SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2510.16079 · ICML · 2025

EvolveR: Self-Evolving LLM Agents through an Experience-Driven Lifecycle

Task, Cheng Yang, Yuxin Wang, Botian Shi, Licheng Wen, Daocheng Fu, Pinlong Cai, Jianbiao Mei, Xuemeng Yang, Rong Wu, Xiaoman Wang, Yufan Shen

arXiv · PDF · Open in the Atlas

Code that ran

5 of 5 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; 1 honoured the contract we drafted, 1 violated it and 0 have no contract verdict; 3 failures of our instrument; 0 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/2510.16079.json · badge: /api/ran/2510.16079.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
Edaizi/EvolveR canonical (classifier label)
classifier basis: census
5 ran · 0 no recorded run
FunctionStatusWhere it lives
cosine_similarity Ran
ran, violated the contract we drafted
Edaizi/EvolveR/evolver/rewards/similarity_utils.py
code served (permissive licence) · get_code("ed0b0e48526b0be6")
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 float64, ndim 0, 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 numpy as np

# Create two 1D vectors for cosine similarity computation
v1 = np.array([1.0, 2.0, 3.0, 4.0])
v2 = np.array([2.0, 3.0, 4.0, 5.0])

ARGS = {
    "v1": v1,
    "v2": v2
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, sklearn — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float64, ndim 0, dtype family float — no output shape is stated: observe_signature.py recorded float ndim 0, which disagrees with the fixture run
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
extract_solution Ran
ran; our contract draft was wrong, not the code
Edaizi/EvolveR/evolver/rewards/reward_outcome.py
code served (permissive licence) · get_code("34d7767e8b83b614")
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 re

def extract_solution(solution_str):
    """Extract the equation from the solution string."""
    answer_pattern = '<answer>(.*?)</answer>'
    match = re.finditer(answer_pattern, solution_str, re.DOTALL)
    matches = list(match)
    if not matches:
        return None
    return matches[-1].group(1).strip()

# Test cases for the function
solution_str_with_answer = """
Some working shown here.
<answer>x = 5</answer>
More text after.
"""

solution_str_no_answer = "Just some text without answer tags"

solution_str_multiple_answers = """
<answer>first answer</answer>
Some more text
<answer>final answer</answer>
"""

ARGS = {
    "solution_str": solution_str_multiple_answers
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
format_reward_fn Ran
ran; our contract draft was wrong, not the code
Edaizi/EvolveR/evolver/rewards/reward_format.py
code served (permissive licence) · get_code("e1ff36770e74419a")
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 FormatRewardOutput. 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 re
from dataclasses import dataclass, field

@dataclass
class FormatRewardOutput:
    reward: float
    metrics: dict = field(default_factory=dict)

def format_reward_fn(solution_str: str, experience_enabled: bool=True) -> FormatRewardOutput:
    think_contents = re.findall('<think>(.*?)</think>', solution_str, re.DOTALL)
    think_count = len([c for c in think_contents if len(c.strip()) >= 3])
    search_experience_contents = re.findall('<search_experience>(.*?)</search_experience>', solution_str, re.DOTALL)
    search_experience_count = len([c for c in search_experience_contents if len(c.strip()) >= 3])
    search_knowledge_contents = re.findall('<search_knowledge>(.*?)</search_knowledge>', solution_str, re.DOTALL)
    search_knowledge_count = len([c for c in search_knowledge_contents if len(c.strip()) >= 3])
    answer_contents = re.findall('<answer>(.*?)</answer>', solution_str, re.DOTALL)
    has_answer = any((len(c.strip()) >= 3 for c in answer_contents))
    if not experience_enabled:
        hallucinated_experience_search = search_experience_count > 0
        total_search_count = search_knowledge_count
        format_valid = think_count > 0 and search_knowledge_count > 0 and has_answer and (not hallucinated_experience_search)
        reward = 1.0 if format_valid else 0.0
        return FormatRewardOutput(reward=reward, metrics={'think_count': think_count, 'search_experience_count': search_experience_count, 'search_knowledge_count': search_knowledge_count, 'total_search_count': total_search_count, 'has_answer': has_answer})
    think_score_mapping = {1: 0.2, 2: 0.3, 3: 0.4, 4: 0.6, 5: 0.8, 6: 1}
    think_score = think_score_mapping.get(think_count, 0.0)
    if think_count > 8:
        think_score = 0.5
    search_score = 0.0
    total_search_count = search_experience_count + search_knowledge_count
    has_search_experience = search_experience_count > 0
    has_search_knowledge = search_knowledge_count > 0
    if has_search_experience and has_search_knowledge:
        diversity_score = 0.5
    elif has_search_experience or has_search_knowledge:
        diversity_score = 0.2
    else:
        diversity_score = 0.0
    count_bonus = min((total_search_count - 1) * 0.1, 0.5) if total_search_count > 1 else 0.0
    search_score = diversity_score + count_bonus
    answer_score = 1.0 if has_answer else 0.0
    is_complete = think_count > 0 and total_search_count > 0 and has_answer
    if not is_complete:
        reward = 0.0
    else:
        think_weight = 1.0
        search_weight = 1.0
        total_score = think_score * think_weight + search_score * search_weight
        total_weight = think_weight + search_weight
        reward = total_score / total_weight
    return FormatRewardOutput(reward=reward, metrics={'think_count': think_count, 'search_experience_count': search_experience_count, 'search_knowledge_count': search_knowledge_count, 'total_search_count': total_search_count, 'has_answer': has_answer, 'think_score': think_score, 'search_score': search_score, 'answer_score': answer_score})

solution_with_all_tags = """
<think>
Let me think about this problem carefully and systematically.
</think>
<search_experience>
I recall from my experience that this approach works well.
</search_experience>
<search_knowledge>
The documentation states that this is the correct method.
</search_knowledge>
<answer>
The answer is 42.
</answer>
"""

ARGS = {
    'solution_str': solution_with_all_tags,
    'experience_enabled': True
}
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 FormatRewardOutput — a FormatRewardOutput output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a FormatRewardOutput 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_similarity_score Ran
ran, honoured the contract we drafted
Edaizi/EvolveR/evolver/rewards/similarity_utils.py
code served (permissive licence) · get_code("d7a998abfdafd523")
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 float32, 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 numpy as np
from unittest.mock import Mock, MagicMock
from openai import OpenAI

# Mock the EmbeddingClient since it requires an actual API connection
class MockEmbeddingClient:
    def __init__(self, api_url='http://127.0.0.1:8000/v1', api_key='empty', model_name='bge_m3'):
        self.api_url = api_url
        self.api_key = api_key
        self.model_name = model_name
        self.client = Mock()
    
    def get_embeddings(self, texts):
        """Return mock embeddings with consistent shape"""
        # Return a numpy array of shape (len(texts), 128) with small random values
        np.random.seed(42)
        return np.random.randn(len(texts), 128).astype(np.float32)

# Create instances for the ARGS dict
text1_str = "hello world"
text2_str = "hello there"
text1_list = ["hello world", "goodbye world"]
text2_list = ["hello there", "see you later"]

# Create a mock embedding client
embedding_client = MockEmbeddingClient()

# Build valid arguments for get_similarity_score
ARGS = {
    'text1': text1_str,
    'text2': text2_str,
    'embedding_client': embedding_client
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, openai, sklearn — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float32, 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
info_gain_reward_fn Ran
ran; our contract draft was wrong, not the code
Edaizi/EvolveR/evolver/rewards/reward_info_gain.py
code served (permissive licence) · get_code("12d53fdb88746504")
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 InfoGainRewardOutput. 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, field
from typing import List, Dict

class MockEmbeddingClient:
    """Mock embedding client with minimal interface."""
    def embed(self, text: str):
        return [0.1, 0.2, 0.3, 0.4]

ARGS = {
    'retrieved_docs': [
        {
            'title': 'Document One',
            'text': 'This is the first retrieved document.'
        },
        {
            'title': 'Document Two',
            'text': 'This is the second retrieved document.'
        },
        [
            {
                'title': 'Nested Document',
                'text': 'This is a nested document.'
            }
        ]
    ],
    'golden_docs': [
        {
            'title': 'Document One',
            'text': 'The golden version of document one.'
        },
        {
            'title': 'Document Three',
            'text': 'This is a golden document not in retrieved.'
        }
    ],
    'answer': 'This is the answer to the question based on the retrieved documents.',
    'embedding_client': MockEmbeddingClient(),
    'config': {
        'max_docs': 10,
        'similarity_threshold': 0.5
    }
}
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 InfoGainRewardOutput — a InfoGainRewardOutput output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a InfoGainRewardOutput 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 5 samples lifted from this paper's linked repositories, 5 carry a function identity and 0 do not. Of the 5 distinct functions with an identity, 5 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 0 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
cosine_similarity 1 other paper
1 undated
no dated paper among those named
openreview_BamDfCP5R3
From Interactions to Principles: Experience-Driven Self-Distillation for Evolving LLM Agents
extract_solution 1 other paper
1 undated
no dated paper among those named
openreview_BamDfCP5R3
From Interactions to Principles: Experience-Driven Self-Distillation for Evolving LLM Agents
format_reward_fn 1 other paper
1 undated
no dated paper among those named
openreview_BamDfCP5R3
From Interactions to Principles: Experience-Driven Self-Distillation for Evolving LLM Agents
get_similarity_score 1 other paper
1 undated
no dated paper among those named
openreview_BamDfCP5R3
From Interactions to Principles: Experience-Driven Self-Distillation for Evolving LLM Agents
info_gain_reward_fn 1 other paper
1 undated
no dated paper among those named
openreview_BamDfCP5R3
From Interactions to Principles: Experience-Driven Self-Distillation for Evolving LLM 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

Current Large Language Model (LLM) agents show strong performance in tool use, but lack the crucial capability to systematically learn from their own experiences. While existing frameworks mainly focus on mitigating external knowledge gaps, they fail to address a more fundamental limitation: the inability to iteratively refine problem-solving strategies. In this work, we introduce EvolveR, a framework designed to enable agent to self-improve through a complete, closedloop experience lifecycle. This lifecycle comprises two key stages: (1) Offline Self-Distillation, where the agent's interaction trajectories are synthesized into a structured repository of abstract, reusable strategic principles; (2) Online Interaction, where the agent interacts with tasks and actively retrieves distilled principles to guide its decision-making, accumulating a diverse set of behavioral trajectories. This loop employs a policy reinforcement mechanism to iteratively update the agent based on its performance. We demonstrate the effectiveness of EvolveR on complex multi-hop question-answering benchmarks, where it achieves superior performance over strong agentic baselines. Our work presents a comprehensive blueprint for agents that learn not only from external data but also from the consequences of their own actions, paving the way for more autonomous and continuously improving systems. Code is available at https://github.com/Edaizi/EvolveR.

For agents

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

get_harvested_code_for_paper("2510.16079")
get_code_for_paper("2510.16079")
have("2510.16079")

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

curl https://syntology.ai/api/ran/2510.16079.json

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

[![Syntology run record](https://syntology.ai/api/ran/2510.16079.svg)](https://syntology.ai/paper/2510.16079)

Connect an agent — have() is free.