SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2406.15927 · 2024

Semantic Entropy Probes: Robust and Cheap Hallucination Detection in LLMs

arXiv · PDF · Open in the Atlas

Code that ran

12 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; 0 honoured the contract we drafted, 0 violated it and 8 have no contract verdict; 4 failures of our instrument; 1 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/2406.15927.json · badge: /api/ran/2406.15927.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
oatml/semantic-entropy-probes community (classifier label)
classifier basis not recorded on the link
12 ran · 1 no recorded run
FunctionStatusWhere it lives
accuracy_at_quantile Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
oatml/semantic-entropy-probes/semantic_uncertainty/uncertainty/utils/eval_utils.py
code served (permissive licence) · get_code("c425058e87c88adb")
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 float64, ndim 0, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Create sample data for the function
np.random.seed(0)

# accuracies: array of boolean or float values indicating correctness
# Should be a 1D array of values between 0 and 1 (or boolean)
accuracies = np.array([1.0, 0.0, 1.0, 1.0, 0.0, 1.0, 1.0, 0.0])

# uncertainties: array of uncertainty scores, same length as accuracies
# Should be a 1D array of values
uncertainties = np.array([0.1, 0.9, 0.2, 0.3, 0.8, 0.15, 0.25, 0.95])

# quantile: a float between 0 and 1 representing the quantile threshold
quantile = 0.5

ARGS = {
    "accuracies": accuracies,
    "uncertainties": uncertainties,
    "quantile": quantile,
}
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 float64, ndim 0, dtype family float — no output shape is stated: observe_signature.py recorded float ndim 0, which disagrees with the fixture run
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
auroc Ran
ran on a synthesized input
oatml/semantic-entropy-probes/semantic_uncertainty/uncertainty/utils/eval_utils.py
code served (permissive licence) · get_code("ccde4062eb35a7ad")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# The auroc function takes y_true and y_score as parameters
# y_true should be binary labels (0/1)
# y_score should be predicted probabilities or scores

# Create small sample data
y_true = np.array([0, 1, 1, 0, 1, 0])
y_score = np.array([0.1, 0.9, 0.8, 0.2, 0.7, 0.3])

ARGS = {
    "y_true": y_true,
    "y_score": y_score
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, sklearn — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
bootstrap Ran
ran; our contract draft was wrong, not the code
oatml/semantic-entropy-probes/semantic_uncertainty/uncertainty/utils/eval_utils.py
code served (permissive licence) · get_code("4b00bf597ecfdaac")
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 function. 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 scipy.stats

def dummy_function(data):
    """A simple statistic function - computes mean"""
    return np.mean(data)

# Create sample data as a numpy array
sample_data = np.array([1.2, 2.3, 1.8, 2.1, 1.9, 2.4, 2.0, 1.7])

# Create RNG
rng = np.random.default_rng(42)

ARGS = {
    'function': dummy_function,
    'rng': rng,
    'n_resamples': 1000
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function 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
calculate_p_true Ran
ran on a synthesized input
oatml/semantic-entropy-probes/semantic_uncertainty/uncertainty/uncertainty_measures/p_true.py
code served (permissive licence) · get_code("ab0a03d09dc95c36")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

class MockModel(nn.Module):
    """Mock model that implements get_p_true method."""
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(10, 1)
    
    def get_p_true(self, prompt):
        """Mock implementation that returns a log probability."""
        # Simulate computing a log probability based on prompt
        # In reality this would use the model to score the prompt
        return -0.5

ARGS = {
    'model': MockModel(),
    'question': 'What is the capital of France?',
    'most_probable_answer': 'Paris',
    'brainstormed_answers': ['Lyon', 'Marseille', 'Nice'],
    'few_shot_prompt': 'Example Q: What is 2+2?\nExample A: 4',
    'hint': False
}
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 float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
construct_fewshot_prompt_from_indices Ran
ran on a synthesized input
oatml/semantic-entropy-probes/semantic_uncertainty/uncertainty/utils/utils.py
code served (permissive licence) · get_code("0b79bb92328a4653")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Create a mock dataset that supports indexing and returns dicts with the expected structure
class MockDataset:
    def __init__(self, num_examples=5):
        self.num_examples = num_examples
    
    def __getitem__(self, idx):
        return {
            'context': f'This is context {idx}. It contains background information.',
            'question': f'What is question {idx}?',
            'answers': {
                'text': [f'Answer to question {idx}']
            }
        }

# Create a simple make_prompt function that combines context, question, and answer
def make_prompt(context, question, answer, brief, brief_always):
    prompt_text = f"Context: {context}\nQuestion: {question}\nAnswer: {answer}\n"
    return prompt_text

# Create instances for the arguments
dataset = MockDataset(num_examples=5)
example_indices = [0, 1, 2]
brief = "You are a helpful assistant.\n"
brief_always = False

ARGS = {
    'dataset': dataset,
    'example_indices': example_indices,
    'brief': brief,
    'brief_always': brief_always,
    'make_prompt': make_prompt
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
context_entails_response Ran
ran on a synthesized input
oatml/semantic-entropy-probes/semantic_uncertainty/uncertainty/uncertainty_measures/semantic_entropy.py
code served (permissive licence) · get_code("0814f57f66c87087")
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 float64, ndim 0, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np
import torch
import torch.nn as nn

torch.manual_seed(0)

class MockModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 2)
    
    def check_implication(self, context, response):
        # Returns a binary value (0 or 1) indicating if context entails response
        # We'll return a simple deterministic value based on input
        return np.random.randint(0, 2)

context = "The cat is on the mat"
responses = [
    "The cat exists",
    "Something is on the mat",
    "The mat is under the cat"
]
model = MockModel()

ARGS = {
    "context": context,
    "responses": responses,
    "model": model
}
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 float64, ndim 0, dtype family float — no output shape is stated: observe_signature.py recorded float ndim 0, which disagrees with the fixture run
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_parser Ran
ran on a synthesized input
oatml/semantic-entropy-probes/semantic_uncertainty/uncertainty/utils/utils.py
code served (permissive licence) · get_code("e5a9b05c0453abbc")
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 ArgumentParser. 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 os
import argparse

# Set a random seed for reproducibility
import torch
torch.manual_seed(0)

# Define ARGS dict mapping all parameters of get_parser to values
ARGS = {
    'stages': ['generate', 'compute']
}
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 ArgumentParser — a ArgumentParser output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a ArgumentParser 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
get_semantic_ids Ran
ran on a synthesized input
oatml/semantic-entropy-probes/semantic_uncertainty/uncertainty/uncertainty_measures/semantic_entropy.py
code served (permissive licence) · get_code("80a6075a64b753ae")
How it ran

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

import torch
import torch.nn as nn

torch.manual_seed(0)

class MockImplicationModel(nn.Module):
    """Mock model that checks implication between two texts."""
    
    def __init__(self):
        super().__init__()
        self.embedding = nn.Linear(10, 8)
    
    def check_implication(self, text1, text2, example=None):
        """
        Returns implication score:
        0 = no implication
        1 = one-way implication
        2 = mutual implication (equivalence)
        """
        # Simple mock: hash-based deterministic behavior for testing
        hash_val = (hash(text1) + hash(text2)) % 3
        return hash_val

# Create test data
strings_list = [
    "The cat is sleeping",
    "A feline is resting",
    "The dog is running",
    "A canine is moving",
    "The bird is flying"
]

model = MockImplicationModel()

ARGS = {
    "strings_list": strings_list,
    "model": model,
    "strict_entailment": False,
    "example": None
}
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 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
logsumexp_by_id Ran
ran on a synthesized input
oatml/semantic-entropy-probes/semantic_uncertainty/uncertainty/uncertainty_measures/semantic_entropy.py
code served (permissive licence) · get_code("b95c0c18affed908")
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 numpy as np

# Construct valid arguments for logsumexp_by_id

# semantic_ids: list of integer IDs that must be consecutive starting from 0
semantic_ids = [0, 0, 1, 1, 2]

# log_likelihoods: array of log probabilities corresponding to each semantic_id
log_likelihoods = np.array([-1.0, -2.5, -0.5, -3.0, -1.5])

# agg: aggregation method (one of 'sum', 'sum_normalized', 'mean')
agg = 'sum'

ARGS = {
    'semantic_ids': semantic_ids,
    'log_likelihoods': log_likelihoods,
    'agg': agg
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 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
md5hash Ran
ran; our contract draft was wrong, not the code
oatml/semantic-entropy-probes/semantic_uncertainty/uncertainty/utils/openai.py
code served (permissive licence) · get_code("a7132c6a5c4b94e4")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned int, ndim 0, dtype family 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 hashlib

# The function takes a single parameter: string
# It encodes the string to UTF-8, computes MD5 hash, and converts to int
# So we need to provide a valid string argument

ARGS = {
    "string": "hello world"
}
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 int, 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
remove_split_layer Ran
ran on a synthesized input
oatml/semantic-entropy-probes/semantic_uncertainty/uncertainty/models/huggingface_models.py
code served (permissive licence) · get_code("50909f87202233f8")
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 dict. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 copy
import logging
from collections import Counter

# Configure logging to see info messages
logging.basicConfig(level=logging.INFO)

# The function takes a device_map_in parameter which is a dictionary
# where keys are layer names (strings with dot notation) and values are devices
# The function uses:
# - copy.deepcopy on device_map_in
# - .keys() on the dict
# - string operations like split, join, startswith
# - Counter on a list of strings
# - .pop() on the dict

# Create a valid device_map_in that represents a split layer scenario
# A split layer is when a layer (first 2 parts of dot notation) appears multiple times
device_map_in = {
    'model.layer.0.weight': 'cuda:0',
    'model.layer.0.bias': 'cuda:1',  # Same layer (model.layer) but different device - this is a split
    'model.layer.1.weight': 'cuda:0',
    'model.layer.1.bias': 'cuda:0',
}

ARGS = {
    'device_map_in': device_map_in,
}
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 dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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
split_dataset Ran
ran; our contract draft was wrong, not the code
oatml/semantic-entropy-probes/semantic_uncertainty/uncertainty/utils/utils.py
code served (permissive licence) · get_code("eb7cd0b37505f35f")
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)

# Create a minimal dataset that satisfies the function's requirements
# The function expects a list-like object where each element has an 'answers' key
# with a 'text' key that contains a list

dataset = [
    {'answers': {'text': ['answer1', 'answer2']}},  # answerable
    {'answers': {'text': []}},  # unanswerable
    {'answers': {'text': ['answer3']}},  # answerable
    {'answers': {'text': []}},  # unanswerable
    {'answers': {'text': ['answer4', 'answer5', 'answer6']}},  # answerable
]

ARGS = {
    'dataset': dataset
}
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_p_ik Not yet run
unverified (harvested, no recorded run)
oatml/semantic-entropy-probes/semantic_uncertainty/uncertainty/uncertainty_measures/p_ik.py
code served (permissive licence) · get_code("f85264d7c7639ee8")
How it ran

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, 4 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 4 later) and 9 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
accuracy_at_quantile 1 other paper
1 later
2605.08863 (2026-05)
Max-pooling Network Revisited: Analyzing the Role of Semantic Probability in Multiple Instance Learning for Hallucination Detection
bootstrap 1 other paper
1 later
2605.08863 (2026-05)
Max-pooling Network Revisited: Analyzing the Role of Semantic Probability in Multiple Instance Learning for Hallucination Detection
md5hash 1 other paper
1 later
2605.08863 (2026-05)
Max-pooling Network Revisited: Analyzing the Role of Semantic Probability in Multiple Instance Learning for Hallucination Detection
split_dataset 1 other paper
1 later
2605.08863 (2026-05)
Max-pooling Network Revisited: Analyzing the Role of Semantic Probability in Multiple Instance Learning for Hallucination Detection

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

We propose semantic entropy probes (SEPs), a cheap and reliable method for uncertainty quantification in Large Language Models (LLMs). Hallucinations, which are plausible-sounding but factually incorrect and arbitrary model generations, present a major challenge to the practical adoption of LLMs. Recent work by Farquhar et al. (2024) proposes semantic entropy (SE), which can detect hallucinations by estimating uncertainty in the space semantic meaning for a set of model generations. However, the 5-to-10-fold increase in computation cost associated with SE computation hinders practical adoption. To address this, we propose SEPs, which directly approximate SE from the hidden states of a single generation. SEPs are simple to train and do not require sampling multiple model generations at test time, reducing the overhead of semantic uncertainty quantification to almost zero. We show that SEPs retain high performance for hallucination detection and generalize better to out-of-distribution data than previous probing methods that directly predict model accuracy. Our results across models and tasks suggest that model hidden states capture SE, and our ablation studies give further insights into the token positions and model layers for which this is the case.

For agents

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

get_harvested_code_for_paper("2406.15927")
get_code_for_paper("2406.15927")
have("2406.15927")

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

curl https://syntology.ai/api/ran/2406.15927.json

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

[![Syntology run record](https://syntology.ai/api/ran/2406.15927.svg)](https://syntology.ai/paper/2406.15927)

Connect an agent — have() is free.