SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2510.03215 · ICLR · 2025

Cache-to-Cache: Direct Semantic Communication Between Large Language Models

Yu Wang, Wanli Ouyang, Guohao Dai, Tianyu Fu, Zihan Min, Hanling Zhang, Jichao Yan

arXiv · PDF · Open in the Atlas

Code that ran

13 of 14 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; 3 honoured the contract we drafted, 1 violated it and 1 has no contract verdict; 8 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/2510.03215.json · badge: /api/ran/2510.03215.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
thu-nics/C2C canonical (classifier label)
classifier basis: census
13 ran · 1 no recorded run
FunctionStatusWhere it lives
all_sharers_mask Ran
ran, honoured the contract we drafted
thu-nics/C2C/rosetta/utils/core.py
code served (permissive licence) · get_code("3a17fd78f7eeb7d5")
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 int, ndim 0, dtype family int, 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
import numpy as np

# The function all_sharers_mask takes a single parameter: num_sharers (int)
# It computes a bitmask with all bits set for the given number of sharers
# For example: 3 sharers -> (1 << 3) - 1 = 8 - 1 = 7 (binary: 111)

ARGS = {
    "num_sharers": 3
}
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 int, ndim 0, dtype family int, 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
build_prompt Ran
ran; our contract draft was wrong, not the code
thu-nics/C2C/rosetta/utils/evaluate.py
code served (permissive licence) · get_code("472ca2ea3d9cecca")
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

ARGS = {
    "dataset": "mmmlu",
    "locale": "SW_KE",
    "question": "What is the capital of Kenya?",
    "choices": "A) Nairobi\nB) Mombasa\nC) Kisumu\nD) Nakuru",
    "use_cot": False,
    "use_template": True,
}
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
capture_init_args Ran
ran on a synthesized input
thu-nics/C2C/rosetta/train/dataset_adapters.py
code served (permissive licence) · get_code("3f66eae199cc340b")
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 type. 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 inspect

def capture_init_args(cls):
    """
    Decorator to capture initialization arguments of a dataset class.
    
    Args:
        cls: The class to decorate
        
    Returns:
        The decorated class with automatic init args capture
    """
    original_init = cls.__init__

    def new_init(self, *args, **kwargs):
        self._init_args = {}
        sig = inspect.signature(original_init)
        param_names = list(sig.parameters.keys())[1:]
        for i, arg in enumerate(args):
            if i < len(param_names):
                self._init_args[param_names[i]] = arg
        self._init_args.update(kwargs)
        original_init(self, *args, **kwargs)
    cls.__init__ = new_init
    return cls


class ExampleDatasetClass:
    """A simple dataset class to decorate"""
    def __init__(self, data, batch_size=32, shuffle=True, name="dataset"):
        self.data = data
        self.batch_size = batch_size
        self.shuffle = shuffle
        self.name = name


ARGS = {
    "cls": ExampleDatasetClass
}
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 type — a type output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a type 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
create_text_length_filter Ran
ran; our contract draft was wrong, not the code
thu-nics/C2C/rosetta/train/dataset_adapters.py
code served (permissive licence) · get_code("4e2d2967c4bfe0b1")
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 function. 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 Any, Dict, Callable

torch.manual_seed(0)

# Define a minimal tokenizer mock that has the required methods
class MockTokenizer:
    def __init__(self):
        self.vocab_size = 1000
    
    def __call__(self, texts, add_special_tokens=False):
        # Return a dict with 'input_ids' key containing token lists
        result = {'input_ids': []}
        for text in texts:
            # Simple mock tokenization: split by spaces and map to token IDs
            tokens = text.split()
            token_ids = [hash(t) % self.vocab_size for t in tokens]
            result['input_ids'].append(token_ids)
        return result
    
    def apply_chat_template(self, chat_list, tokenize=False, add_generation_prompt=False):
        # Simple mock: join chat messages
        result = " ".join([str(item) for item in chat_list])
        return result

# Define a text extractor function
def text_extractor(sample: Dict[str, Any]) -> str:
    """Extract text from a sample dict"""
    if 'text' in sample:
        return sample['text']
    elif 'content' in sample:
        return sample['content']
    return ""

# Set up arguments for the function
ARGS = {
    'max_length': 100,
    'text_extractor': text_extractor,
    'tokenizer': MockTokenizer(),
    'use_tokens': 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 function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function 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
extract_answer_from_content Ran
ran; our contract draft was wrong, not the code
thu-nics/C2C/rosetta/utils/evaluate.py
code served (permissive licence) · get_code("487b0fcf699da856")
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
from typing import Dict, Any, List, Tuple, Optional

# The function takes a single parameter: text (str)
# Let's create several test cases to demonstrate the function works

# Test case 1: Simple answer with "Answer:" prefix
text_1 = "Answer: A"

# Test case 2: Answer with multiple formats
text_2 = "The answer is B"

# Test case 3: Swahili format
text_3 = "Jibu: C"

# Test case 4: Standalone letter at end
text_4 = "The correct choice is D."

# Test case 5: Letter in middle of text
text_5 = "Based on the analysis, I choose B"

# Test case 6: Mathematical content without answer indicator (should return None)
text_6 = "x + y = 2x - 5, solve for A"

# Test case 7: Mathematical content with answer indicator
text_7 = "x + y = 2x - 5, the answer is C"

# Test case 8: Multiple letters, should return last one
text_8 = "Options are A, B, C, and D. I select B"

# Test case 9: Empty or no valid answer
text_9 = "This is just random text with no answer"

# Test case 10: Uppercase ANSWER prefix
text_10 = "ANSWER: D"

# For the function, we only need to provide the 'text' parameter
ARGS = {
    "text": "Answer: A"
}
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
hybrid_to_dynamic Ran
ran, honoured the contract we drafted
thu-nics/C2C/rosetta/model/wrapper.py
code served (permissive licence) · get_code("7e4691822f4b29b8")
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 DynamicCache. 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 transformers.cache_utils import DynamicCache

torch.manual_seed(0)

# Create a mock hybrid cache object that has key_cache and value_cache attributes
class MockHybridCache:
    def __init__(self):
        # Simulate a cache with 2 layers
        # Each layer has key and value tensors
        # Shape: (batch_size, num_heads, seq_len, head_dim)
        self.key_cache = [
            torch.randn(2, 8, 4, 3),  # layer 0
            torch.randn(2, 8, 4, 3),  # layer 1
        ]
        self.value_cache = [
            torch.randn(2, 8, 4, 3),  # layer 0
            torch.randn(2, 8, 4, 3),  # layer 1
        ]

hybrid_cache_obj = MockHybridCache()

ARGS = {
    "hybrid_cache": hybrid_cache_obj
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type DynamicCache — a DynamicCache output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a DynamicCache output the recorded ndim 6 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
k_nearest_sources Ran
ran; our contract draft was wrong, not the code
thu-nics/C2C/rosetta/train/model_utils.py
code served (permissive licence) · get_code("7560f1f6daf446f2")
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

torch.manual_seed(0)

ARGS = {
    "num_target_layers": 4,
    "num_source_layers": 6,
    "k": 3,
}
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
last_aligned_sources Ran
ran; our contract draft was wrong, not the code
thu-nics/C2C/rosetta/train/model_utils.py
code served (permissive licence) · get_code("22b2f2f6f8147c86")
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
from typing import Dict, Any, List

torch.manual_seed(0)

# Based on the function signature, we need:
# - num_target_layers: int
# - num_source_layers: int
# - k: int (optional, defaults to 1)

ARGS = {
    "num_target_layers": 11,
    "num_source_layers": 33,
    "k": 1
}
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
mask_to_sharers Ran
ran, honoured the contract we drafted
thu-nics/C2C/rosetta/utils/core.py
code served (permissive licence) · get_code("7a7ba05d0e27c078")
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

# Test case 1: mask = 1 (binary: 001)
# Expected: [1]
test_mask_1 = 1

# Test case 2: mask = 3 (binary: 011)
# Expected: [1, 2]
test_mask_2 = 3

# Test case 3: mask = 5 (binary: 101)
# Expected: [1, 3]
test_mask_3 = 5

# Test case 4: mask = 7 (binary: 111)
# Expected: [1, 2, 3]
test_mask_4 = 7

# Test case 5: mask = 0 (edge case)
# Expected: []
test_mask_5 = 0

# Test case 6: mask = 15 (binary: 1111)
# Expected: [1, 2, 3, 4]
test_mask_6 = 15

# Main argument for the function
ARGS = {
    "mask": 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 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
parse_answer Ran
ran; our contract draft was wrong, not the code
thu-nics/C2C/rosetta/utils/evaluate.py
code served (permissive licence) · get_code("6a9ab19801c22746")
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.

import torch
import numpy as np

torch.manual_seed(0)

ARGS = {
    "answer_str": "0213"
}
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 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_dataset Ran
ran; our contract draft was wrong, not the code
thu-nics/C2C/rosetta/train/dataset_adapters.py
code served (permissive licence) · get_code("a335ebc9764afbf4")
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 type. 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

# Clear the registry for clean execution
DATASET_REGISTRY = {}

# Define a minimal dataset class for testing
class SimpleDataset:
    """Minimal dataset class for testing the register_dataset decorator."""
    def __init__(self):
        self.data = torch.randn(2, 4)
    
    def __len__(self):
        return 2
    
    def __getitem__(self, idx):
        return self.data[idx]

# Test case 1: Using as decorator without arguments (cls passed directly)
test_cls_1 = SimpleDataset

# Test case 2: Using as decorator with custom name
test_cls_2 = SimpleDataset

# Test case 3: Using with name argument
custom_name = "CustomDataset"

# Construct ARGS for the function
# The function signature is: register_dataset(cls=None, name=None)
# We'll test the case where cls is provided without name (most common decorator usage)
ARGS = {
    "cls": SimpleDataset,
    "name": None
}
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 type — a type output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a type 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
sample_token Ran
ran; our contract draft was wrong, not the code
thu-nics/C2C/rosetta/model/sampling.py
code served (permissive licence) · get_code("687e2c81ca6fa045")
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 Tensor, ndim 1, dtype family int, shape (2,) 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 torch.nn.functional as F

torch.manual_seed(0)

# Create a small logits tensor for testing
# Shape: [batch_size=2, vocab_size=4]
logits = torch.randn(2, 4)

# Define arguments for sample_token function
ARGS = {
    "logits": logits,
    "temperature": 1.0,
    "top_p": 1.0,
    "top_k": -1
}
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 Tensor, ndim 1, dtype family int, shape (2,) — shape (2,) 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
sharers_to_mask Ran
ran, violated the contract we drafted
thu-nics/C2C/rosetta/utils/core.py
code served (permissive licence) · get_code("f1bff5c661343925")
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 int, ndim 0, dtype family int, 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
from typing import List

torch.manual_seed(0)

# Test case 1: Single sharer
test1_sharer_indices = [1]

# Test case 2: Multiple sharers
test2_sharer_indices = [1, 2, 3]

# Test case 3: Non-contiguous sharers
test3_sharer_indices = [1, 3]

# Test case 4: Empty list (edge case)
test4_sharer_indices = []

# Use one of the test cases for ARGS
ARGS = {
    "sharer_indices": [1, 2, 3]
}
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 int, ndim 0, dtype family int, 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
clone_kv_cache Not yet run
unverified (harvested, no recorded run)
thu-nics/C2C/rosetta/model/wrapper.py
code served (permissive licence) · get_code("4b5ab2ca4513e83b")
How it ran

no invocation recorded.

Also present in other papers' code

Of 14 samples lifted from this paper's linked repositories, 14 carry a function identity and 0 do not. Of the 14 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 14 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

Multi-LLM systems harness the complementary strengths of diverse Large Language Models, achieving performance and efficiency gains that are not attainable by a single model. In existing designs, LLMs communicate through text, forcing internal representations to be transformed into output token sequences. This process both loses rich semantic information and incurs token-by-token generation latency. Motivated by these limitations, we ask: Can LLMs communicate beyond text? Oracle experiments show that enriching the KV-Cache semantics can improve response quality without increasing cache size, supporting KV-Cache as an effective medium for inter-model communication. Thus, we propose Cache-to-Cache (C2C), a new paradigm for direct semantic communication between LLMs. C2C uses a neural network to project and fuse the source model's KV-cache with that of the target model to enable direct semantic transfer. A learnable gating mechanism selects the target layers that benefit from cache communication. Compared with text communication, C2C utilizes the deep, specialized semantics from both models, while avoiding explicit intermediate text generation. Experiments show that C2C achieves 6.4-14.2% higher average accuracy than individual models. It further outperforms the text communication paradigm by approximately 3.1-5.4%, while delivering an average 2.5× speedup in latency. Our code is available at https://github.com/thu-nics/C2C.

For agents

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

get_harvested_code_for_paper("2510.03215")
get_code_for_paper("2510.03215")
have("2510.03215")

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

curl https://syntology.ai/api/ran/2510.03215.json

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

[![Syntology run record](https://syntology.ai/api/ran/2510.03215.svg)](https://syntology.ai/paper/2510.03215)

Connect an agent — have() is free.