SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2410.06003 · NeurIPS · 2024

Is the MMI Criterion Necessary for Interpretability? Degenerating Non-causal Features to Plain Noise for Self-Rationalization

Jun Wang, Wei Liu, Haozhao Wang, Ruixuan Li, Yuankai Zhang, Zhiying Deng, Zhongyu Niu

arXiv · PDF · Open in the Atlas

Code that ran

10 of 10 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (2 of them constructed an object rather than computing a result; 4 honoured the contract we drafted, 1 violated it and 2 have no contract verdict; 3 failures of our instrument; 0 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; 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/2410.06003.json · badge: /api/ran/2410.06003.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
jugechengzi/Rationalization-MRD canonical (classifier label)
classifier basis: census
8 ran · 0 no recorded run
jugechengzi/rationalization-mrd canonical (classifier label)
classifier basis: classifier_2025
2 ran · 0 no recorded run
FunctionStatusWhere it lives
SelectItem Ran
ran on a synthesized input · constructed an object; no result was computed
jugechengzi/rationalization-mrd/model.py
code served (permissive licence) · get_code("0ff73f4686a15dfd")
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, SelectItem: 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

torch.manual_seed(0)

# The SelectItem module takes an item_index parameter
# and in forward(), it indexes into the inputs using that index
# So inputs should be a tuple, list, or sequence-like object

# Create a simple test case where inputs is a tuple of tensors
item_index = 0
test_inputs = (
    torch.randn(2, 4),  # First tensor
    torch.randn(2, 3),  # Second tensor
)

ARGS = {
    'item_index': item_index
}
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 SelectItem — 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 SelectItem 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
Sp_norm_rnp Ran
ran on a synthesized input · constructed an object; no result was computed
jugechengzi/rationalization-mrd/model.py
code served (permissive licence) · get_code("4b2ad6b5073236b1")
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, Sp_norm_rnp: 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
import numpy as np

torch.manual_seed(0)

class Args:
    def __init__(self):
        self.vocab_size = 100
        self.embedding_dim = 8
        self.hidden_dim = 16
        self.num_layers = 1
        self.num_class = 3
        self.dropout = 0.1
        self.pretrained_embedding = None

ARGS = {
    'args': Args()
}
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 Sp_norm_rnp — 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 Sp_norm_rnp 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
classfy Ran
ran, honoured the contract we drafted
jugechengzi/Rationalization-MRD/train_util.py
code served (permissive licence) · get_code("91596484f00a8c53")
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, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 2)
    
    def forward(self, inputs, masks):
        # inputs shape: (batch_size, 4)
        # masks shape: (batch_size, 4)
        # Return logits of shape (batch_size, 2) for binary classification
        return self.linear(inputs)

# Create a simple dataset as a list of batches
# Each batch is (inputs, masks, labels)
dataset = [
    (
        torch.randn(2, 4),  # inputs: batch_size=2, feature_dim=4
        torch.randn(2, 4),  # masks: batch_size=2, feature_dim=4
        torch.tensor([0, 1])  # labels: batch_size=2, binary labels
    ),
    (
        torch.randn(2, 4),
        torch.randn(2, 4),
        torch.tensor([1, 0])
    )
]

model = SimpleModel()
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
device = torch.device('cpu')

args = {}

ARGS = {
    'model': model,
    'optimizer': optimizer,
    'dataset': dataset,
    'device': device,
    'args': args
}
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, ndim 1, 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
compute_micro_stats Ran
ran; our contract draft was wrong, not the code
jugechengzi/Rationalization-MRD/metric.py
code served (permissive licence) · get_code("560b7aab59832a78")
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, 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

torch.manual_seed(0)

# Create small binary tensors for labels and predictions
batch_size = 2
sequence_length = 4

# labels: binary indicators of rationale (0 or 1)
labels = torch.randint(0, 2, (batch_size, sequence_length))

# predictions: soft probabilities in [0, 1] range
predictions = torch.rand(batch_size, sequence_length)

ARGS = {
    "labels": labels,
    "predictions": predictions,
}
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, ndim 1, dtype family int — 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
dev_adv_causal Ran
ran, honoured the contract we drafted
jugechengzi/Rationalization-MRD/train_util.py
code served (permissive licence) · get_code("e84f787c88cbb39b")
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, 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 torch
import torch.nn as nn
from torch.utils.tensorboard import SummaryWriter
import tempfile

torch.manual_seed(0)

# Define a minimal model that implements the methods called in dev_adv_causal
class MinimalModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 2)
    
    def get_rationale(self, inputs, masks):
        # Returns rationale tensor with same shape as inputs
        batch_size = inputs.shape[0]
        return torch.rand(batch_size, 3, 4)
    
    def adversarial_pred_logit(self, inputs, masks, rationales):
        # Returns logits for binary classification
        batch_size = inputs.shape[0]
        return self.linear(rationales.mean(dim=1))
    
    def train_one_step(self, inputs, masks):
        # Returns logits for binary classification
        batch_size = inputs.shape[0]
        return self.linear(inputs.mean(dim=1))
    
    def pred_forward_logit(self, inputs, masks, rationales):
        # Returns logits for binary classification
        batch_size = inputs.shape[0]
        return self.linear(rationales.mean(dim=1))
    
    def eval(self):
        super().eval()

# Create a minimal dataset as a list of batches
dataset = [
    (torch.randn(2, 3, 4), torch.ones(2, 3), torch.tensor([0, 1])),
    (torch.randn(2, 3, 4), torch.ones(2, 3), torch.tensor([1, 0])),
]

# Create model
model = MinimalModel()

# Setup device
device = 'cpu'

# Create args dict with minimal configuration
args = {
    'learning_rate': 0.001,
    'batch_size': 2,
}

# Create a temporary directory for tensorboard
temp_dir = tempfile.mkdtemp()
writer = SummaryWriter(temp_dir)
writer_epoch = [writer, 0]

ARGS = {
    'model': model,
    'dataset': dataset,
    'device': device,
    'args': args,
    'writer_epoch': writer_epoch,
}
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, ndim 2, 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
get_continuity_loss Ran
ran, honoured the contract we drafted
jugechengzi/Rationalization-MRD/metric.py
code served (permissive licence) · get_code("c92373379a9ed7e6")
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 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 torch

torch.manual_seed(0)

# Construct a small tensor with shape (batch_size=2, sequence_length=4)
# The function computes differences along sequence dimension and takes mean
z = torch.randn(2, 4)

ARGS = {
    "z": z
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 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_embeddings Ran
ran; our contract draft was wrong, not the code
jugechengzi/Rationalization-MRD/embedding.py
code served (permissive licence) · get_code("a7fd4428a7a26977")
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 gzip
import numpy as np
import tempfile
import os

# Create a temporary gzip file with embedding data
temp_dir = tempfile.mkdtemp()
embedding_path = os.path.join(temp_dir, "embeddings.txt.gz")

# Write sample embedding data to gzip file
with gzip.open(embedding_path, 'wt') as f:
    f.write("word1 0.1 0.2 0.3 0.4\n")
    f.write("word2 0.5 0.6 0.7 0.8\n")
    f.write("word3 0.9 1.0 1.1 1.2\n")

ARGS = {
    "embedding_path": embedding_path
}
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
get_glove_embedding Ran
ran; our contract draft was wrong, not the code
jugechengzi/Rationalization-MRD/embedding.py
code served (permissive licence) · get_code("f5f0244064d12e86")
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 tempfile
import os

# Create a temporary GloVe embedding file
temp_dir = tempfile.mkdtemp()
glove_file = os.path.join(temp_dir, "glove.txt")

# Write a small GloVe format file
with open(glove_file, 'w', encoding='utf-8') as f:
    f.write("hello 0.1 0.2 0.3 0.4\n")
    f.write("world 0.5 0.6 0.7 0.8\n")
    f.write("python 0.9 1.0 1.1 1.2\n")

ARGS = {
    "glove_embedding_path": glove_file
}
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
get_sparsity_loss Ran
ran, violated the contract we drafted
jugechengzi/Rationalization-MRD/metric.py
code served (permissive licence) · get_code("7d30b2c623f59a22")
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 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 torch

torch.manual_seed(0)

# z: (batch_size, sequence_length) - tensor of values to sum
z = torch.randn(2, 4)

# mask: (batch_size, seq_length) - tensor used to normalize the sum
mask = torch.ones(2, 4)

# level: sparsity level (a float scalar)
level = 0.5

ARGS = {
    "z": z,
    "mask": mask,
    "level": level
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 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
validate_dev_sentence Ran
ran, honoured the contract we drafted
jugechengzi/Rationalization-MRD/validate_util.py
code served (permissive licence) · get_code("b57d1110e6ff819d")
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, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class DummyModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 2)
    
    def train_one_step(self, inputs, masks):
        # inputs shape: (batch_size, seq_len, hidden_dim)
        # Return logits of shape (batch_size, 2) for binary classification
        batch_size = inputs.shape[0]
        return self.linear(inputs[:, 0, :])  # Use first token, output (batch_size, 2)

class DummyWriter:
    def add_scalar(self, tag, value, step):
        pass

# Create model
model = DummyModel()

# Create dev_loader: list of batches with (inputs, masks, labels)
# inputs: (batch_size, seq_len, hidden_dim) = (2, 3, 4)
# masks: (batch_size, seq_len) = (2, 3)
# labels: (batch_size,) = (2,) with values 0 or 1
dev_loader = [
    (
        torch.randn(2, 3, 4),
        torch.ones(2, 3),
        torch.tensor([0, 1])
    ),
    (
        torch.randn(2, 3, 4),
        torch.ones(2, 3),
        torch.tensor([1, 0])
    )
]

device = 'cpu'

writer = DummyWriter()
epoch = 1
writer_epoch = (writer, epoch)

ARGS = {
    'model': model,
    'dev_loader': dev_loader,
    'device': device,
    'writer_epoch': writer_epoch
}
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, ndim 1, 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

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, 7 are also present in other papers whose code we hold (6 with at least one paper dated earlier, 0 the same month, 6 later) and 3 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
compute_micro_stats 3 other papers
2 dated earlier, 1 later
2309.13391 (2023-09)
D-Separation for Causal Self-Explanation
2312.04103 (2023-12)
Enhancing the Rationale-Input Alignment for Self-explaining Rationalization
2503.06202 (2025-03)
Breaking Free from MMI: A New Frontier in Rationalization by Probing Input Utilization
get_continuity_loss 3 other papers
2 dated earlier, 1 later
2309.13391 (2023-09)
D-Separation for Causal Self-Explanation
2312.04103 (2023-12)
Enhancing the Rationale-Input Alignment for Self-explaining Rationalization
2503.06202 (2025-03)
Breaking Free from MMI: A New Frontier in Rationalization by Probing Input Utilization
get_embeddings 3 other papers
2 dated earlier, 1 later
2309.13391 (2023-09)
D-Separation for Causal Self-Explanation
2312.04103 (2023-12)
Enhancing the Rationale-Input Alignment for Self-explaining Rationalization
2503.06202 (2025-03)
Breaking Free from MMI: A New Frontier in Rationalization by Probing Input Utilization
get_glove_embedding 3 other papers
2 dated earlier, 1 later
2309.13391 (2023-09)
D-Separation for Causal Self-Explanation
2312.04103 (2023-12)
Enhancing the Rationale-Input Alignment for Self-explaining Rationalization
2503.06202 (2025-03)
Breaking Free from MMI: A New Frontier in Rationalization by Probing Input Utilization
get_sparsity_loss 3 other papers
2 dated earlier, 1 later
2309.13391 (2023-09)
D-Separation for Causal Self-Explanation
2312.04103 (2023-12)
Enhancing the Rationale-Input Alignment for Self-explaining Rationalization
2503.06202 (2025-03)
Breaking Free from MMI: A New Frontier in Rationalization by Probing Input Utilization
SelectItem
brick of this paper's assembly: helper
1 other paper
1 later · 1 only as an assembly brick
2505.02118 (2025-05)
Adversarial Cooperative Rationalization: The Risk of Spurious Correlations in Even Clean Datasets
classfy 1 other paper
1 dated earlier
2309.13391 (2023-09)
D-Separation for Causal Self-Explanation

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

An important line of research in the field of explainability is to extract a small subset of crucial rationales from the full input. The most widely used criterion for rationale extraction is the maximum mutual information (MMI) criterion. However, in certain datasets, there are spurious features non-causally correlated with the label and also get high mutual information, complicating the loss landscape of MMI. Although some penalty-based methods have been developed to penalize the spurious features (e.g., invariance penalty, intervention penalty, etc) to help MMI work better, these are merely remedial measures. In the optimization objectives of these methods, spurious features are still distinguished from plain noise, which hinders the discovery of causal rationales. This paper aims to develop a new criterion that treats spurious features as plain noise, allowing the model to work on datasets rich in spurious features as if it were working on clean datasets, thereby making rationale extraction easier. We theoretically observe that removing either plain noise or spurious features from the input does not alter the conditional distribution of the remaining components relative to the task label. However, significant changes in the conditional distribution occur only when causal features are eliminated. Based on this discovery, the paper proposes a criterion for Maximizing the Remaining Discrepancy (MRD). Experiments on six widely used datasets show that our MRD criterion improves rationale quality (measured by the overlap with human-annotated rationales) by up to 10.4% as compared to several recent competitive MMI variants. Code: https://github.com/jugechengzi/Rationalization-MRD.

For agents

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

get_harvested_code_for_paper("2410.06003")
get_code_for_paper("2410.06003")
have("2410.06003")

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

curl https://syntology.ai/api/ran/2410.06003.json

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

[![Syntology run record](https://syntology.ai/api/ran/2410.06003.svg)](https://syntology.ai/paper/2410.06003)

Connect an agent — have() is free.