SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2603.01169 · 2026

TripleSumm: Adaptive Triple-Modality Fusion for Video Summarization

Joonseok Lee, Yejin Kim, Hyemin Jeong, Mingu Kang, Sumin Kim, Yoori Oh

arXiv · PDF · Open in the Atlas

Code that ran

8 of 10 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 8 have no contract verdict; 0 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/2603.01169.json · badge: /api/ran/2603.01169.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
smkim37/TripleSumm community
Papers with Code archive
8 ran · 2 no recorded run
FunctionStatusWhere it lives
CrossModalAttention Ran
ran on a synthesized input
smkim37/TripleSumm/models/model.py
code served (permissive licence) · get_code("5b3de883fca30aef")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) 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 torch.nn as nn

class CrossModalAttention(nn.Module):

    def __init__(self, input_dim, num_heads, dropout):
        super().__init__()
        assert input_dim % num_heads == 0, 'input_dim must be divisible by num_heads'
        self.input_dim = input_dim
        self.num_heads = num_heads
        self.head_dim = input_dim // num_heads
        self.scale = self.head_dim ** (-0.5)
        self.q_proj = nn.Linear(input_dim, input_dim, bias=False)
        self.k_proj = nn.Linear(input_dim, input_dim, bias=False)
        self.v_proj = nn.Linear(input_dim, input_dim, bias=False)
        self.out_proj = nn.Linear(input_dim, input_dim, bias=False)
        self.dropout = nn.Dropout(dropout)

    def forward(self, query, context):
        B_q, T_q, _ = query.shape
        B_c, T_c, _ = context.shape
        q = self.q_proj(query).reshape(B_q, T_q, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
        k = self.k_proj(context).reshape(B_c, T_c, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
        v = self.v_proj(context).reshape(B_c, T_c, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
        scores = q @ k.transpose(-2, -1) * self.scale
        attn_probs = torch.nn.functional.softmax(scores, dim=-1)
        attn_weights = attn_probs
        attn_probs = self.dropout(attn_probs)
        out = attn_probs @ v
        out = out.transpose(1, 2).reshape(B_q, T_q, self.input_dim)
        out = self.out_proj(out)
        return (out, attn_weights)


INIT_ARGS = {
    'input_dim': 4,
    'num_heads': 2,
    'dropout': 0.1,
}

ARGS = {
    'query': torch.randn(2, 3, 4),
    'context': torch.randn(2, 5, 4),
}

if __name__ == '__main__':
    model = CrossModalAttention(**INIT_ARGS)
    output, attn_weights = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Attention weights shape: {attn_weights.shape}")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
CrossModalAttentionLayer Ran
ran on a synthesized input
smkim37/TripleSumm/models/model.py
code served (permissive licence) · get_code("336c9365fc0d1a0c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) 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 torch.nn as nn
import torch.nn.functional as F

class CrossModalAttention(nn.Module):

    def __init__(self, input_dim, num_heads, dropout):
        super().__init__()
        assert input_dim % num_heads == 0, 'input_dim must be divisible by num_heads'
        self.input_dim = input_dim
        self.num_heads = num_heads
        self.head_dim = input_dim // num_heads
        self.scale = self.head_dim ** (-0.5)
        self.q_proj = nn.Linear(input_dim, input_dim, bias=False)
        self.k_proj = nn.Linear(input_dim, input_dim, bias=False)
        self.v_proj = nn.Linear(input_dim, input_dim, bias=False)
        self.out_proj = nn.Linear(input_dim, input_dim, bias=False)
        self.dropout = nn.Dropout(dropout)

    def forward(self, query, context):
        B_q, T_q, _ = query.shape
        B_c, T_c, _ = context.shape
        q = self.q_proj(query).reshape(B_q, T_q, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
        k = self.k_proj(context).reshape(B_c, T_c, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
        v = self.v_proj(context).reshape(B_c, T_c, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
        scores = q @ k.transpose(-2, -1) * self.scale
        attn_probs = F.softmax(scores, dim=-1)
        attn_weights = attn_probs
        attn_probs = self.dropout(attn_probs)
        out = attn_probs @ v
        out = out.transpose(1, 2).reshape(B_q, T_q, self.input_dim)
        out = self.out_proj(out)
        return (out, attn_weights)

class SwiGLU(nn.Module):

    def __init__(self, input_dim, dropout):
        super().__init__()
        hidden_dim = int(input_dim * 2 * (2 / 3))
        self.w1 = nn.Linear(input_dim, hidden_dim, bias=False)
        self.w2 = nn.Linear(input_dim, hidden_dim, bias=False)
        self.w3 = nn.Linear(hidden_dim, input_dim, bias=False)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        x = F.silu(self.w1(x)) * self.w2(x)
        x = self.w3(x)
        x = self.dropout(x)
        return x

class CrossModalAttentionLayer(nn.Module):

    def __init__(self, input_dim, num_heads, dropout):
        super().__init__()
        self.cma = CrossModalAttention(input_dim=input_dim, num_heads=num_heads, dropout=dropout)
        self.ffn = SwiGLU(input_dim=input_dim, dropout=dropout)
        self.norm_q = nn.LayerNorm(input_dim)
        self.norm_c = nn.LayerNorm(input_dim)
        self.norm2 = nn.LayerNorm(input_dim)
        self.dropout = nn.Dropout(dropout)

    def forward(self, query, context):
        attn_out, attn_weights = self.cma(self.norm_q(query), self.norm_c(context))
        query = query + self.dropout(attn_out)
        ffn_out = self.ffn(self.norm2(query))
        query = query + self.dropout(ffn_out)
        return (query, attn_weights)

INIT_ARGS = {
    'input_dim': 4,
    'num_heads': 2,
    'dropout': 0.1,
}

ARGS = {
    'query': torch.randn(2, 3, 4),
    'context': torch.randn(2, 5, 4),
}

if __name__ == '__main__':
    model = CrossModalAttentionLayer(**INIT_ARGS)
    output, attn_weights = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Attention weights shape: {attn_weights.shape}")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
CrossModalFusionBlock Ran
ran on a synthesized input
smkim37/TripleSumm/models/model.py
code served (permissive licence) · get_code("ba5914b0fa6f6b9f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) 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 torch.nn as nn
import torch.nn.functional as F

class SwiGLU(nn.Module):
    def __init__(self, input_dim, dropout):
        super().__init__()
        hidden_dim = int(input_dim * 2 * (2 / 3))
        self.w1 = nn.Linear(input_dim, hidden_dim, bias=False)
        self.w2 = nn.Linear(input_dim, hidden_dim, bias=False)
        self.w3 = nn.Linear(hidden_dim, input_dim, bias=False)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        x = F.silu(self.w1(x)) * self.w2(x)
        x = self.w3(x)
        x = self.dropout(x)
        return x

class CrossModalAttention(nn.Module):
    def __init__(self, input_dim, num_heads, dropout):
        super().__init__()
        assert input_dim % num_heads == 0, 'input_dim must be divisible by num_heads'
        self.input_dim = input_dim
        self.num_heads = num_heads
        self.head_dim = input_dim // num_heads
        self.scale = self.head_dim ** (-0.5)
        self.q_proj = nn.Linear(input_dim, input_dim, bias=False)
        self.k_proj = nn.Linear(input_dim, input_dim, bias=False)
        self.v_proj = nn.Linear(input_dim, input_dim, bias=False)
        self.out_proj = nn.Linear(input_dim, input_dim, bias=False)
        self.dropout = nn.Dropout(dropout)

    def forward(self, query, context):
        B_q, T_q, _ = query.shape
        B_c, T_c, _ = context.shape
        q = self.q_proj(query).reshape(B_q, T_q, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
        k = self.k_proj(context).reshape(B_c, T_c, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
        v = self.v_proj(context).reshape(B_c, T_c, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
        scores = q @ k.transpose(-2, -1) * self.scale
        attn_probs = F.softmax(scores, dim=-1)
        attn_weights = attn_probs
        attn_probs = self.dropout(attn_probs)
        out = attn_probs @ v
        out = out.transpose(1, 2).reshape(B_q, T_q, self.input_dim)
        out = self.out_proj(out)
        return (out, attn_weights)

class CrossModalAttentionLayer(nn.Module):
    def __init__(self, input_dim, num_heads, dropout):
        super().__init__()
        self.cma = CrossModalAttention(input_dim=input_dim, num_heads=num_heads, dropout=dropout)
        self.ffn = SwiGLU(input_dim=input_dim, dropout=dropout)
        self.norm_q = nn.LayerNorm(input_dim)
        self.norm_c = nn.LayerNorm(input_dim)
        self.norm2 = nn.LayerNorm(input_dim)
        self.dropout = nn.Dropout(dropout)

    def forward(self, query, context):
        attn_out, attn_weights = self.cma(self.norm_q(query), self.norm_c(context))
        query = query + self.dropout(attn_out)
        ffn_out = self.ffn(self.norm2(query))
        query = query + self.dropout(ffn_out)
        return (query, attn_weights)

class CrossModalFusionBlock(nn.Module):
    def __init__(self, input_dim, num_heads, dropout):
        super().__init__()
        self.num_heads = num_heads
        self.cma_layer = CrossModalAttentionLayer(input_dim=input_dim, num_heads=num_heads, dropout=dropout)

    def forward(self, fusion, visual, text, audio):
        query = fusion
        context = torch.stack((visual, text, audio), dim=2)
        B, T, D = query.shape
        _, _, N, _ = context.shape
        query_reshaped = query.reshape(B * T, 1, D)
        context_reshaped = context.reshape(B * T, N, D)
        query_update, attn_weights = self.cma_layer(query_reshaped, context_reshaped)
        fusion_update = query_update.reshape(B, T, D)
        attn_weights = attn_weights.squeeze(2).view(B, T, self.num_heads, N)
        return (fusion_update, attn_weights)

INIT_ARGS = {
    'input_dim': 4,
    'num_heads': 2,
    'dropout': 0.1,
}

ARGS = {
    'fusion': torch.randn(2, 3, 4),
    'visual': torch.randn(2, 3, 4),
    'text': torch.randn(2, 3, 4),
    'audio': torch.randn(2, 3, 4),
}

if __name__ == '__main__':
    model = CrossModalFusionBlock(**INIT_ARGS)
    fusion_update, attn_weights = model(**ARGS)
    print(f"fusion_update shape: {fusion_update.shape}")
    print(f"attn_weights shape: {attn_weights.shape}")
    print("Success!")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
ModalityEmbedding Ran
ran on a synthesized input
smkim37/TripleSumm/models/model.py
code served (permissive licence) · get_code("79a184b673464a91")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 3, dtype family float, shape (2, 5, 8) 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

class ModalityEmbedding(nn.Module):

    def __init__(self, input_dim):
        super().__init__()
        self.embedding = nn.Embedding(num_embeddings=4, embedding_dim=input_dim)

    def forward(self, x, modality_index):
        batch_size, seq_len, _ = x.shape
        modality_tensor = torch.full((batch_size, seq_len), modality_index, dtype=torch.long, device=x.device)
        return self.embedding(modality_tensor)


INIT_ARGS = {
    'input_dim': 8
}

ARGS = {
    'x': torch.randn(2, 5, 8),
    'modality_index': 2
}

if __name__ == '__main__':
    model = ModalityEmbedding(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Output:\n{output}")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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 3, dtype family float, shape (2, 5, 8) — shape (2, 5, 8) 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
SinusoidalTemporalPE Ran
ran on a synthesized input
smkim37/TripleSumm/models/model.py
code served (permissive licence) · get_code("773afa0f598197d1")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 3, dtype family float, shape (2, 6, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import math

class SinusoidalTemporalPE(nn.Module):

    def __init__(self, input_dim, max_seq_len, dropout):
        super().__init__()
        assert input_dim % 2 == 0, 'input_dim must be even'
        position = torch.arange(max_seq_len).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, input_dim, 2) * (-math.log(10000.0) / input_dim))
        pe = torch.zeros(1, max_seq_len, input_dim)
        pe[0, :, 0::2] = torch.sin(position * div_term)
        pe[0, :, 1::2] = torch.cos(position * div_term)
        self.register_buffer('pe', pe)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        x = x + self.pe[:, :x.size(1), :]
        x = self.dropout(x)
        return x


INIT_ARGS = {
    'input_dim': 4,
    'max_seq_len': 8,
    'dropout': 0.1
}

ARGS = {
    'x': torch.randn(2, 6, 4)
}

if __name__ == '__main__':
    model = SinusoidalTemporalPE(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Input shape: {ARGS['x'].shape}")
    print(f"Output shape: {output.shape}")
    print("Test passed!")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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 3, dtype family float, shape (2, 6, 4) — shape (2, 6, 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
SwiGLU Ran
ran on a synthesized input
smkim37/TripleSumm/models/model.py
code served (permissive licence) · get_code("7b38a0ea68d56f6f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 2, dtype family float, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

INIT_ARGS = {
    'input_dim': 4,
    'dropout': 0.1,
}

ARGS = {
    'x': torch.randn(2, 4),
}
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
WindowedSelfAttention Ran
ran on a synthesized input
smkim37/TripleSumm/models/model.py
code served (permissive licence) · get_code("e5eb2b6899045502")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 3, dtype family float, shape (2, 5, 8) 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

# Initialize the class with small, distinct dimensions
INIT_ARGS = {
    'input_dim': 8,
    'num_heads': 2,
    'dropout': 0.1,
    'window_size': 3
}

# Arguments for the forward pass
ARGS = {
    'x': torch.randn(2, 5, 8),  # batch=2, seq_len=5, input_dim=8
    'mask': torch.ones(2, 5, dtype=torch.bool)  # batch=2, seq_len=5
}

# Construct and call the class
if __name__ == '__main__':
    from torch.nn import Module
    
    # Import the class (in practice it would be imported from the module)
    class WindowedSelfAttention(nn.Module):

        def __init__(self, input_dim, num_heads, dropout, window_size):
            super().__init__()
            assert input_dim % num_heads == 0, 'input_dim must be divisible by num_heads'
            assert window_size >= 0, 'window_size must be non-negative'
            assert window_size % 2 == 1 or window_size == 0, 'window_size must be odd or 0'
            self.input_dim = input_dim
            self.num_heads = num_heads
            self.head_dim = input_dim // num_heads
            self.scale = self.head_dim ** (-0.5)
            self.use_window = window_size > 0
            self.window_size = window_size
            self.radius = window_size // 2
            self.qkv_proj = nn.Linear(input_dim, input_dim * 3, bias=False)
            self.out_proj = nn.Linear(input_dim, input_dim, bias=False)
            self.dropout = nn.Dropout(dropout)

        @staticmethod
        def _edge_pad(x, radius):
            if radius == 0:
                return x
            left = x[:, :1, :].expand(-1, radius, -1)
            right = x[:, -1:, :].expand(-1, radius, -1)
            return torch.cat((left, x, right), dim=1)

        def forward(self, x, mask=None):
            B, T, D = x.shape
            if self.use_window:
                qkv = self.qkv_proj(x)
                q, k, v = qkv.chunk(3, dim=-1)
                q = q.reshape(B, T, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
                q = q.unsqueeze(-2)
                padded_k = self._edge_pad(k, self.radius)
                padded_k = padded_k.reshape(B, T + 2 * self.radius, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
                k_windows = padded_k.unfold(2, self.window_size, 1)
                padded_v = self._edge_pad(v, self.radius)
                padded_v = padded_v.reshape(B, T + 2 * self.radius, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
                v_windows = padded_v.unfold(2, self.window_size, 1)
                attn_scores = q @ k_windows * self.scale
                attn_scores = attn_scores.squeeze(-2)
                if mask is not None:
                    mask_windows = (~mask).view(B, T, 1)
                    padded_mask = torch.nn.functional.pad(mask_windows, (0, 0, self.radius, self.radius), value=True)
                    attn_mask = padded_mask.unfold(1, self.window_size, 1).squeeze(2).unsqueeze(1)
                    attn_scores = attn_scores.masked_fill(attn_mask, -1000000000.0)
                attn_probs = torch.nn.functional.softmax(attn_scores, dim=-1)
                attn_probs = self.dropout(attn_probs)
                v_windows_permuted = v_windows.permute(0, 1, 2, 4, 3)
                out = (attn_probs.unsqueeze(-2) @ v_windows_permuted).squeeze(-2)
                out = out.transpose(1, 2).reshape(B, T, D)
                out = self.out_proj(out)
            else:
                qkv = self.qkv_proj(x).reshape(B, T, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
                q, k, v = qkv.unbind(0)
                scores = q @ k.transpose(-2, -1) * self.scale
                if mask is not None:
                    attn_mask = (~mask).view(B, 1, 1, T).expand(-1, self.num_heads, T, -1)
                    scores = scores.masked_fill(attn_mask, -1000000000.0)
                attn_probs = torch.nn.functional.softmax(scores, dim=-1)
                attn_probs = self.dropout(attn_probs)
                out = attn_probs @ v
                out = out.transpose(1, 2).reshape(B, T, D)
                out = self.out_proj(out)
            return out

    model = WindowedSelfAttention(**INIT_ARGS)
    output = model(**ARGS)
    print(f"Output shape: {output.shape}")
    print(f"Expected shape: {ARGS['x'].shape}")
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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 3, dtype family float, shape (2, 5, 8) — shape (2, 5, 8) 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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
WindowedSelfAttentionLayer Ran
ran on a synthesized input
smkim37/TripleSumm/models/model.py
code served (permissive licence) · get_code("4edea45e6a96a4db")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, ndim 3, 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 torch.nn as nn
from collections import OrderedDict

INIT_ARGS = OrderedDict(
    input_dim=4,
    num_heads=2,
    dropout=0.1,
    window_size=3
)

# Construct input tensor and optional mask
B, T, D = 2, 6, 4  # batch, sequence length, feature dim
x = torch.randn(B, T, D)
mask = torch.zeros(B, T, dtype=torch.bool)
mask[0, 3:] = True  # mask last 3 positions in first batch element

ARGS = OrderedDict(
    x=x,
    mask=mask
)
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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 3, dtype family float — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
MultiScaleTemporalBlock Not yet run
unverified (harvested, no recorded run)
smkim37/TripleSumm/models/model.py
code served (permissive licence) · get_code("f59fc18b12fb43d6")
How it ran

no invocation recorded.

TripleSumm Not yet run
unverified (harvested, no recorded run)
smkim37/TripleSumm/models/model.py
code served (permissive licence) · get_code("0b2c4c77f75b0c04")
How it ran

no invocation recorded.

Also present in other papers' code

Of 10 samples lifted from this paper's linked repositories, 10 carry a function identity and 0 do not. Of the 10 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 10 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

The exponential growth of video content necessitates effective video summarization to efficiently extract key information from long videos. However, current approaches struggle to fully comprehend complex videos, primarily because they employ static or modality-agnostic fusion strategies. These methods fail to account for the dynamic, frame-dependent variations in modality saliency inherent in video data. To overcome these limitations, we propose TripleSumm, a novel architecture that adaptively weights and fuses the contributions of visual, text, and audio modalities at the frame level. Furthermore, a significant bottleneck for research into multimodal video summarization has been the lack of comprehensive benchmarks. Addressing this bottleneck, we introduce MoSu (Most Replayed Multimodal Video Summarization), the first large-scale benchmark that provides all three modalities. Extensive experiments demonstrate that Triple-Summ achieves state-of-the-art performance, outperforming existing methods by a significant margin on four benchmarks, including MoSu. Our code and dataset are available at https://github.com/smkim37/TripleSumm.

For agents

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

get_harvested_code_for_paper("2603.01169")
get_code_for_paper("2603.01169")
have("2603.01169")

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

curl https://syntology.ai/api/ran/2603.01169.json

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

[![Syntology run record](https://syntology.ai/api/ran/2603.01169.svg)](https://syntology.ai/paper/2603.01169)

Connect an agent — have() is free.