SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2410.06554 · EMNLP · 2024

The Accuracy Paradox in RLHF: When Better Reward Models Don't Yield Better Language Models

Xiaoyu Shen, Dawei Zhu, Xinghao Chen, Wei Zhang, Yanjun Chen, Yirong Sun

arXiv · PDF · Open in the Atlas

Code that ran

5 of 9 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 5 failures of our instrument; 4 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/2410.06554.json · badge: /api/ran/2410.06554.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
EIT-NLP/AccuracyParadox-RLHF canonical (classifier label)
classifier basis: census
5 ran · 4 no recorded run
FunctionStatusWhere it lives
create_position_ids_from_input_ids Ran
ran; our contract draft was wrong, not the code
EIT-NLP/AccuracyParadox-RLHF/reward_modeling/my_longformer.py
code served (permissive licence) · get_code("ce0ed06559b93e81")
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 2, dtype family int, 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

torch.manual_seed(0)

# Create input_ids tensor: shape (batch_size=2, seq_length=4)
# Use token IDs where 0 is the padding token
input_ids = torch.tensor([
    [0, 5, 6, 7],
    [0, 8, 0, 9]
], dtype=torch.long)

# padding_idx indicates which token ID represents padding
padding_idx = 0

ARGS = {
    "input_ids": input_ids,
    "padding_idx": padding_idx,
}
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 2, dtype family int, 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: none by the harness; any seed is in the setup block
reduce_mean Ran
ran; our contract draft was wrong, not the code
EIT-NLP/AccuracyParadox-RLHF/fgrlhf/utils.py
code served (permissive licence) · get_code("e5bacba887a721ef")
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)

# Create small test tensors
value = torch.randn(2, 4)  # batch_size=2, dim=4
mask = torch.ones(2, 4)    # same shape as value

# Test with axis=None
ARGS = {
    "value": value,
    "mask": mask,
    "axis": 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 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
reduce_std Ran
ran; our contract draft was wrong, not the code
EIT-NLP/AccuracyParadox-RLHF/fgrlhf/utils.py
code served (permissive licence) · get_code("174bfbbeb1f29e12")
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)

# Create small test tensors
# value: the tensor to reduce
# mask: binary mask to select which elements to include
value = torch.randn(2, 4)
mask = torch.ones(2, 4)

# Set some mask values to 0 to have meaningful masking
mask[0, 2] = 0
mask[1, 3] = 0

ARGS = {
    "value": value,
    "mask": mask,
}
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
reduce_sum Ran
ran; our contract draft was wrong, not the code
EIT-NLP/AccuracyParadox-RLHF/fgrlhf/utils.py
code served (permissive licence) · get_code("5874b0a8910897fd")
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)

# Create small tensors for testing
value = torch.randn(2, 4)  # Shape (2, 4)
mask = torch.randint(0, 2, (2, 4)).float()  # Shape (2, 4), binary mask as floats

ARGS = {
    "value": value,
    "mask": mask,
    "axis": 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 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
tril_bfloat16 Ran
ran; our contract draft was wrong, not the code
EIT-NLP/AccuracyParadox-RLHF/reward_modeling/my_longformer.py
code served (permissive licence) · get_code("d37f5dbfe7b33d1e")
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. 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 small bfloat16 tensor for testing
# Using distinct dimensions: 2x4 to make shapes unambiguous
input_tensor = torch.randn(2, 4, dtype=torch.bfloat16)

# Define arguments
ARGS = {
    "input_tensor": input_tensor,
    "diagonal": 0
}
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, shape (2, 4) — shape (2, 4) as a later run recorded it
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_rouge_scores Not yet run
unverified (harvested, no recorded run)
EIT-NLP/AccuracyParadox-RLHF/fgrlhf/evaluators.py
code served (permissive licence) · get_code("b3257abd722edb60")
How it ran

no invocation recorded.

postprocess_text Not yet run
unverified (harvested, no recorded run)
EIT-NLP/AccuracyParadox-RLHF/fgrlhf/evaluators.py
code served (permissive licence) · get_code("94a55f838c975848")
How it ran

no invocation recorded.

split_text_to_sentences Not yet run
unverified (harvested, no recorded run)
EIT-NLP/AccuracyParadox-RLHF/fgrlhf/reward_utils.py
code served (permissive licence) · get_code("e7f95d91543ac720")
How it ran

no invocation recorded.

split_text_to_subsentences Not yet run
unverified (harvested, no recorded run)
EIT-NLP/AccuracyParadox-RLHF/fgrlhf/reward_utils.py
code served (permissive licence) · get_code("11589c4018cfc261")
How it ran

no invocation recorded.

Also present in other papers' code

Of 9 samples lifted from this paper's linked repositories, 9 carry a function identity and 0 do not. Of the 9 distinct functions with an identity, 9 are also present in other papers whose code we hold (8 with at least one paper dated earlier, 0 the same month, 3 later) and 0 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
create_position_ids_from_input_ids 8 other papers
7 dated earlier, 1 later
2021.emnlp-main.154 (2021 (year only))
Reconstruction Attack on Instance Encoding for Language Understanding
2104.08247 (2021-04)
What to Pre-Train on? Efficient Intermediate Task Selection
2201.03514 (2022-01)
Black-Box Tuning for Language-Model-as-a-Service
and 5 more not named here
reduce_mean 7 other papers
7 dated earlier · 3 only as an assembly brick
2210.03078 (2022-10)
RAINIER: Reinforced Knowledge Introspector for Commonsense Question Answering
2212.10409 (2022-12)
ClarifyDelphi: Reinforced Clarification Questions with Defeasibility Rewards for Social and Moral Situations
2305.13829 (2023-05)
Learning from Mistakes via Cooperative Study Assistant for Large Language Models
and 4 more not named here
reduce_sum 7 other papers
7 dated earlier · 3 only as an assembly brick
2210.03078 (2022-10)
RAINIER: Reinforced Knowledge Introspector for Commonsense Question Answering
2212.10409 (2022-12)
ClarifyDelphi: Reinforced Clarification Questions with Defeasibility Rewards for Social and Moral Situations
2305.13829 (2023-05)
Learning from Mistakes via Cooperative Study Assistant for Large Language Models
and 4 more not named here
reduce_std 5 other papers
5 dated earlier · 1 only as an assembly brick
2210.03078 (2022-10)
RAINIER: Reinforced Knowledge Introspector for Commonsense Question Answering
2212.10409 (2022-12)
ClarifyDelphi: Reinforced Clarification Questions with Defeasibility Rewards for Social and Moral Situations
2305.13829 (2023-05)
Learning from Mistakes via Cooperative Study Assistant for Large Language Models
and 2 more not named here
postprocess_text 2 other papers
1 dated earlier, 1 later
2402.04315 (2024-02)
Training Language Models to Generate Text with Citations via Fine-grained Rewards
2025.findings-emnlp.228 (2025 (year only))
FG-PRM: Fine-grained Hallucination Detection and Mitigation in Language Model Mathematical Reasoning
split_text_to_sentences 2 other papers
2 dated earlier
2402.04315 (2024-02)
Training Language Models to Generate Text with Citations via Fine-grained Rewards
2403.06754 (2024-03)
ALaRM: Align Language Models via Hierarchical Rewards Modeling
split_text_to_subsentences 2 other papers
2 dated earlier
2402.04315 (2024-02)
Training Language Models to Generate Text with Citations via Fine-grained Rewards
2403.06754 (2024-03)
ALaRM: Align Language Models via Hierarchical Rewards Modeling
get_rouge_scores 1 other paper
1 later
2025.findings-emnlp.228 (2025 (year only))
FG-PRM: Fine-grained Hallucination Detection and Mitigation in Language Model Mathematical Reasoning
tril_bfloat16 1 other paper
1 dated earlier
2403.06754 (2024-03)
ALaRM: Align Language Models via Hierarchical Rewards Modeling

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

Reinforcement Learning from Human Feedback significantly enhances Natural Language Processing by aligning language models with human expectations. A critical factor in this alignment is the strength of reward models used during training. This study explores whether stronger reward models invariably lead to better language models. In this paper, through experiments on relevance, factuality, and completeness tasks using the QA-FEEDBACK dataset and reward models based on Longformer, we uncover a surprising paradox: language models trained with moderately accurate reward models outperform those guided by highly accurate ones. This challenges the widely held belief that stronger reward models always lead to better language models, and opens up new avenues for future research into the key factors driving model performance and how to choose the most suitable reward models.

For agents

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

get_harvested_code_for_paper("2410.06554")
get_code_for_paper("2410.06554")
have("2410.06554")

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

curl https://syntology.ai/api/ran/2410.06554.json

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

[![Syntology run record](https://syntology.ai/api/ran/2410.06554.svg)](https://syntology.ai/paper/2410.06554)

Connect an agent — have() is free.