SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2502.19645 · 2025

Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success

arXiv · PDF · Open in the Atlas

Code that ran

4 of 4 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, 0 violated it and 0 have no contract verdict; 4 failures of our instrument; 0 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; 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/2502.19645.json · badge: /api/ran/2502.19645.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
moojink/openvla-oft extension (classifier label)
classifier basis: classifier_2025
4 ran · 0 no recorded run
FunctionStatusWhere it lives
get_run_id Ran
ran; our contract draft was wrong, not the code
moojink/openvla-oft/vla-scripts/finetune.py
code served (permissive licence) · get_code("8b2a2aa060e1e24f")
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) 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 torch.nn as nn

class FinetuneConfig:
    def __init__(
        self,
        run_id_override=None,
        resume=False,
        vla_path="models/vla_base_model",
        dataset_name="dataset_v1",
        batch_size=4,
        grad_accumulation_steps=2,
        learning_rate=1e-4,
        use_lora=False,
        lora_rank=8,
        lora_dropout=0.1,
        image_aug=False,
        run_id_note=None,
    ):
        self.run_id_override = run_id_override
        self.resume = resume
        self.vla_path = vla_path
        self.dataset_name = dataset_name
        self.batch_size = batch_size
        self.grad_accumulation_steps = grad_accumulation_steps
        self.learning_rate = learning_rate
        self.use_lora = use_lora
        self.lora_rank = lora_rank
        self.lora_dropout = lora_dropout
        self.image_aug = image_aug
        self.run_id_note = run_id_note

cfg_basic = FinetuneConfig()

ARGS = {
    "cfg": cfg_basic,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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
load_checkpoint Ran
ran; our contract draft was wrong, not the code
moojink/openvla-oft/vla-scripts/finetune.py
code served (permissive licence) · get_code("8c56534935f5a294")
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) 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 os
import torch
import tempfile

torch.manual_seed(0)

# Create a temporary directory for checkpoint storage
temp_dir = tempfile.mkdtemp()

# Create a minimal state dict with both DDP-prefixed and non-prefixed keys
state_dict_with_ddp = {
    'module.layer1.weight': torch.randn(4, 3),
    'module.layer1.bias': torch.randn(4),
    'layer2.weight': torch.randn(3, 2),
    'layer2.bias': torch.randn(3),
}

# Save the checkpoint
module_name = 'test_model'
step = 100
checkpoint_path = os.path.join(temp_dir, f'{module_name}--{step}_checkpoint.pt')
torch.save(state_dict_with_ddp, checkpoint_path)

# Define arguments for load_checkpoint
ARGS = {
    'module_name': module_name,
    'path': temp_dir,
    'step': step,
    'device': 'cpu',
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — 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
load_initial_states Ran
ran; our contract draft was wrong, not the code
moojink/openvla-oft/experiments/robot/libero/run_libero_eval.py
code served (permissive licence) · get_code("99b7442f72fa092c")
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) and returned tuple, 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 json
import logging
from dataclasses import dataclass
from enum import Enum
from pathlib import Path
from typing import Optional, Union
from io import StringIO

class TaskSuite(str, Enum):
    LIBERO_SPATIAL = 'libero_spatial'
    LIBERO_OBJECT = 'libero_object'
    LIBERO_GOAL = 'libero_goal'
    LIBERO_10 = 'libero_10'
    LIBERO_90 = 'libero_90'

logger = logging.getLogger(__name__)

@dataclass
class GenerateConfig:
    model_family: str = 'openvla'
    pretrained_checkpoint: Union[str, Path] = ''
    use_l1_regression: bool = True
    use_diffusion: bool = False
    num_diffusion_steps_train: int = 50
    num_diffusion_steps_inference: int = 50
    use_film: bool = False
    num_images_in_input: int = 2
    use_proprio: bool = True
    center_crop: bool = True
    num_open_loop_steps: int = 8
    lora_rank: int = 32
    unnorm_key: Union[str, Path] = ''
    load_in_8bit: bool = False
    load_in_4bit: bool = False
    task_suite_name: str = TaskSuite.LIBERO_SPATIAL
    num_steps_wait: int = 10
    num_trials_per_task: int = 50
    initial_states_path: str = 'DEFAULT'
    env_img_res: int = 256
    run_id_note: Optional[str] = None
    local_log_dir: str = './experiments/logs'
    use_wandb: bool = False
    wandb_entity: str = 'your-wandb-entity'
    wandb_project: str = 'your-wandb-project'
    seed: int = 7

def log_message(message: str, log_file=None):
    """Log a message to console and optionally to a log file."""
    logger.info(message)
    if log_file:
        log_file.write(message + '\n')
        log_file.flush()

def load_initial_states(cfg: GenerateConfig, task_suite, task_id: int, log_file=None):
    """Load initial states for the given task."""
    initial_states = task_suite.get_task_init_states(task_id)
    if cfg.initial_states_path != 'DEFAULT':
        with open(cfg.initial_states_path, 'r') as f:
            all_initial_states = json.load(f)
        log_message(f'Using initial states from {cfg.initial_states_path}', log_file)
        return (initial_states, all_initial_states)
    else:
        log_message('Using default initial states', log_file)
        return (initial_states, None)


class MockTaskSuite:
    """Mock task suite that implements the required interface."""
    def get_task_init_states(self, task_id: int):
        """Return mock initial states for a task."""
        return {
            'state_0': [0.1, 0.2, 0.3],
            'state_1': [0.4, 0.5, 0.6]
        }


# Create configuration
cfg = GenerateConfig(
    model_family='openvla',
    pretrained_checkpoint='',
    use_l1_regression=True,
    use_diffusion=False,
    num_diffusion_steps_train=50,
    num_diffusion_steps_inference=50,
    use_film=False,
    num_images_in_input=2,
    use_proprio=True,
    center_crop=True,
    num_open_loop_steps=8,
    lora_rank=32,
    unnorm_key='',
    load_in_8bit=False,
    load_in_4bit=False,
    task_suite_name=TaskSuite.LIBERO_SPATIAL,
    num_steps_wait=10,
    num_trials_per_task=50,
    initial_states_path='DEFAULT',
    env_img_res=256,
    run_id_note=None,
    local_log_dir='./experiments/logs',
    use_wandb=False,
    wandb_entity='your-wandb-entity',
    wandb_project='your-wandb-project',
    seed=7
)

# Create mock task suite
task_suite = MockTaskSuite()

# Create a log file (StringIO for testing)
log_file = StringIO()

ARGS = {
    'cfg': cfg,
    'task_suite': task_suite,
    'task_id': 0,
    'log_file': log_file
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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 tuple, ndim 1, 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
remove_ddp_in_checkpoint Ran
ran; our contract draft was wrong, not the code
moojink/openvla-oft/vla-scripts/finetune.py
code served (permissive licence) · get_code("c61f2447224a1296")
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) 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 torch
import torch.nn as nn

torch.manual_seed(0)

# Create a sample state_dict that mimics a DDP-trained model
# Some parameters have 'module.' prefix, some don't
state_dict = {
    'module.layer1.weight': torch.randn(4, 3),
    'module.layer1.bias': torch.randn(4),
    'module.layer2.weight': torch.randn(5, 4),
    'layer3.weight': torch.randn(2, 5),  # This one doesn't have 'module.' prefix
    'layer3.bias': torch.randn(2),
}

ARGS = {
    'state_dict': state_dict
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — 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

Also present in other papers' code

Of 4 samples lifted from this paper's linked repositories, 4 carry a function identity and 0 do not. Of the 4 distinct functions with an identity, 2 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 2 later) and 2 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
load_checkpoint 1 other paper
1 later
2603.04639 (2026-03)
RoboMME: Benchmarking and Understanding Memory for Robotic Generalist Policies
remove_ddp_in_checkpoint 1 other paper
1 later
2603.04639 (2026-03)
RoboMME: Benchmarking and Understanding Memory for Robotic Generalist Policies

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

Recent vision-language-action models (VLAs) build upon pretrained vision-language models and leverage diverse robot datasets to demonstrate strong task execution, language following ability, and semantic generalization. Despite these successes, VLAs struggle with novel robot setups and require fine-tuning to achieve good performance, yet how to most effectively fine-tune them is unclear given many possible strategies. In this work, we study key VLA adaptation design choices such as different action decoding schemes, action representations, and learning objectives for fine-tuning, using OpenVLA as our representative base model. Our empirical analysis informs an Optimized Fine-Tuning (OFT) recipe that integrates parallel decoding, action chunking, a continuous action representation, and a simple L1 regression-based learning objective to altogether improve inference efficiency, policy performance, and flexibility in the model's input-output specifications. We propose OpenVLA-OFT, an instantiation of this recipe, which sets a new state of the art on the LIBERO simulation benchmark, significantly boosting OpenVLA's average success rate across four task suites from 76.5% to 97.1% while increasing action generation throughput by 26$\times$. In real-world evaluations, our fine-tuning recipe enables OpenVLA to successfully execute dexterous, high-frequency control tasks on a bimanual ALOHA robot and outperform other VLAs ($π_0$ and RDT-1B) fine-tuned using their default recipes, as well as strong imitation learning policies trained from scratch (Diffusion Policy and ACT) by up to 15% (absolute) in average success rate. We release code for OFT and pretrained model checkpoints at https://openvla-oft.github.io/.

For agents

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

get_harvested_code_for_paper("2502.19645")
get_code_for_paper("2502.19645")
have("2502.19645")

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

curl https://syntology.ai/api/ran/2502.19645.json

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

[![Syntology run record](https://syntology.ai/api/ran/2502.19645.svg)](https://syntology.ai/paper/2502.19645)

Connect an agent — have() is free.