SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2112.08688 · 2021

Evidentiality-guided Generation for Knowledge-Intensive NLP Tasks

arXiv · PDF · Open in the Atlas

Code that ran

10 of 13 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; 1 honoured the contract we drafted, 2 violated it and 1 has no contract verdict; 6 failures of our instrument; 3 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/2112.08688.json · badge: /api/ran/2112.08688.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
akariasai/evidentiality_qa canonical (classifier label)
classifier basis not recorded on the link
10 ran · 3 no recorded run
FunctionStatusWhere it lives
check_answer Ran
ran, violated the contract we drafted
akariasai/evidentiality_qa/evi_gen/src/evaluation.py
pointer only (licence: MIT); the stored copy get_code serves, from facebookresearch/contriever, is pointer only (licence: NOASSERTION) · get_code("637b51b8df78832a")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from facebookresearch/contriever, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

cross_attention_forward Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
akariasai/evidentiality_qa/evi_gen/src/model.py
code served (permissive licence) · get_code("6bef6bdd6eb44070")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Sonnet 4.5, prompt python_fixture_v1) 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

torch.manual_seed(0)

# Define a minimal self object that has all the required attributes and methods
class CrossAttentionModule(nn.Module):
    def __init__(self):
        super().__init__()
        self.n_heads = 2
        self.d_kv = 3
        self.inner_dim = self.n_heads * self.d_kv  # 6
        self.dropout = 0.1
        self.has_relative_attention_bias = True
        self.score_storage = None
        
        # Linear layers
        input_dim = 8
        kv_dim = 10
        self.q = nn.Linear(input_dim, self.inner_dim, bias=False)
        self.k = nn.Linear(kv_dim, self.inner_dim, bias=False)
        self.v = nn.Linear(kv_dim, self.inner_dim, bias=False)
        self.o = nn.Linear(self.inner_dim, input_dim, bias=False)
        
    def compute_bias(self, qlen, klen):
        # Return a position bias tensor of shape (1, n_heads, qlen, klen)
        return torch.randn(1, self.n_heads, qlen, klen)
    
    def forward(self, input, mask=None, kv=None, position_bias=None, past_key_value_state=None, head_mask=None, query_length=None, use_cache=False, output_attentions=False):
        return cross_attention_forward(self, input, mask, kv, position_bias, past_key_value_state, head_mask, query_length, use_cache, output_attentions)

def cross_attention_forward(self, input, mask=None, kv=None, position_bias=None, past_key_value_state=None, head_mask=None, query_length=None, use_cache=False, output_attentions=False):
    """
    This only works for computing cross attention over the input
    """
    assert kv != None
    assert head_mask == None
    assert position_bias != None or self.has_relative_attention_bias
    bsz, qlen, dim = input.size()
    n_heads, d_heads = (self.n_heads, self.d_kv)
    klen = kv.size(1)
    q = self.q(input).view(bsz, -1, n_heads, d_heads).transpose(1, 2)
    if past_key_value_state == None:
        k = self.k(kv).view(bsz, -1, n_heads, d_heads).transpose(1, 2)
        v = self.v(kv).view(bsz, -1, n_heads, d_heads).transpose(1, 2)
    else:
        k, v = past_key_value_state
    scores = torch.einsum('bnqd,bnkd->bnqk', q, k)
    if mask is not None:
        scores += mask
    if position_bias is None:
        position_bias = self.compute_bias(qlen, klen)
    scores += position_bias
    if self.score_storage is None:
        self.score_storage = scores
    attn = F.softmax(scores.float(), dim=-1).type_as(scores)
    attn = F.dropout(attn, p=self.dropout, training=self.training)
    output = torch.matmul(attn, v)
    output = output.transpose(1, 2).contiguous().view(bsz, -1, self.inner_dim)
    output = self.o(output)
    if use_cache:
        output = (output,) + ((k, v),)
    else:
        output = (output,) + (None,)
    if output_attentions:
        output = output + (attn,)
    if self.has_relative_attention_bias:
        output = output + (position_bias,)
    return output

# Create the module instance
self_module = CrossAttentionModule()

# Create input tensors
bsz = 2
qlen = 5
input_dim = 8
klen = 7
kv_dim = 10

input_tensor = torch.randn(bsz, qlen, input_dim)
kv_tensor = torch.randn(bsz, klen, kv_dim)

# Create position_bias (1, n_heads, qlen, klen)
position_bias_tensor = torch.randn(1, self_module.n_heads, qlen, klen)

# Create mask (broadcastable to scores shape: bsz, n_heads, qlen, klen)
mask_tensor = torch.randn(1, 1, qlen, klen)

ARGS = {
    'self': self_module,
    'input': input_tensor,
    'mask': mask_tensor,
    'kv': kv_tensor,
    'position_bias': position_bias_tensor,
    'past_key_value_state': None,
    'head_mask': None,
    'query_length': None,
    'use_cache': False,
    'output_attentions': False
}
Setup block drafted by
a language model, Claude Sonnet 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 — 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
encode_passages Ran
ran, honoured the contract we drafted
akariasai/evidentiality_qa/evi_gen/src/data.py
code served (permissive licence) · get_code("bc0dae797248182b")
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 4, 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)

class StandInTokenizer:
    def batch_encode_plus(self, texts, max_length=None, pad_to_max_length=False, return_tensors=None, truncation=False):
        # Simulate tokenizer output
        batch_size = len(texts)
        seq_length = max_length if max_length else 10
        input_ids = torch.randint(0, 1000, (batch_size, seq_length))
        attention_mask = torch.ones(batch_size, seq_length, dtype=torch.long)
        
        result = {
            'input_ids': input_ids,
            'attention_mask': attention_mask
        }
        return result

# Create batch_text_passages: list of lists of strings
# Outer list has 2 items (batch), each item is a list of 3 text passages
batch_text_passages = [
    ["passage one text", "passage two text", "passage three text"],
    ["passage four text", "passage five text", "passage six text"]
]

tokenizer = StandInTokenizer()

max_length = 10

ARGS = {
    "batch_text_passages": batch_text_passages,
    "tokenizer": tokenizer,
    "max_length": max_length
}
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 4, 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
get_checkpoint_path Ran
ran; our contract draft was wrong, not the code
akariasai/evidentiality_qa/evi_gen/src/util.py
pointer only (licence: MIT); the stored copy get_code serves, from upper9527/gerea, is pointer only (licence: NONE) · get_code("e3a0e6ac26d46ae2")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from upper9527/gerea, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

has_answer Ran
ran, violated the contract we drafted
akariasai/evidentiality_qa/evi_gen/src/evaluation.py
pointer only (licence: MIT); the stored copy get_code serves, from facebookresearch/contriever, is pointer only (licence: NOASSERTION) · get_code("fa4ecbff4b9a354b")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from facebookresearch/contriever, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

init_logger Ran
ran; our contract draft was wrong, not the code
akariasai/evidentiality_qa/evi_gen/src/util.py
pointer only (licence: MIT); the stored copy get_code serves, from upper9527/gerea, is pointer only (licence: NONE) · get_code("1700ba94a8342336")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from upper9527/gerea, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

load_data Ran
ran; our contract draft was wrong, not the code
akariasai/evidentiality_qa/evi_gen/src/data.py
pointer only (licence: MIT); the stored copy get_code serves, from facebookresearch/FiD, is pointer only (licence: NOASSERTION) · get_code("148af9b729d03eba")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from facebookresearch/FiD, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

load_data_class Ran
ran on a synthesized input
akariasai/evidentiality_qa/evi_gen/src/data.py
code served (permissive licence) · get_code("977e5af0d6932a3e")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import tempfile
import os

# Create temporary JSON and JSONL files for testing
temp_dir = tempfile.mkdtemp()

# Create a .json file with sample data
json_file = os.path.join(temp_dir, "data.json")
json_data = [
    {"text": "example 1", "label": 0},
    {"text": "example 2", "label": 1},
    {"text": "example 3", "label": 0}
]
with open(json_file, 'w') as f:
    json.dump(json_data, f)

# Create a .jsonl file with sample data
jsonl_file = os.path.join(temp_dir, "data.jsonl")
with open(jsonl_file, 'w') as f:
    f.write('{"text": "line 1", "label": 0}\n')
    f.write('{"text": "line 2", "label": 1}\n')
    f.write('{"text": "line 3", "label": 0}\n')

# Define arguments dictionary
ARGS = {
    "data_path": json_file,
    "global_rank": -1,
    "world_size": -1
}
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 list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
select_examples_NQ Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
akariasai/evidentiality_qa/evi_gen/src/preprocess.py
code served (permissive licence) · get_code("e2f8b9900fd662b1")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

# Analyze function usage:
# - data: indexed with k (integer), accessed with data[k]['question'] and data[k]['answer']
# - index: iterable of integers k
# - passages: indexed with idx (integer), accessed as passages[idx][0] and passages[idx][1]
# - passages_index: dict-like, accessed with str(i) as key, returns iterable of idx values

# Build small stand-in data structures
data = {
    0: {'question': 'What is AI?', 'answer': ['Artificial Intelligence']},
    1: {'question': 'What is ML?', 'answer': ['Machine Learning']},
}

index = [0, 1]

passages = {
    0: ('First passage text about AI', 'AI Title'),
    1: ('Second passage text about ML', 'ML Title'),
    2: ('Third passage text', 'Third Title'),
}

passages_index = {
    '0': [0, 1],
    '1': [2],
}

ARGS = {
    'data': data,
    'index': index,
    'passages': passages,
    'passages_index': passages_index,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
select_examples_TQA Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
akariasai/evidentiality_qa/evi_gen/src/preprocess.py
code served (permissive licence) · get_code("e544cd4b39e5ec61")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Construct the 'data' parameter - a list of dicts with question/answer structure
data = [
    {
        'Question': 'What is the capital of France?',
        'QuestionId': 'q1',
        'Answer': {
            'Value': 'PARIS',
            'Aliases': ['Paris', 'paris']
        }
    },
    {
        'Question': 'What is the largest planet?',
        'QuestionId': 'q2',
        'Answer': {
            'Value': 'Jupiter',
            'Aliases': ['Jupiter', 'jupiter', 'JUPITER']
        }
    },
    {
        'Question': 'What is 2+2?',
        'QuestionId': 'q3',
        'Answer': {
            'Value': 'FOUR',
            'Aliases': ['four', '4']
        }
    }
]

# Construct the 'index' parameter - indices to select from data
index = [0, 2]

# Construct the 'passages' parameter - list of [text, title] pairs
passages = [
    ['Paris is the capital and most populous city of France.', 'Paris'],
    ['Jupiter is the fifth planet from the Sun.', 'Jupiter'],
    ['Two plus two equals four.', 'Arithmetic'],
    ['Additional passage text here.', 'Extra']
]

# Construct the 'passages_index' parameter - maps QuestionId to list of passage indices
passages_index = {
    'q1': [0, 1],
    'q2': [1, 2],
    'q3': [2, 3]
}

ARGS = {
    'data': data,
    'index': index,
    'passages': passages,
    'passages_index': passages_index
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
calculate_matches Not yet run
unverified (harvested, no recorded run)
akariasai/evidentiality_qa/evi_gen/src/evaluation.py
pointer only (licence: MIT); the stored copy get_code serves, from facebookresearch/contriever, is pointer only (licence: NOASSERTION) · get_code("7a4988952dd3a127")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from facebookresearch/contriever, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

get_options Not yet run
unverified (harvested, no recorded run)
akariasai/evidentiality_qa/evi_gen/src/options.py
code served (permissive licence) · get_code("9c4b02a23802ff58")
How it ran

no invocation recorded.

load Not yet run
unverified (harvested, no recorded run)
akariasai/evidentiality_qa/evi_gen/src/util.py
pointer only (licence: MIT); the stored copy get_code serves, from upper9527/gerea, is pointer only (licence: NONE) · get_code("a16b92bf7d32bad5")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from upper9527/gerea, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

Also present in other papers' code

Of 13 samples lifted from this paper's linked repositories, 13 carry a function identity and 0 do not. Of the 13 distinct functions with an identity, 12 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 12 later) and 1 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
calculate_matches 13 other papers
13 later
2401.03568 (2024-01)
Agent AI: Surveying the Horizons of Multimodal Interaction
2401.16979 (2024-01)
Re3val: Reinforced and Reranked Generative Retrieval
2402.02503 (2024-02)
GeReA: Question-Aware Prompt Captions for Knowledge-based Visual Question Answering
and 10 more not named here
check_answer 13 other papers
13 later
2401.03568 (2024-01)
Agent AI: Surveying the Horizons of Multimodal Interaction
2401.16979 (2024-01)
Re3val: Reinforced and Reranked Generative Retrieval
2402.02503 (2024-02)
GeReA: Question-Aware Prompt Captions for Knowledge-based Visual Question Answering
and 10 more not named here
has_answer 11 other papers
11 later
2401.03568 (2024-01)
Agent AI: Surveying the Horizons of Multimodal Interaction
2401.16979 (2024-01)
Re3val: Reinforced and Reranked Generative Retrieval
2402.02503 (2024-02)
GeReA: Question-Aware Prompt Captions for Knowledge-based Visual Question Answering
and 8 more not named here
cross_attention_forward 10 other papers
1 dated earlier, 9 later · 1 only as an assembly brick
2108.06314 (2021-08)
A Dataset for Answering Time-Sensitive Questions
2206.15474 (2022-06)
Forecasting Future World Events with Neural Networks
2207.12888 (2022-07)
LaKo: Knowledge-driven Visual Question Answering via Late Knowledge-to-Text Injection
and 7 more not named here
encode_passages 4 other papers
4 later
2401.03568 (2024-01)
Agent AI: Surveying the Horizons of Multimodal Interaction
2401.16979 (2024-01)
Re3val: Reinforced and Reranked Generative Retrieval
2402.02503 (2024-02)
GeReA: Question-Aware Prompt Captions for Knowledge-based Visual Question Answering
and 1 more not named here
get_options 4 other papers
4 later
2401.16979 (2024-01)
Re3val: Reinforced and Reranked Generative Retrieval
2402.05391 (2024-02)
Knowledge Graphs Meet Multi-Modal Learning: A Comprehensive Survey
2403.06095 (2024-03)
RepoHyper: Search-Expand-Refine on Semantic Graphs for Repository-Level Code Completion
and 1 more not named here
select_examples_NQ 3 other papers
3 later
2401.16979 (2024-01)
Re3val: Reinforced and Reranked Generative Retrieval
2402.02503 (2024-02)
GeReA: Question-Aware Prompt Captions for Knowledge-based Visual Question Answering
2402.05391 (2024-02)
Knowledge Graphs Meet Multi-Modal Learning: A Comprehensive Survey
select_examples_TQA 3 other papers
3 later
2401.16979 (2024-01)
Re3val: Reinforced and Reranked Generative Retrieval
2402.02503 (2024-02)
GeReA: Question-Aware Prompt Captions for Knowledge-based Visual Question Answering
2402.05391 (2024-02)
Knowledge Graphs Meet Multi-Modal Learning: A Comprehensive Survey
get_checkpoint_path 2 other papers
2 later
2401.16979 (2024-01)
Re3val: Reinforced and Reranked Generative Retrieval
2402.02503 (2024-02)
GeReA: Question-Aware Prompt Captions for Knowledge-based Visual Question Answering
init_logger 2 other papers
2 later
2401.16979 (2024-01)
Re3val: Reinforced and Reranked Generative Retrieval
2402.02503 (2024-02)
GeReA: Question-Aware Prompt Captions for Knowledge-based Visual Question Answering
load 2 other papers
2 later
2401.16979 (2024-01)
Re3val: Reinforced and Reranked Generative Retrieval
2402.02503 (2024-02)
GeReA: Question-Aware Prompt Captions for Knowledge-based Visual Question Answering
load_data 2 other papers
2 later
2401.16979 (2024-01)
Re3val: Reinforced and Reranked Generative Retrieval
2025.findings-acl.1365 (2025 (year only))
ECoRAG: Evidentiality-guided Compression for Long Context RAG

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

Retrieval-augmented generation models have shown state-of-the-art performance across many knowledge-intensive NLP tasks such as open question answering and fact verification. These models are trained to generate the final output given the retrieved passages, which can be irrelevant to the original query, leading to learning spurious cues or answer memorization. This work introduces a method to incorporate the evidentiality of passages -- whether a passage contains correct evidence to support the output -- into training the generator. We introduce a multi-task learning framework to jointly generate the final output and predict the evidentiality of each passage, leveraging a new task-agnostic method to obtain silver evidentiality labels for supervision. Our experiments on five datasets across three knowledge-intensive tasks show that our new evidentiality-guided generator significantly outperforms its direct counterpart with the same-size model and advances the state of the art on FaVIQ-Ambig. We attribute these improvements to both the auxiliary multi-task learning and silver evidentiality mining techniques.

For agents

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

get_harvested_code_for_paper("2112.08688")
get_code_for_paper("2112.08688")
have("2112.08688")

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

curl https://syntology.ai/api/ran/2112.08688.json

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

[![Syntology run record](https://syntology.ai/api/ran/2112.08688.svg)](https://syntology.ai/paper/2112.08688)

Connect an agent — have() is free.