SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2605.16883 · ICML · 2026

SE-GA: Memory-Augmented Self-Evolution for GUI Agents

Zhuosheng Zhang, Lanjun Wang, Shilong Jin

arXiv · PDF · Open in the Atlas

Code that ran

10 of 11 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (8 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 8 have no contract verdict; 2 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/2605.16883.json · badge: /api/ran/2605.16883.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
jinshilong-dev/SE-GA community
Papers with Code archive
10 ran · 1 no recorded run
FunctionStatusWhere it lives
EmbeddingProvider Ran
ran on a synthesized input · constructed an object; no result was computed
jinshilong-dev/SE-GA/gui_sega/ttme/memory/ttme.py
code served (permissive licence) · get_code("9f3f739c1707ce64")
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, EmbeddingProvider: 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 os
from unittest.mock import Mock, MagicMock

# Mock the OpenAI client to avoid actual API calls
class MockEmbeddingResponse:
    def __init__(self):
        self.data = [Mock(embedding=[0.1, 0.2, 0.3, 0.4])]

class MockEmbeddingsCreate:
    def __call__(self, model, input):
        return MockEmbeddingResponse()

class MockOpenAIClient:
    def __init__(self, api_key=None, base_url=None):
        self.api_key = api_key
        self.base_url = base_url
        self.embeddings = Mock()
        self.embeddings.create = MockEmbeddingsCreate()

# Patch the OpenAI import
import sys
mock_openai_module = Mock()
mock_openai_module.OpenAI = MockOpenAIClient
sys.modules['openai'] = mock_openai_module

ARGS = {
    'api_key': 'test-api-key-12345',
    'base_url': 'https://api.openai.com/v1',
    'model': 'text-embedding-3-small'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
openai — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type EmbeddingProvider — 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 EmbeddingProvider 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
EpisodicEntry Ran
ran on a synthesized input · constructed an object; no result was computed
jinshilong-dev/SE-GA/gui_sega/ttme/memory/ttme.py
code served (permissive licence) · get_code("a9321b3855e5417c")
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, EpisodicEntry: 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 numpy as np
from typing import List, Optional, Dict, Any
from dataclasses import dataclass, field, asdict

torch.manual_seed(0)

@dataclass
class EpisodicEntry:
    observation: str
    action: str
    next_observation: str

    def to_dict(self) -> Dict[str, Any]:
        return asdict(self)

    @classmethod
    def from_dict(cls, d: Dict[str, Any]) -> 'EpisodicEntry':
        return cls(**d)

ARGS = {
    "observation": "The agent is at position (0, 0) facing north",
    "action": "move_forward",
    "next_observation": "The agent is at position (0, 1) facing north"
}
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 EpisodicEntry — 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 EpisodicEntry 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
EpisodicMemory Ran
ran on a synthesized input · constructed an object; no result was computed
jinshilong-dev/SE-GA/gui_sega/ttme/memory/ttme.py
code served (permissive licence) · get_code("c3c51c8ad6873b35")
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, EpisodicMemory: 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.

from dataclasses import dataclass, asdict
from typing import List, Dict, Any

@dataclass
class EpisodicEntry:
    observation: str
    action: str
    next_observation: str

    def to_dict(self) -> Dict[str, Any]:
        return asdict(self)

    @classmethod
    def from_dict(cls, d: Dict[str, Any]) -> 'EpisodicEntry':
        return cls(**d)

class EpisodicMemory:
    """
    Episodic Memory (M_EPI): Short-term working memory that tracks the
    immediate task progress and executed historical actions.

    Data structure: M_EPI^t = [m_k]_{k=1}^{t-1}
    where m_k = <o_k, a_k, o_{k+1}>

    Retrieval: Sliding window with fixed horizon H.
    C_epi^t = [m_k]_{k=epsilon}^{t-1}, epsilon = max(1, t-H)
    """

    def __init__(self, horizon: int=10):
        self.horizon = horizon
        self._entries: List[EpisodicEntry] = []

    @property
    def entries(self) -> List[EpisodicEntry]:
        return self._entries

    def add(self, observation: str, action: str, next_observation: str):
        entry = EpisodicEntry(observation=observation, action=action, next_observation=next_observation)
        self._entries.append(entry)

    def retrieve(self) -> List[EpisodicEntry]:
        t = len(self._entries)
        epsilon = max(0, t - self.horizon)
        return self._entries[epsilon:]

    def build_context(self) -> str:
        windowed = self.retrieve()
        if not windowed:
            return ''
        lines = []
        for i, entry in enumerate(windowed):
            lines.append(f'Step {i + 1}: Observation: {entry.observation} | Action: {entry.action} | Result: {entry.next_observation}')
        return '\n'.join(lines)

    def clear(self):
        self._entries = []

    def __len__(self) -> int:
        return len(self._entries)

    def to_list(self) -> List[Dict[str, Any]]:
        return [e.to_dict() for e in self._entries]

    def from_list(self, data: List[Dict[str, Any]]):
        self._entries = [EpisodicEntry.from_dict(d) for d in data]


ARGS = {
    "horizon": 5
}
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 EpisodicMemory — 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 EpisodicMemory 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
ExperientialEntry Ran
ran on a synthesized input · constructed an object; no result was computed
jinshilong-dev/SE-GA/gui_sega/ttme/memory/ttme.py
code served (permissive licence) · get_code("3ab42b7bedec2705")
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, ExperientialEntry: 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
from typing import List, Optional, Dict, Any
from dataclasses import dataclass, field, asdict

torch.manual_seed(0)

@dataclass
class ExperientialEntry:
    trajectory: str
    reflective_summary: str
    intent_embedding: List[float] = field(default_factory=list)
    task_embedding: List[float] = field(default_factory=list)

    def to_dict(self) -> Dict[str, Any]:
        return asdict(self)

    @classmethod
    def from_dict(cls, d: Dict[str, Any]) -> 'ExperientialEntry':
        return cls(**d)


ARGS = {
    "trajectory": "agent moved from state A to state B via action sequence [move_left, move_right]",
    "reflective_summary": "learned that moving left then right returns to similar state",
    "intent_embedding": [0.1, 0.2, 0.3, 0.4],
    "task_embedding": [0.5, 0.6, 0.7, 0.8],
}
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 ExperientialEntry — 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 ExperientialEntry 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
ExperientialMemory Ran
ran on a synthesized input · constructed an object; no result was computed
jinshilong-dev/SE-GA/gui_sega/ttme/memory/ttme.py
code served (permissive licence) · get_code("bc04176bfa8de566")
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, ExperientialMemory: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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
from typing import List, Optional
from PIL import Image
from unittest.mock import Mock, MagicMock

torch.manual_seed(0)

# Create a mock EmbeddingProvider that doesn't require API keys
class MockEmbeddingProvider:
    def __init__(self, api_key: str = None, base_url: str = None, model: str = 'text-embedding-3-small'):
        self.model = model
        self.embedding_dim = 4
    
    def get_embedding(self, text: str) -> List[float]:
        # Return a deterministic embedding based on text length
        np.random.seed(hash(text) % (2**32))
        embedding = np.random.randn(self.embedding_dim).tolist()
        return embedding

# Create a mock VisualEncoder
class MockVisualEncoder:
    def __init__(self, model_name: str = 'clip-ViT-B-32', device: str = 'cpu'):
        self.model_name = model_name
        self.device = device
        self.embedding_dim = 3
    
    def encode_image(self, image: Image.Image) -> List[float]:
        # Return a deterministic embedding for images
        np.random.seed(42)
        embedding = np.random.randn(self.embedding_dim).tolist()
        return embedding

# Create mock objects
embedding_provider = MockEmbeddingProvider(api_key="test_key", base_url="http://localhost", model="text-embedding-3-small")
visual_encoder = MockVisualEncoder(model_name='clip-ViT-B-32', device='cpu')

ARGS = {
    'embedding_provider': embedding_provider,
    'visual_encoder': visual_encoder,
    'lambda_weight': 0.7,
    'top_k': 3
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy, openai, sentence_transformers, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ExperientialMemory — 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 ExperientialMemory 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
SemanticEntry Ran
ran on a synthesized input · constructed an object; no result was computed
jinshilong-dev/SE-GA/gui_sega/ttme/memory/ttme.py
code served (permissive licence) · get_code("6fd06ce921aa3b41")
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, SemanticEntry: 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 numpy as np
from typing import List, Optional, Dict, Any
from dataclasses import dataclass, field, asdict

torch.manual_seed(0)

@dataclass
class SemanticEntry:
    description: str
    embedding: List[float] = field(default_factory=list)

    def to_dict(self) -> Dict[str, Any]:
        return asdict(self)

    @classmethod
    def from_dict(cls, d: Dict[str, Any]) -> 'SemanticEntry':
        return cls(**d)

ARGS = {
    "description": "This is a semantic entry for a machine learning concept",
    "embedding": [0.1, 0.2, 0.3, 0.4, 0.5]
}
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 SemanticEntry — 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 SemanticEntry 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
SemanticMemory Ran
ran on a synthesized input · constructed an object; no result was computed
jinshilong-dev/SE-GA/gui_sega/ttme/memory/ttme.py
code served (permissive licence) · get_code("8cf7ee9fb618e448")
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, SemanticMemory: 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 math
from typing import List, Optional, Dict, Any
from dataclasses import dataclass, field, asdict


class MockEmbeddingProvider:
    """Mock embedding provider that returns consistent embeddings without API calls"""
    
    def __init__(self, api_key: str=None, base_url: str=None, model: str='text-embedding-3-small'):
        self.model = model
        self.embedding_dim = 4
        self._embedding_cache = {}
    
    def get_embedding(self, text: str) -> List[float]:
        text = text.replace('\n', ' ').strip()
        if not text:
            raise ValueError('Text cannot be empty')
        
        # Return consistent embeddings based on text hash
        if text not in self._embedding_cache:
            hash_val = hash(text)
            import random
            rng = random.Random(hash_val)
            embedding = [rng.uniform(-1, 1) for _ in range(self.embedding_dim)]
            # Normalize
            norm = math.sqrt(sum(x*x for x in embedding))
            embedding = [x/norm for x in embedding]
            self._embedding_cache[text] = embedding
        
        return self._embedding_cache[text]


# Create the mock embedding provider
embedding_provider = MockEmbeddingProvider(
    api_key="mock_key",
    base_url="http://localhost",
    model="text-embedding-3-small"
)

# Construct ARGS dictionary for SemanticMemory
ARGS = {
    "embedding_provider": embedding_provider,
    "top_k": 3
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
openai — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type SemanticMemory — 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 SemanticMemory 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
VisualEncoder Ran
ran on a synthesized input · constructed an object; no result was computed
jinshilong-dev/SE-GA/gui_sega/ttme/memory/ttme.py
code served (permissive licence) · get_code("201b4e8d12b7955c")
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, VisualEncoder: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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
from PIL import Image
from unittest.mock import Mock, MagicMock

torch.manual_seed(0)

# Create a mock SentenceTransformer that behaves like the real one
mock_model = Mock()

# Mock the encode method to return a numpy array that can be converted to list
def mock_encode(image, convert_to_numpy=True):
    # Return a small embedding vector
    return np.array([0.1, 0.2, 0.3, 0.4], dtype=np.float32)

mock_model.encode = mock_encode

# Create a small test image
test_image = Image.new('RGB', (32, 32), color='red')

ARGS = {
    'model_name': 'clip-ViT-B-32',
    'device': 'cpu'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy, sentence_transformers, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type VisualEncoder — 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 VisualEncoder 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
cosine_similarity Ran
ran; our contract draft was wrong, not the code
jinshilong-dev/SE-GA/gui_sega/ttme/memory/ttme.py
code served (permissive licence) · get_code("95334fc548831d33")
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 math
from typing import List

torch_imported = False
try:
    import torch
    torch_imported = True
except ImportError:
    pass

# Define ARGS dict with valid arguments for cosine_similarity
ARGS = {
    'vec1': [1.0, 2.0, 3.0, 4.0],
    'vec2': [2.0, 3.0, 4.0, 5.0],
}
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 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
top_k_indices Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
jinshilong-dev/SE-GA/gui_sega/ttme/memory/ttme.py
code served (permissive licence) · get_code("72ccb7048c526c5a")
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 int; 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
import numpy as np
from typing import List, Dict, Any

torch.manual_seed(0)

# Construct arguments for top_k_indices function
scores = [0.9, 0.1, 0.8, 0.2, 0.7, 0.3]
k = 3

ARGS = {
    "scores": scores,
    "k": k
}
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 list, ndim 1, dtype family int — 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
TTMEMemory Not yet run
unverified (harvested, no recorded run)
jinshilong-dev/SE-GA/gui_sega/ttme/memory/ttme.py
code served (permissive licence) · get_code("3d7a1b0348834580")
How it ran

no invocation recorded.

Also present in other papers' code

Of 11 samples lifted from this paper's linked repositories, 11 carry a function identity and 0 do not. Of the 11 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 11 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

Autonomous Graphical User Interface (GUI) agents often struggle with multi-step tasks due to constrained context windows and static policies that fail to adapt to dynamic environments. To address these limitations, this work proposes the Self-Evolving GUI Agent (SE-GA), a novel framework that integrates hierarchical memory structures with an iterative self-improvement mechanism. At the core of our approach is Test-Time Memory Extension (TTME), which facilitates long-term planning by dynamically retrieving episodic, semantic, and experiential memories to provide salient contexts during inference. To ensure continuous learning, we introduce Memory-Augmented Self-Evolution (MASE), which is a training pipeline that adopts the data collected by TTME to stabilize and enhance the agent's foundational policy. Extensive evaluations across both offline and online benchmarks demonstrate SE-GA achieves stateof-the-art performance, reaching success rates of 89.0% on ScreenSpot and 75.8% on the challenging AndroidControl-High dataset. Furthermore, significant improvements on the AndroidWorld benchmark highlight the superior generalization to dynamic environments. Open source code: https://github.com/jinshilong-dev/SE-GA

For agents

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

get_harvested_code_for_paper("2605.16883")
get_code_for_paper("2605.16883")
have("2605.16883")

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

curl https://syntology.ai/api/ran/2605.16883.json

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

[![Syntology run record](https://syntology.ai/api/ran/2605.16883.svg)](https://syntology.ai/paper/2605.16883)

Connect an agent — have() is free.