SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2410.03577 · ICML · 2025

Look Twice Before You Answer: Memory-Space Visual Retracing for Hallucination Mitigation in Multimodal Large Language Models

Jia Liu, Yizhou Wang, Xuming Hu, Xin Zou, Yibo Yan, Yuanhuiyi Lyu, Kening Zheng, Sirui Huang, Junkai Chen, Peijie Jiang, Chang Tang

arXiv · PDF · Open in the Atlas

Code that ran

6 of 8 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 6 failures of our instrument; 2 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2410.03577.json · badge: /api/ran/2410.03577.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
1zhou-Wang/MemVR canonical (classifier label)
classifier basis: census
6 ran · 2 no recorded run
FunctionStatusWhere it lives
apply_rotary_pos_emb Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
1zhou-Wang/MemVR/modeling/modeling_chatglm.py
code served (permissive licence) · get_code("1cadbcf04f8bc964")
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 4, 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)

# Analyze the function signature and usage:
# - x: torch.Tensor with shape (b, np, sq, hn)
#   The function accesses x.size(0), x.size(1), x.size(2), x.size(3)
#   It slices x[..., :rot_dim] and x[..., rot_dim:]
#   rot_dim = rope_cache.shape[-2] * 2
#   So hn must be >= rot_dim
#
# - rope_cache: torch.Tensor with shape that allows:
#   rope_cache[:, :sq] slicing (has at least 2 dims)
#   rope_cache.shape[-2] exists and is used to compute rot_dim
#   rope_cache.view(-1, 1, sq, xshaped.size(3), 2) where xshaped.size(3) = rot_dim // 2

# Let's construct minimal tensors:
# batch size b = 2
# num_heads np = 3
# sequence_length sq = 4
# head_dim hn = 8

b = 2
np_heads = 3
sq = 4
hn = 8

# rope_cache needs shape where:
# - rope_cache.shape[-2] * 2 = rot_dim (should be <= hn)
# - rope_cache can be sliced as rope_cache[:, :sq]
# - rope_cache.view(-1, 1, sq, rot_dim//2, 2) works
# If rot_dim = 8, then rope_cache.shape[-2] = 4
# rope_cache should have shape compatible with [:, :sq] slicing
# and final view requires shape to have product = -1 * 1 * sq * 4 * 2 = -1 * 8 * sq

# A shape like (seq_len, hidden_dim, rot_dim//2, 2) = (4, ?, 4, 2) works
# Let's use (sq, 1, 4, 2) for rope_cache
rope_cache = torch.randn(sq, 1, 4, 2)

# x has shape (b, np, sq, hn)
x = torch.randn(b, np_heads, sq, hn)

ARGS = {
    "x": x,
    "rope_cache": rope_cache,
}
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 Tensor, ndim 4, dtype family float — no output shape is stated: observe_signature.py could not observe it (import:OSError)
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
collate_fn Ran
ran; our contract draft was wrong, not the code
1zhou-Wang/MemVR/eval/llava_model_vqa_loader.py
code served (permissive licence) · get_code("20e4f665698a3d18")
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

torch.manual_seed(0)

# The collate_fn expects a batch which is a list/iterable of tuples
# Each tuple contains (input_ids, image_tensor, image_size)
# The function unpacks these, stacks input_ids and image_tensors, and returns image_sizes as-is

# Create sample data items
item1 = (
    torch.randint(0, 100, (4,)),  # input_ids: small tensor
    torch.randn(3, 8, 8),          # image_tensor: 3-channel image, 8x8
    (8, 8)                         # image_size: tuple
)

item2 = (
    torch.randint(0, 100, (4,)),  # input_ids: same shape
    torch.randn(3, 8, 8),          # image_tensor: same shape
    (8, 8)                         # image_size: tuple
)

# Create batch as a list of tuples
batch = [item1, item2]

ARGS = {
    "batch": batch
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
collate_fn Ran
ran; our contract draft was wrong, not the code
1zhou-Wang/MemVR/eval/glm_model_vqa_loader.py
code served (permissive licence) · get_code("f5ff8b7846971cf9")
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

torch.manual_seed(0)

# The function expects batch to be an iterable of tuples with 4 elements each:
# (input_ids, image_tensors, image_sizes, images)
# It unpacks them with zip(*batch)
# It stacks input_ids tensors along dim=0
# image_tensors is set to None in the function
# image_sizes and images are kept as tuples

# Create sample data for a batch of 2 items
batch_item_1 = (
    torch.tensor([1, 2, 3, 4]),  # input_ids
    torch.randn(3, 4),  # image_tensors
    (224, 224),  # image_sizes
    "image_path_1.jpg"  # images
)

batch_item_2 = (
    torch.tensor([5, 6, 7, 8]),  # input_ids
    torch.randn(3, 4),  # image_tensors
    (256, 256),  # image_sizes
    "image_path_2.jpg"  # images
)

batch = [batch_item_1, batch_item_2]

ARGS = {
    "batch": batch
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_chunk Ran
ran; our contract draft was wrong, not the code
1zhou-Wang/MemVR/eval/glm_model_vqa.py
code served (permissive licence) · get_code("42a46570620cd9fa")
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) 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 math

def split_list(lst, n):
    """Split a list into n (roughly) equal-sized chunks"""
    chunk_size = math.ceil(len(lst) / n)
    return [lst[i:i + chunk_size] for i in range(0, len(lst), chunk_size)]

def get_chunk(lst, n, k):
    chunks = split_list(lst, n)
    return chunks[k]

# Setup code
lst = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
n = 3
k = 1

ARGS = {
    "lst": lst,
    "n": n,
    "k": k
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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 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
split_list Ran
ran; our contract draft was wrong, not the code
1zhou-Wang/MemVR/eval/glm_model_vqa.py
code served (permissive licence) · get_code("076c252c52cbb161")
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) and returned list. 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 math

def split_list(lst, n):
    """Split a list into n (roughly) equal-sized chunks"""
    chunk_size = math.ceil(len(lst) / n)
    return [lst[i:i + chunk_size] for i in range(0, len(lst), chunk_size)]

# Setup code
lst = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
n = 3

ARGS = {
    "lst": lst,
    "n": n
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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 list — a list 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
split_tensor_along_last_dim Ran
ran; our contract draft was wrong, not the code
1zhou-Wang/MemVR/modeling/modeling_chatglm.py
code served (permissive licence) · get_code("d7e378ee7899fbde")
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) and returned tuple, ndim 4, 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

torch.manual_seed(0)

# Create a small tensor with shape (2, 4, 8)
# Last dimension is 8, which is divisible by 2 and 4
tensor = torch.randn(2, 4, 8)

ARGS = {
    "tensor": tensor,
    "num_partitions": 2,
    "contiguous_split_chunks": False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 4, dtype family float — a tuple 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
default_init Not yet run
unverified (harvested, no recorded run)
1zhou-Wang/MemVR/modeling/modeling_chatglm.py
code served (permissive licence) · get_code("237019e36b9481d9")
How it ran

no invocation recorded.

forward Not yet run
unverified (harvested, no recorded run)
1zhou-Wang/MemVR/memvr.py
code served (permissive licence) · get_code("eb84abf07fedcf77")
How it ran

no invocation recorded.

Also present in other papers' code

Of 8 samples lifted from this paper's linked repositories, 8 carry a function identity and 0 do not. Of the 8 distinct functions with an identity, 6 are also present in other papers whose code we hold (6 with at least one paper dated earlier, 3 the same month, 6 later) and 2 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
split_list 197 other papers
109 dated earlier, 15 the same month, 66 later, 6 same year, order not determinable, 1 undated
2112.05682 (2021-12)
Self-attention Does Not Need $O(n^2)$ Memory
2306.17107 (2023-06)
LLaVAR: Enhanced Visual Instruction Tuning for Text-Rich Image Understanding
2307.10490 (2023-07)
Abusing Images and Sounds for Indirect Instruction Injection in Multi-Modal LLMs
and 194 more not named here
get_chunk 189 other papers
104 dated earlier, 15 the same month, 64 later, 6 same year, order not determinable
2112.05682 (2021-12)
Self-attention Does Not Need $O(n^2)$ Memory
2306.17107 (2023-06)
LLaVAR: Enhanced Visual Instruction Tuning for Text-Rich Image Understanding
2307.10490 (2023-07)
Abusing Images and Sounds for Indirect Instruction Injection in Multi-Modal LLMs
and 186 more not named here
collate_fn 54 other papers
21 dated earlier, 10 the same month, 23 later
2308.13566 (2023-08)
MLLM-DataEngine: An Iterative Refinement Approach for MLLM
2310.08588 (2023-10)
Octopus: Embodied Vision-Language Programmer from Environmental Feedback
2311.07362 (2023-11)
Volcano: Mitigating Multimodal Hallucination through Self-Feedback Guided Revision
and 51 more not named here
split_tensor_along_last_dim 8 other papers
6 dated earlier, 2 later
2307.13528 (2023-07)
FacTool: Factuality Detection in Generative AI -- A Tool Augmented Framework for Multi-Task and Multi-Domain Scenarios
2311.18445 (2023-11)
VTimeLLM: Empower LLM to Grasp Video Moments
2401.18058 (2024-01)
LongAlign: A Recipe for Long Context Alignment of Large Language Models
and 5 more not named here
default_init 6 other papers
5 dated earlier, 1 later
2307.13528 (2023-07)
FacTool: Factuality Detection in Generative AI -- A Tool Augmented Framework for Multi-Task and Multi-Domain Scenarios
2311.18445 (2023-11)
VTimeLLM: Empower LLM to Grasp Video Moments
2401.18058 (2024-01)
LongAlign: A Recipe for Long Context Alignment of Large Language Models
and 3 more not named here
apply_rotary_pos_emb 2 other papers
1 dated earlier, 1 later
2408.10943 (2024-08)
SysBench: Can Large Language Models Follow System Messages?
2502.20766 (2025-02)
FlexPrefill: A CONTEXT-AWARE SPARSE ATTENTION MECHANISM FOR EFFICIENT LONG-SEQUENCE IN-FERENCE

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

Despite their impressive capabilities, multimodal large language models (MLLMs) are prone to hallucinations, i.e., the generated content that is nonsensical or unfaithful to input sources. Unlike in LLMs, hallucinations in MLLMs often stem from the sensitivity of text decoder to visual tokens, leading to a phenomenon akin to "amnesia" about visual information. To address this issue, we propose MemVR, a novel decoding paradigm inspired by common cognition: when the memory of an image seen the moment before is forgotten, people will look at it again for factual answers. Following this principle, we treat visual tokens as supplementary evidence, re-injecting them into the MLLM through Feed Forward Network (FFN) as "key-value memory" at the middle trigger layer. This look-twice mechanism occurs when the model exhibits high uncertainty during inference, effectively enhancing factual alignment. Comprehensive experimental evaluations demonstrate that MemVR significantly mitigates hallucination across various MLLMs and excels in general benchmarks without incurring additional time overhead. The implementation is available from https://github.com/1zhou-Wang/MemVR.

For agents

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

get_harvested_code_for_paper("2410.03577")
get_code_for_paper("2410.03577")
have("2410.03577")

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

curl https://syntology.ai/api/ran/2410.03577.json

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

[![Syntology run record](https://syntology.ai/api/ran/2410.03577.svg)](https://syntology.ai/paper/2410.03577)

Connect an agent — have() is free.