SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2602.06409 · ICML · 2026

VENOMREC: Cross-Modal Interactive Poisoning for Targeted Promotion in Multimodal LLM Recommender Systems

Wei Yang, Bryan Lim, Jiaming Zhang, Longtao Huang, Cyril Leung, Tianxiang Chen, Guowei Guan, Yurong Hao, Tiantong Wu, Fuyao Zhang

arXiv · PDF · Open in the Atlas

Code that ran

19 of 20 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (5 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 2 violated it and 5 have no contract verdict; 11 failures of our instrument; 1 no recorded run), measured 2026-11-21; 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/2602.06409.json · badge: /api/ran/2602.06409.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
GuoweiGuan666/VenomRec community
Papers with Code archive
19 ran · 1 no recorded run
FunctionStatusWhere it lives
AdaptiveBudgetController Ran
ran on a synthesized input · constructed an object; no result was computed
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("0c34cb25c626a7ee")
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, AdaptiveBudgetController: 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 dataclasses import dataclass, field
from typing import Dict, List, Optional

# Set seed for reproducibility
torch.manual_seed(0)

# Create valid arguments for AdaptiveBudgetController
ARGS = {
    'img_eps': 0.05,
    'txt_ratio': 0.15,
    'embed_eps': 0.01,
    'min_img_eps': 0.005,
    'max_img_eps': 0.2,
    'min_txt_ratio': 0.05,
    'max_txt_ratio': 0.4,
    'min_embed_eps': 0.0,
    'max_embed_eps': 0.05,
    'up_factor': 1.15,
    'down_factor': 0.85,
    'history': [],
    'prev_info_nce': None,
    'prev_sim': None,
    'prev_alignment': None,
    'prev_bce': None,
    'prev_hit': None,
    'bce_tolerance': 0.001,
    'sim_tolerance': 0.0001,
    'align_tolerance': 0.0001,
    'info_tolerance': 0.0001,
    'hit_tolerance': 0.001,
}
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 AdaptiveBudgetController — 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 AdaptiveBudgetController 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
ImagePerturber Ran
ran on a synthesized input · constructed an object; no result was computed
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("f62a318737248f50")
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, ImagePerturber: 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 numpy as np

ARGS = {
    'eps': 0.1,
    'iters': 3,
    'psnr_min': 30.0,
    'strategy': 'cosine'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ImagePerturber — 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 ImagePerturber 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
PerturbationResult Ran
ran on a synthesized input · constructed an object; no result was computed
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("2f01dc3fa37414ad")
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, PerturbationResult: 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 torch.nn as nn
from dataclasses import dataclass
from typing import List

torch.manual_seed(0)

@dataclass
class PerturbationRound:
    """Record collected after each interaction round."""
    round_idx: int
    sim_after_image: float
    sim_target: float
    dist_target: float
    img_eps_used: float
    txt_ratio: float
    img_budget: float
    txt_budget: float
    info_nce: float
    psnr_img: float
    bce_loss: float
    hit_prob: float
    txt_embed_delta_norm: float | None = None
    stop_reason: str | None = None

@dataclass
class PerturbationResult:
    """Summary returned for every poisoned target."""
    target_id: str
    rounds: List[PerturbationRound]
    final_image: List[float]
    final_text: str

# Create sample PerturbationRound objects
round_1 = PerturbationRound(
    round_idx=0,
    sim_after_image=0.95,
    sim_target=0.87,
    dist_target=0.15,
    img_eps_used=0.05,
    txt_ratio=0.3,
    img_budget=1.0,
    txt_budget=0.5,
    info_nce=0.12,
    psnr_img=28.5,
    bce_loss=0.08,
    hit_prob=0.92,
    txt_embed_delta_norm=0.04,
    stop_reason=None
)

round_2 = PerturbationRound(
    round_idx=1,
    sim_after_image=0.97,
    sim_target=0.91,
    dist_target=0.10,
    img_eps_used=0.08,
    txt_ratio=0.35,
    img_budget=0.92,
    txt_budget=0.42,
    info_nce=0.09,
    psnr_img=30.2,
    bce_loss=0.05,
    hit_prob=0.96,
    txt_embed_delta_norm=0.02,
    stop_reason="budget_exhausted"
)

# Create sample final image as a list of floats
final_image = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8]

# Create sample final text
final_text = "perturbed text representation"

# Create the target_id
target_id = "target_001"

ARGS = {
    "target_id": target_id,
    "rounds": [round_1, round_2],
    "final_image": final_image,
    "final_text": final_text
}
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 PerturbationResult — 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 PerturbationResult 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
PerturbationRound Ran
ran on a synthesized input · constructed an object; no result was computed
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("a42303b353d7d5eb")
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, PerturbationRound: 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 dataclasses import dataclass

torch.manual_seed(0)

ARGS = {
    "round_idx": 5,
    "sim_after_image": 0.8234,
    "sim_target": 0.9102,
    "dist_target": 0.1543,
    "img_eps_used": 0.0314,
    "txt_ratio": 0.45,
    "img_budget": 1.0,
    "txt_budget": 1.0,
    "info_nce": 0.2891,
    "psnr_img": 28.45,
    "bce_loss": 0.3421,
    "hit_prob": 0.7654,
    "txt_embed_delta_norm": 0.0521,
    "stop_reason": "max_iterations",
}
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 PerturbationRound — 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 PerturbationRound 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
TextPerturber Ran
ran on a synthesized input · constructed an object; no result was computed
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("6bf57cee2c1d71a5")
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, TextPerturber: 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 numpy as np
from typing import Dict, Any, List

# Create a minimal tokenizer stand-in
class MinimalTokenizer:
    def encode(self, text: str) -> List[int]:
        """Simple tokenizer that splits on whitespace and returns token ids."""
        words = text.split()
        return list(range(len(words)))
    
    def decode(self, token_ids: List[int]) -> str:
        """Decode token ids back to string."""
        return " ".join(str(tid) for tid in token_ids)

# Define a simple domain synonyms dict (the function references DOMAIN_SYNONYMS)
DOMAIN_SYNONYMS = {
    'good': ['great', 'excellent', 'wonderful'],
    'bad': ['poor', 'terrible', 'awful'],
    'product': ['item', 'thing', 'object'],
}

# Build minimal arguments
tokenizer = MinimalTokenizer()
text = "this is a good product and bad thing"
tokens = tokenizer.encode(text)

# Create mask - True for replaceable tokens
mask = [False, False, False, True, True, False, True, False]

# Create embeddings (num_tokens x embedding_dim)
token_embeddings = np.random.randn(len(tokens), 4).astype(np.float32)

# Simple score function that returns a score
def dummy_score_fn(candidate: str, original: str) -> float:
    return 0.5

# Keywords can be a dict or iterable
keywords_dict = {'good': 'great', 'bad': 'poor', 'product': 'item'}

ARGS = {
    'ratio': 0.3,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type TextPerturber — 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 TextPerturber 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
_cosine_aligned_step Ran
ran; our contract draft was wrong, not the code
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("9219366ca8d18fe3")
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 numpy as np
import math

# Construct arguments for _cosine_aligned_step
x_feat = [1.0, 2.0, 3.0, 4.0]
target_feat = [1.5, 2.5, 3.5, 4.5]
mask = [True, True, False, True]
eps = 0.5

ARGS = {
    'x_feat': x_feat,
    'target_feat': target_feat,
    'mask': mask,
    'eps': eps,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — 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
_count_tokens Ran
ran, honoured the contract we drafted
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("b5c0d9f50eef2dc1")
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 torch.nn as nn

torch.manual_seed(0)

# Create a minimal adapter class with a tokenizer
class MinimalTokenizer:
    def encode(self, text):
        # Simple tokenizer that splits and returns token list
        return text.split()
    
    def __call__(self, text, return_tensors=None):
        # Return format compatible with the function
        tokens = text.split()
        if return_tensors is None:
            return {'input_ids': tokens}
        return {'input_ids': tokens}

class MinimalAdapter:
    def __init__(self):
        self.tokenizer = MinimalTokenizer()

# Test case 1: adapter with a tokenizer
adapter1 = MinimalAdapter()
text1 = "hello world test"

ARGS = {
    "adapter": adapter1,
    "text": text1
}
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
_embedding_delta Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("278036de1273f106")
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 2, 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
import logging

# Set seed for reproducibility
np.random.seed(0)

# Define test arguments for _embedding_delta
ARGS = {
    "embeddings": [
        [0.1, 0.2, 0.3, 0.4],
        [0.5, 0.6, 0.7, 0.8],
        [0.9, 1.0, 1.1, 1.2],
        [1.3, 1.4, 1.5, 1.6]
    ],
    "mask": [True, False, True, False],
    "embed_eps": 0.05
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 2, 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
_ensure_tokens Ran
ran, violated the contract we drafted
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("e0d9491e3a23066d")
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

class MockTokenizer:
    """Stand-in tokenizer with encode method"""
    def encode(self, text: str):
        # Simple mock: split text and return list of token ids
        return [int(ord(c)) for c in text.split()]

# Test case 1: tokens provided (should ignore tokenizer and text)
tokens_case = [101, 102, 103, 104]

# Test case 2: tokens is None, need tokenizer and text
tokenizer = MockTokenizer()
text = "hello world"

ARGS = {
    "tokens": tokens_case,
    "tokenizer": tokenizer,
    "text": text,
}
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
_strip_markers Ran
ran; our contract draft was wrong, not the code
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("5a99df7d61dff180")
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

SUBWORD_MARKERS = {'▁', '##'}

def _strip_markers(tok: str) -> str:
    t = tok
    for m in SUBWORD_MARKERS:
        t = t.replace(m, '')
    return t

# The function takes a single string parameter 'tok' and removes subword markers from it
# Test with a token that contains markers
ARGS = {
    'tok': '▁hello##world'
}
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
_token_to_str Ran
ran; our contract draft was wrong, not the code
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("f0b2b02805bc76b3")
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 torch.nn as nn

torch.manual_seed(0)

# The function uses:
# - tokenizer.decode([tok_id]) method (if it has decode attribute)
# - tokenizer can be None
# - tok_id needs to be convertible to string

class MockTokenizer:
    """Minimal tokenizer with decode method"""
    def decode(self, token_ids):
        return f"decoded_{'_'.join(str(t) for t in token_ids)}"

ARGS = {
    "tokenizer": MockTokenizer(),
    "tok_id": 42
}
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
_topk_mask Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("0354dc50f89eccae")
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 math
import numpy as np

# Setup for _topk_mask function
# The function takes:
# - values: np.ndarray (any shape)
# - ratio: float (between 0.0 and 1.0)

# Create test data
values = np.array([[5.2, 1.3, 8.9, 2.1],
                   [3.4, 7.6, 0.5, 9.2]])
ratio = 0.5

ARGS = {
    "values": values,
    "ratio": ratio
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — 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
compute_cosine Ran
ran; our contract draft was wrong, not the code
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("7ffc28aa3dfef9f7")
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 float, 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

# compute_cosine takes two iterables of floats and computes cosine similarity
# It converts them to numpy arrays, computes norms, and returns the dot product divided by the product of norms

# Create two simple iterables of floats
a = [1.0, 2.0, 3.0, 4.0]
b = [5.0, 6.0, 7.0, 8.0]

ARGS = {
    "a": a,
    "b": b,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, 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
cosine_similarity Ran
ran; our contract draft was wrong, not the code
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("601d2b794f6f8f1f")
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 float, 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

# Construct arguments for cosine_similarity function
vec_a = [1.0, 2.0, 3.0, 4.0]
vec_b = [5.0, 6.0, 7.0, 8.0]

ARGS = {
    "a": vec_a,
    "b": vec_b,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, 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
get_visual_text_masks Ran
ran; our contract draft was wrong, not the code
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("919a02925f1b0f3f")
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 numpy as np
import math

class MockAdapter:
    """Mock adapter that computes cross-modal saliency."""
    def compute_cross_modal_saliency(self, image_feat, text):
        """Return a dict with visual_scores and text_scores."""
        # Return scores matching the expected dimensions
        visual_scores = np.array([0.5, 0.3, 0.8, 0.2], dtype=float)
        text_scores = np.array([0.7, 0.4, 0.6], dtype=float)
        return {
            'visual_scores': visual_scores,
            'text_scores': text_scores
        }

# Create test data
adapter = MockAdapter()
image_feat = np.array([0.1, 0.2, 0.3, 0.4], dtype=float)  # 4 visual tokens
text = "hello world python"  # 3 tokens when split
token_count = 3
pop_vec = [0.1, 0.2, 0.3]

ARGS = {
    'adapter': adapter,
    'image_feat': image_feat,
    'text': text,
    'vis_ratio': 0.15,
    'txt_ratio': 0.15,
    'token_count': token_count,
    'pop_vec': pop_vec,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — 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
info_nce_loss Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("b84770e3b3129208")
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 float, 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

# Set seed for reproducibility
np.random.seed(0)

# Create small distinct vectors for testing
target = [0.1, 0.2, 0.3, 0.4]
positive = [0.15, 0.25, 0.35, 0.45]
negatives = [
    [0.9, 0.8, 0.7, 0.6],
    [0.5, 0.4, 0.3, 0.2],
    [0.05, 0.1, 0.15, 0.2]
]
temperature = 0.07

ARGS = {
    "target": target,
    "positive": positive,
    "negatives": negatives,
    "temperature": temperature
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, 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_replaceable_token Ran
ran, violated the contract we drafted
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("110f179872f95a82")
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.

# Setup code to construct valid arguments for is_replaceable_token

ARGS = {
    'tok': '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 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
masked_pgd_image Ran
ran; our contract draft was wrong, not the code
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("b5da9bff43fcd7f8")
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 numpy as np
import math
import logging

# Set up logging
logging.basicConfig(level=logging.DEBUG)

# Create small test data
torch.manual_seed(0)

# x_img_or_feat: Sequence[float] - original image or feature vector
x_img_or_feat = [1.0, 2.0, 3.0, 4.0, 5.0]

# mask: Sequence[bool] - which elements can be perturbed
mask = [True, False, True, False, True]

# target_feat: Sequence[float] - target feature vector to match
target_feat = [1.5, 2.2, 3.5, 4.1, 5.3]

# eps: float - maximum perturbation magnitude
eps = 0.5

# iters: int - number of iterations
iters = 10

# psnr_min: float - minimum PSNR threshold for early stopping
psnr_min = 20.0

# peak: Optional[float] - peak value for PSNR calculation (data range)
peak = 10.0

ARGS = {
    "x_img_or_feat": x_img_or_feat,
    "mask": mask,
    "target_feat": target_feat,
    "eps": eps,
    "iters": iters,
    "psnr_min": psnr_min,
    "peak": peak,
}
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 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
normalize_token Ran
ran; our contract draft was wrong, not the code
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("f9c1dce0553ff207")
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 = {
    "tok": "#Hello_World#"
}
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
interactive_perturb_target Not yet run
unverified (harvested, no recorded run)
GuoweiGuan666/VenomRec/attack/ours/dcip_ieos/interactive.py
code served (permissive licence) · get_code("d7bad37ca3ef4af1")
How it ran

no invocation recorded.

Also present in other papers' code

Of 20 samples lifted from this paper's linked repositories, 20 carry a function identity and 0 do not. Of the 20 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 20 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 2026-11-21.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Multimodal large language models (MLLMs) are pushing recommender systems (RecSys) toward content-grounded retrieval and ranking via cross-modal fusion. We find that while crossmodal consensus often mitigates conventional poisoning that manipulates interaction logs or perturbs a single modality, it also introduces a new attack surface where synchronised multimodal poisoning can reliably steer fused representations along stable semantic directions during fine-tuning. To characterise this threat, we formalise cross-modal interactive poisoning and propose VENOMREC, which performs Exposure Alignment to identify high-exposure regions in the joint embedding space and Cross-modal Interactive Perturbation to craft attention-guided coupled token-patch edits. Experiments on four realworld multimodal datasets demonstrate that VEN-OMREC consistently outperforms strong baselines, achieving 0.73 mean ER@20 and improving over the strongest baseline by +0.52 absolute ER points on average, while maintaining comparable recommendation utility. Code is available at https://github.com/GuoweiGuan666/VenomRec.

For agents

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

get_harvested_code_for_paper("2602.06409")
get_code_for_paper("2602.06409")
have("2602.06409")

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

curl https://syntology.ai/api/ran/2602.06409.json

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

[![Syntology run record](https://syntology.ai/api/ran/2602.06409.svg)](https://syntology.ai/paper/2602.06409)

Connect an agent — have() is free.