SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2607.26784 · 2026

SkillRise: Agentic Reinforcement Learning for Cross-Task Skill Evolution

Yuxin Chen, Zihan Wang, Yongliang Shen, Zishan Xu, Weiwen Liu, Xunliang Cai, Zhengxi Lu, Yueqing Sun, Yuquan Lu, Yifu Guo, Qi Gu, Ziang Ye, and 4 more

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 (3 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 0 violated it and 3 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/2607.26784.json · badge: /api/ran/2607.26784.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
Within-yao/SkillRise community
Papers with Code archive
10 ran · 1 no recorded run
FunctionStatusWhere it lives
EnvironmentManagerBase Ran
ran on a synthesized input · constructed an object; no result was computed
Within-yao/SkillRise/agent_system/environments/skillrise_alfworld/env_manager.py
code served (permissive licence) · get_code("5016ca4ef4d18a0a")
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, EnvironmentManagerBase: 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 collections import defaultdict

torch.manual_seed(0)

# Create a minimal mock environment
class MockEnv:
    def __init__(self, num_envs=2):
        self.num_envs = num_envs
    
    def reset(self):
        # Return observations and infos
        obs = np.random.rand(self.num_envs, 4, 3).astype(np.float32)
        infos = [{'step': 0} for _ in range(self.num_envs)]
        return obs, infos
    
    def step(self, actions):
        # Return next_obs, rewards, dones, infos
        next_obs = np.random.rand(self.num_envs, 4, 3).astype(np.float32)
        rewards = np.random.rand(self.num_envs).astype(np.float32)
        dones = np.zeros(self.num_envs, dtype=bool)
        infos = [{'won': False, 'task_score': 0.5} for _ in range(self.num_envs)]
        return next_obs, rewards, dones, infos
    
    def close(self):
        pass

# Create a minimal projection function
def projection_f(text_actions):
    # Map text actions to environment actions and validity flags
    num_actions = len(text_actions)
    actions = np.array([0] * num_actions, dtype=np.int64)
    valids = np.array([1.0] * num_actions, dtype=np.float32)
    return actions, valids

# Create a minimal config
class Config:
    def __init__(self):
        self.num_envs = 2
        self.max_steps = 100

# Construct the arguments
ARGS = {
    'envs': MockEnv(num_envs=2),
    'projection_f': projection_f,
    'config': Config()
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type EnvironmentManagerBase — 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 EnvironmentManagerBase 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
GroupLoader Ran
ran on a synthesized input · constructed an object; no result was computed
Within-yao/SkillRise/agent_system/environments/skillrise_alfworld/env_manager.py
code served (permissive licence) · get_code("f2045b8a81d620b5")
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, GroupLoader: 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 json
import os
import tempfile
import random

# Create a temporary directory and group file for testing
tmpdir = tempfile.mkdtemp()
os.environ['ALFWORLD_DATA'] = tmpdir

# Create the necessary directory structure
split = 'train'
json_dir = os.path.join(tmpdir, 'json_2.1.1', split)
os.makedirs(json_dir, exist_ok=True)

# Create sample task directories and game files
task_dir_1 = 'pick_and_place_simple-State_0'
trial_1 = 'trial_1'
task_path_1 = os.path.join(json_dir, task_dir_1, trial_1)
os.makedirs(task_path_1, exist_ok=True)
with open(os.path.join(task_path_1, 'game.tw-pddl'), 'w') as f:
    f.write('dummy game file 1')

task_dir_2 = 'pick_and_place_simple-State_1'
trial_2 = 'trial_1'
task_path_2 = os.path.join(json_dir, task_dir_2, trial_2)
os.makedirs(task_path_2, exist_ok=True)
with open(os.path.join(task_path_2, 'game.tw-pddl'), 'w') as f:
    f.write('dummy game file 2')

# Create a group file (JSONL format)
group_file = os.path.join(tmpdir, 'groups.jsonl')
with open(group_file, 'w') as f:
    group_1 = {
        'K': 2,
        'tasks': [
            {'task_dir': task_dir_1, 'trial': trial_1},
            {'task_dir': task_dir_2, 'trial': trial_2}
        ]
    }
    f.write(json.dumps(group_1) + '\n')
    
    group_2 = {
        'K': 2,
        'tasks': [
            {'task_dir': task_dir_1, 'trial': trial_1},
            {'task_dir': task_dir_2, 'trial': trial_2}
        ]
    }
    f.write(json.dumps(group_2) + '\n')

ARGS = {
    'group_file': group_file,
    'batch_groups': 2,
    'seed': 42,
    'split': 'train'
}
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 GroupLoader — 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 GroupLoader 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
SimpleMemory Ran
ran on a synthesized input · constructed an object; no result was computed
Within-yao/SkillRise/agent_system/environments/skillrise_alfworld/env_manager.py
code served (permissive licence) · get_code("14d332dcdd01f1b5")
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, SimpleMemory: 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.

class SimpleMemory:
    """
    Memory manager: responsible for storing & fetching per‑environment history records.
    """

    def __init__(self):
        self._data = None
        self.keys = None
        self.batch_size = 0

    def __len__(self):
        return len(self._data)

    def __getitem__(self, idx):
        return self._data[idx]

    def reset(self, batch_size: int):
        if self._data is not None:
            self._data.clear()
        self._data = [[] for _ in range(batch_size)]
        self.batch_size = batch_size
        self.keys = None

    def store(self, record: dict):
        """
        Store a new record (one step of history) for each environment instance.
        """
        if self.keys is None:
            self.keys = list(record.keys())
        assert self.keys == list(record.keys())
        for env_idx in range(self.batch_size):
            self._data[env_idx].append({k: record[k][env_idx] for k in self.keys})

    def fetch(self, history_length: int, obs_key: str='text_obs', action_key: str='action'):
        """
        Fetch and format recent interaction history for each environment instance.
        """
        memory_contexts, valid_lengths, dones = ([], [], [])
        for env_idx in range(self.batch_size):
            recent = self._data[env_idx][-history_length:]
            valid_len = len(recent)
            start_idx = len(self._data[env_idx]) - valid_len
            lines = []
            for j, rec in enumerate(recent):
                step_num = start_idx + j + 1
                act = rec[action_key]
                obs = rec[obs_key]
                lines.append(f'Action {step_num}: {act}\nObservation {step_num}: {obs}')
                if 'dones' in rec.keys() and rec['dones']:
                    valid_len = step_num
                    break
            memory_contexts.append('\n'.join(lines))
            valid_lengths.append(valid_len)
        return (memory_contexts, valid_lengths)

    def _fetch(self, history_length: int, obs_key: str='text_obs', action_key: str='action'):
        """
        Fetch and format recent interaction history for each environment instance.
        """
        memory_contexts, obs_contexts, action_contexts, valid_lengths = ([], [], [], [])
        for env_idx in range(self.batch_size):
            recent = self._data[env_idx][-history_length:]
            valid_len = len(recent)
            start_idx = len(self._data[env_idx]) - valid_len
            obs_lines, action_lines, context_lines = ([], [], [])
            for j, rec in enumerate(recent):
                step_num = start_idx + j + 1
                act = rec[action_key]
                obs = rec[obs_key]
                obs_lines.append(f"Observation {step_num}: '{obs}'")
                action_lines.append(f"Step {step_num}: '{act}'")
                context_lines.append(f"[Observation {step_num}: '{obs}', Step {step_num}: '{act}']")
            obs_contexts.append('\n'.join(obs_lines))
            action_contexts.append('\n'.join(action_lines))
            memory_contexts.append('\n'.join(context_lines))
            valid_lengths.append(valid_len)
        return (memory_contexts, obs_contexts, action_contexts, valid_lengths)


# Setup for SimpleMemory instantiation
ARGS = {}
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 SimpleMemory — 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 SimpleMemory 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_reflections Ran
ran; our contract draft was wrong, not the code
Within-yao/SkillRise/agent_system/environments/skillrise_alfworld/env_manager.py
code served (permissive licence) · get_code("b680620d9ddaabe8")
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

def _format_reflections(reflections_dict, types_outcomes):
    """reflections_dict: {pos: text}; types_outcomes: {pos: (task_type, won)}."""
    if not reflections_dict:
        return '\n(No earlier experience yet.)'
    parts = []
    for pos in sorted(reflections_dict):
        tt, won = types_outcomes.get(pos, ('a task', False))
        parts.append('\n- After an earlier task ({task_type}, outcome: {outcome}), you noted:\n{reflection}'.format(task_type=tt, outcome='SUCCESS' if won else 'FAILURE', reflection=(reflections_dict[pos] or '').strip()))
    return ''.join(parts)

ARGS = {
    'reflections_dict': {
        0: 'The model converged quickly with learning rate 0.01',
        1: 'Batch normalization helped reduce variance',
        3: 'Increasing hidden layer size improved accuracy'
    },
    'types_outcomes': {
        0: ('classification', True),
        1: ('regression', True),
        3: ('hyperparameter_tuning', False)
    }
}
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
_format_skill Ran
ran; our contract draft was wrong, not the code
Within-yao/SkillRise/agent_system/environments/skillrise_alfworld/env_manager.py
code served (permissive licence) · get_code("1990312b38c5e39c")
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.

ARGS = {
    'skill': 'machine learning expertise'
}
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
_split_root Ran
ran; our contract draft was wrong, not the code
Within-yao/SkillRise/agent_system/environments/skillrise_alfworld/env_manager.py
code served (permissive licence) · get_code("072456c415283959")
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

# Set up environment variable for ALFWORLD_DATA
os.environ['ALFWORLD_DATA'] = '/tmp/alfworld_data'

# The function _split_root takes a single parameter: split
# It uses the split parameter as a path component in os.path.join
# The split parameter is a string with default value 'train'

ARGS = {
    'split': 'train'
}
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
get_reflect_prompt Ran
ran; our contract draft was wrong, not the code
Within-yao/SkillRise/agent_system/environments/skillrise_alfworld/env_manager.py
code served (permissive licence) · get_code("84fd6da202de7aec")
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

# Call manual_seed for reproducibility
torch.manual_seed(0)

# Build the ARGS dictionary with valid arguments for get_reflect_prompt
ARGS = {
    'phase': 'play',
    'turn_idx': 0,
    'init_observation': 'You are in a kitchen. You see a table with a cup on it.',
    'curr_traj': 'Action: move forward\nObservation: You moved closer to the table.\nAction: pick up cup\nObservation: You picked up the cup.',
    'admissible_actions': 'move forward, move backward, turn left, turn right, pick up, put down, open, close',
    'reflections_text': 'From earlier attempts:\n- Always check receptacles before placing items\n- Objects are typically found in their expected locations\n- Opening drawers first helps locate smaller items',
    'won': False,
}
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
get_skillrise_prompt Ran
ran; our contract draft was wrong, not the code
Within-yao/SkillRise/agent_system/environments/skillrise_alfworld/env_manager.py
code served (permissive licence) · get_code("45a6f77cbc6355c7")
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

# Define test arguments for get_skillrise_prompt function
ARGS = {
    'phase': 'play',
    'turn_idx': 0,
    'init_observation': 'You are in a kitchen. You see a table with a cup on it.',
    'curr_traj': 'Action: move to table\nObservation: You moved closer to the table.',
    'admissible_actions': 'move to table, pick up cup, open cabinet, close cabinet',
    'skill': 'When picking up objects, first move to the object location, then execute the pick-up action.',
    'won': False,
}
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
resolve_gamefile Ran
ran; our contract draft was wrong, not the code
Within-yao/SkillRise/agent_system/environments/skillrise_alfworld/env_manager.py
code served (permissive licence) · get_code("cae4d6bdb53d3044")
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

# Set up environment variable for ALFWORLD_DATA
os.environ['ALFWORLD_DATA'] = '/tmp/alfworld_data'

# Define arguments for resolve_gamefile function
ARGS = {
    'task_dir': 'kitchen',
    'trial': 'trial_1',
    'split': 'train'
}
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
to_numpy Ran
ran, honoured the contract we drafted
Within-yao/SkillRise/agent_system/environments/skillrise_alfworld/env_manager.py
code served (permissive licence) · get_code("21853363fde13d45")
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 ndarray, ndim 2, dtype family float, shape (2, 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
import numpy as np

torch.manual_seed(0)

# Test case 1: torch.Tensor
tensor_data = torch.randn(2, 4)

# Test case 2: numpy.ndarray
numpy_data = np.array([[1, 2, 3, 4], [5, 6, 7, 8]])

# Test case 3: list (will be converted to numpy array)
list_data = [[1, 2, 3, 4], [5, 6, 7, 8]]

# Use the tensor case as the primary argument
ARGS = {
    "data": tensor_data
}
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 ndarray, ndim 2, dtype family float, shape (2, 4) — shape (2, 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
SkillRiseAlfWorldEnvironmentManager Not yet run
unverified (harvested, no recorded run)
Within-yao/SkillRise/agent_system/environments/skillrise_alfworld/env_manager.py
code served (permissive licence) · get_code("d9be5f536bfd266d")
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

Large language model agents often encounter related yet distinct tasks that share reusable solution patterns. Yet standard agentic reinforcement learning treats tasks as independent episodes, while existing approaches to skill learning either focus on repeated attempts of one task or use pipelines with multiple stages that entangle extraction, retrieval, and execution. We introduce SkillRise, a unified reinforcement learning framework for learning skills across tasks. SkillRise organizes related instances into progressively challenging sequences and uses a single policy to alternate between task solving and curating an evolving skill document passed directly to the next task. Decoupled credit assignment across tasks supervises solving with the current task outcome and curation with discounted downstream outcomes. Experiments on ALFWorld, WebShop, and Sci-enceWorld show that SkillRise achieves the strongest Pass@1 performance among the compared methods, with gains over the strongest baseline ranging from 2.3 to 8.5 percentage points. Although trained across distinct tasks, its learned curation policy remains effective for repeated attempts on the same task. Further analysis reveals scaling at test time across tasks: performance improves with longer sequences of related tasks even when each task is attempted only once. This trend suggests that SkillRise reuses transferable skills across tasks rather than benefiting from repeated sampling of the same task. SkillRise further retains strong performance while substantially reducing the runtime overhead of skill learning pipelines with multiple stages. Together, these results provide a simple and efficient training paradigm for LLM agents to extract, refine, and reuse transferable skills across tasks. Our code is publicly available at https://github.com/Within-yao/SkillRise.

For agents

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

get_harvested_code_for_paper("2607.26784")
get_code_for_paper("2607.26784")
have("2607.26784")

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

curl https://syntology.ai/api/ran/2607.26784.json

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

[![Syntology run record](https://syntology.ai/api/ran/2607.26784.svg)](https://syntology.ai/paper/2607.26784)

Connect an agent — have() is free.