SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2604.11554 · 2026

Relax: An Asynchronous Reinforcement Learning Engine for Omni-Modal Post-Training at Scale

Jia Liu, Lei Zhang, Rui Yang, Minghao Li, Xiaoyan Yu, Jiaxing Li, Weihang Chen, Liujie Zhang, Benzhe Ning, Lumeng Wu, Weiqi Hu

arXiv · PDF · Open in the Atlas

Code that ran

11 of 13 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; 9 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/2604.11554.json · badge: /api/ran/2604.11554.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
redai-infra/Relax canonical (classifier label)
classifier basis: census
11 ran · 2 no recorded run
FunctionStatusWhere it lives
agentic_trace_events Ran
ran; our contract draft was wrong, not the code
redai-infra/Relax/relax/agentic/profile.py
code served (permissive licence) · get_code("3bbaae40add33da2")
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 torch
import numpy as np

TRACE_KEY = 'agentic_trace'
EVENTS_KEY = 'events'

# Construct a valid metadata dict to pass to agentic_trace_events
metadata = {
    TRACE_KEY: {
        EVENTS_KEY: {}
    }
}

ARGS = {
    'metadata': metadata
}
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 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
is_model_uri Ran
ran, violated the contract we drafted
redai-infra/Relax/relax/utils/model_source.py
code served (permissive licence) · get_code("22f3bcebf5ebdc62")
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 re

_MODEL_URI_PREFIX = re.compile('^[A-Za-z][A-Za-z0-9+.-]*://')

def is_model_uri(path: str) -> bool:
    """Return whether a model path has a hierarchical URI-shaped prefix."""
    return _MODEL_URI_PREFIX.match(path) is not None


ARGS = {
    "path": "s3://my-bucket/model.pth"
}
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 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
is_s3_uri Ran
ran, violated the contract we drafted
redai-infra/Relax/relax/utils/s3_model_loader.py
code served (permissive licence) · get_code("8535621d145eddb0")
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

ARGS = {
    "uri": "s3://my-bucket/my-key"
}
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
merge_agentic_trace Ran
ran; our contract draft was wrong, not the code
redai-infra/Relax/relax/agentic/profile.py
code served (permissive licence) · get_code("ae680471939332ba")
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 copy
from typing import Any

TRACE_KEY = 'agentic_trace'
EVENTS_KEY = 'events'

def merge_agentic_trace(base_trace: dict[str, Any] | None, overlay_trace: dict[str, Any] | None) -> dict[str, Any]:
    if base_trace is not None and (not isinstance(base_trace, dict)):
        raise TypeError(f'{TRACE_KEY} must be a dict, got {type(base_trace)}')
    if overlay_trace is not None and (not isinstance(overlay_trace, dict)):
        raise TypeError(f'{TRACE_KEY} overlay must be a dict, got {type(overlay_trace)}')
    merged = copy.deepcopy(base_trace or {})
    if not overlay_trace:
        return merged
    for key, value in overlay_trace.items():
        if key == EVENTS_KEY:
            if not isinstance(value, dict):
                raise TypeError(f'{TRACE_KEY}.{EVENTS_KEY} overlay must be a dict, got {type(value)}')
            base_events = merged.get(EVENTS_KEY)
            if base_events is not None and (not isinstance(base_events, dict)):
                raise TypeError(f'{TRACE_KEY}.{EVENTS_KEY} must be a dict, got {type(base_events)}')
            events = copy.deepcopy(base_events or {})
            events.update(copy.deepcopy(value))
            merged[EVENTS_KEY] = events
            continue
        merged[key] = copy.deepcopy(value)
    return merged

# Setup valid arguments for merge_agentic_trace
base_trace_example = {
    'status': 'running',
    'events': {
        'event_1': {'type': 'start', 'timestamp': 0.0},
        'event_2': {'type': 'step', 'timestamp': 1.0}
    },
    'metadata': {'version': 1}
}

overlay_trace_example = {
    'status': 'completed',
    'events': {
        'event_2': {'type': 'step', 'timestamp': 1.5},
        'event_3': {'type': 'end', 'timestamp': 2.0}
    },
    'extra_field': 'extra_value'
}

ARGS = {
    'base_trace': base_trace_example,
    'overlay_trace': overlay_trace_example
}
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
model_source_path_aliases Ran
ran; our contract draft was wrong, not the code
redai-infra/Relax/relax/utils/model_source.py
code served (permissive licence) · get_code("c7e7d601ab85b83e")
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 set. 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 argparse import Namespace

# Create a Namespace object with the attributes used by model_source_path_aliases
args = Namespace()

# Set the _model_source_original_hf_checkpoint attribute
args._model_source_original_hf_checkpoint = "models/bert-base-uncased"

# Define the source_uri parameter
source_uri = "s3://my-bucket/model-checkpoint"

# Construct the ARGS dictionary
ARGS = {
    'args': args,
    'source_uri': source_uri,
}
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 set — a set output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a set 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
normalize_model_path Ran
ran; our contract draft was wrong, not the code
redai-infra/Relax/relax/utils/model_source.py
code served (permissive licence) · get_code("d3041d5314d3922f")
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 os
import re

_MODEL_URI_PREFIX = re.compile('^[A-Za-z][A-Za-z0-9+.-]*://')

def is_model_uri(path: str) -> bool:
    """Return whether a model path has a hierarchical URI-shaped prefix."""
    return _MODEL_URI_PREFIX.match(path) is not None

def normalize_model_path(path: str) -> str:
    """Normalize local path spelling without rewriting model URIs."""
    return path if is_model_uri(path) else os.path.normpath(path)

# Test cases covering both URI paths and local paths
ARGS = {
    "path": "s3://bucket/model.pt"
}
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
register_genrm Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
redai-infra/Relax/relax/core/optional_roles.py
code served (permissive licence) · get_code("6520cfb3d52e8d65")
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.

from argparse import Namespace

torch_manual_seed = __import__('torch').manual_seed(0)

ARGS = {
    'config': Namespace(genrm_model_path=None),
    'algo': {}
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
relax — 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
register_vcs_handler Ran
ran; our contract draft was wrong, not the code
redai-infra/Relax/relax/_version.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from DeepExperience/HyperEyes, is pointer only (licence: NONE) · get_code("45333e6d0a056314")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from DeepExperience/HyperEyes, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

run_command Ran
ran; our contract draft was wrong, not the code
redai-infra/Relax/relax/_version.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from DeepExperience/HyperEyes, is pointer only (licence: NONE) · get_code("e36817d1c9cffe87")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from DeepExperience/HyperEyes, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

vanilla_tis_function Ran
ran; our contract draft was wrong, not the code
redai-infra/Relax/relax/backends/megatron/loss.py
code served (permissive licence) · get_code("eb755a361a0a18f5")
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. 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 simple args object with the required attributes
class Args:
    def __init__(self):
        self.tis_clip_low = 0.5
        self.tis_clip = 2.0

args = Args()

# Create small test tensors
batch_size = 2
tensor_dim = 4

# pg_loss: tensor that will be multiplied by tis_weights
pg_loss = torch.randn(batch_size, dtype=torch.float32)

# train_log_probs: list of tensors to be concatenated
train_log_probs = [
    torch.randn(1, dtype=torch.float32),
    torch.randn(1, dtype=torch.float32)
]

# rollout_log_probs: list of tensors to be concatenated
rollout_log_probs = [
    torch.randn(1, dtype=torch.float32),
    torch.randn(1, dtype=torch.float32)
]

# loss_masks: list of tensors (returned as-is)
loss_masks = [
    torch.ones(1, dtype=torch.float32),
    torch.ones(1, dtype=torch.float32)
]

ARGS = {
    'args': args,
    'pg_loss': pg_loss,
    'train_log_probs': train_log_probs,
    'rollout_log_probs': rollout_log_probs,
    'loss_masks': loss_masks
}
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 — a tuple output has no array shape of its own; none is 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
versions_from_parentdir Ran
ran; our contract draft was wrong, not the code
redai-infra/Relax/relax/_version.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from DeepExperience/HyperEyes, is pointer only (licence: NONE) · get_code("74664b7fba1494f5")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from DeepExperience/HyperEyes, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

register_extra_roles Not yet run
unverified (harvested, no recorded run)
redai-infra/Relax/relax/core/optional_roles.py
code served (permissive licence) · get_code("9b607beef3d5ebbe")
How it ran

no invocation recorded.

register_sft_rollout Not yet run
unverified (harvested, no recorded run)
redai-infra/Relax/relax/core/optional_roles.py
code served (permissive licence) · get_code("71fbbe2b5e2ed61b")
How it ran

no invocation recorded.

Also present in other papers' code

Of 13 samples lifted from this paper's linked repositories, 13 carry a function identity and 0 do not. Of the 13 distinct functions with an identity, 3 are also present in other papers whose code we hold (3 with at least one paper dated earlier, 0 the same month, 3 later) and 10 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
register_vcs_handler 3 other papers
2 dated earlier, 1 later
2001.00888 (2020-01)
Towards Scalable Dataframe Systems
2307.15295 (2023-07)
Sensitivity of the IceCube Upgrade to Atmospheric Neutrino Oscillations
2605.07177 (2026-05)
HyperEyes: Dual-Grained Efficiency-Aware Reinforcement Learning for Parallel Multimodal Search Agents
run_command 3 other papers
2 dated earlier, 1 later
2001.00888 (2020-01)
Towards Scalable Dataframe Systems
2307.15295 (2023-07)
Sensitivity of the IceCube Upgrade to Atmospheric Neutrino Oscillations
2605.07177 (2026-05)
HyperEyes: Dual-Grained Efficiency-Aware Reinforcement Learning for Parallel Multimodal Search Agents
versions_from_parentdir 3 other papers
2 dated earlier, 1 later
2001.00888 (2020-01)
Towards Scalable Dataframe Systems
2307.15295 (2023-07)
Sensitivity of the IceCube Upgrade to Atmospheric Neutrino Oscillations
2605.07177 (2026-05)
HyperEyes: Dual-Grained Efficiency-Aware Reinforcement Learning for Parallel Multimodal Search Agents

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

Reinforcement learning (RL) post-training has proven effective at unlocking reasoning, self-reflection, and tool-use capabilities in large language models. As models extend to omni-modal inputs and agentic multi-turn workflows, RL training systems face three interdependent challenges: heterogeneous data flows, operational robustness at scale, and the staleness-throughput tradeoff. We present Relax (Reinforcement Engine Leveraging Agentic X-modality), an open-source RL training engine that addresses these challenges through three co-designed architectural layers. First, an omni-native architecture builds multimodal support into the full stack-from data preprocessing and modality-aware parallelism to inference generation-rather than retrofitting it onto a text-centric pipeline. Second, each RL role runs as an independent, fault-isolated service that can be scaled, recovered, and upgraded without global coordination. Third, service-level decoupling enables asynchronous training via the TransferQueue data bus, where a single staleness parameter smoothly interpolates among on-policy, near-on-policy, and fully asynchronous execution. Relax achieves a 1.20× end-to-end speedup over veRL on Qwen3-4B on-policy training. Its fully async mode delivers a 1.76× speedup over colocate on Qwen3-4B and a 2.00× speedup on Qwen3-Omni-30B, while all modes converge to the same reward level. Relax supports R3 (Rollout Routing Replay) [1] for MoE models with only 1.9% overhead, compared to 32% degradation in veRL under the same configuration. It further demonstrates stable omni-modal RL convergence on Qwen3-Omni across image, text, and audio, sustaining over 2,000 steps on video without degradation. Relax is available at https://github.com/redai-infra/Relax.

For agents

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

get_harvested_code_for_paper("2604.11554")
get_code_for_paper("2604.11554")
have("2604.11554")

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

curl https://syntology.ai/api/ran/2604.11554.json

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

[![Syntology run record](https://syntology.ai/api/ran/2604.11554.svg)](https://syntology.ai/paper/2604.11554)

Connect an agent — have() is free.