codex_raw Ran
ran; our contract draft was wrong, not the code
EvolvingLMMs-Lab/SkillOpt-Lite/skillopt/model/codex_harness.py
code served (permissive licence) · get_code("9e6ac1f0dd1091da")
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 dict. 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.

# Parse the function to understand what it needs
# The function takes a raw string parameter and parses it
# It looks for lines that are in the set {'user', 'codex', 'exec'}
# These are marker lines that denote step boundaries

# Create a sample raw string with the expected markers and content
raw_input = """some initial content
user
first step content
here
codex
second step content
more content
exec
final step content
trailing content"""

ARGS = {
    'raw': raw_input
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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
render_skill_md Ran
ran; our contract draft was wrong, not the code
EvolvingLMMs-Lab/SkillOpt-Lite/skillopt/model/codex_harness.py
code served (permissive licence) · get_code("291720a2ed08338f")
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 str. 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 numpy as np

torch.manual_seed(0)

# Analyzing the function:
# - skill_content: str - main content, gets stripped and used as body
# - name: str - used in f-string
# - description: str - used in f-string
# - preamble: str - optional, gets stripped and conditionally added

ARGS = {
    'skill_content': 'def solve(state):\n    return state.apply_action("think")',
    'name': 'my-custom-skill',
    'description': 'A custom skill for solving tasks with reflection',
    'preamble': 'This skill enables step-by-step reasoning and action planning.'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str 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
flatten_config Not yet run
unverified (harvested, no recorded run)
EvolvingLMMs-Lab/SkillOpt-Lite/skillopt/config.py
code served (permissive licence) · get_code("f9c45943c6e80129")
How it ran

no invocation recorded.

Also present in other papers' code

Of 9 samples lifted from this paper's linked repositories, 9 carry a function identity and 0 do not. Of the 9 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 9 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

While skill optimization for autonomous agents has gained traction, existing methods rely on complex pipelines. This leaves a fundamental question unaddressed: What constitutes a minimal viable pipeline for skill optimization, where every component is justified by theory or empirical necessity? We formalize skill optimization via Zeroth-Order (ZO) optimization, mapping classical counterparts (central difference, trust regions) to recent literature. Noting that unlike blind numerical perturbations in classical ZO, skill trajectories serve as interpretable debugging feedback. Grounded in Claude Code philosophy and PAC learning, we establish three principles for convergence and generalization: file-system-based trajectory exploration, consensus attribute mining, and independent validation gating. Eliminating redundancies, we propose SkillOpt-Lite. It accelerates convergence and outperforms full SkillOpt: improving LiveMath by +8.8 points on GPT-5.5 and +25.4 points on GPT-5.4-nano, allowing the nano model to surpass standard GPT-5.4 optimized by SkillOpt. Finally, we integrate our framework into production coding agents like VSCode Copilot, enabling developers to evolve agent skills via one line of vibe. Because our framework treats all agent components simply as standard editable code, this minimal pipeline naturally generalizes to full harness optimization (HarnessOpt). On SpreadsheetBench, HarnessOpt enables GPT-5.4-nano to achieve 0.7758 accuracy, outperforming the larger GPT-5.5 running standard pipelines (0.7620). Code is available at https://github.com/EvolvingLMMs-Lab/SkillOpt-Lite.

For agents

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

get_harvested_code_for_paper("2607.03451")
get_code_for_paper("2607.03451")
have("2607.03451")

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

curl https://syntology.ai/api/ran/2607.03451.json

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

[![Syntology run record](https://syntology.ai/api/ran/2607.03451.svg)](https://syntology.ai/paper/2607.03451)

Connect an agent — have() is free.