FlowingFlowing
← Back to all development cases
Development case · blind model judgment pending independent human review; not a formal Benchmark conclusion.
AutocompleteTie22 / 162 · f0e3700def31e87e

Deciphering genomic codes using advanced NLP techniques: a scoping review

Quantitative biology · 2411.16084v1

FLOWING EVIDENCE BENCHMARK

How do we tell whether Evidence helps?

At the same writing position, with the same model and task, how does supplying retrieved paper passages change the first output? We compare matched versions and retain ties, unusable outputs and incomplete reviews.

SAME DRAFT · EVIDENCE ON OR OFF

01 · FIXED WRITING POSITION

MANUSCRIPT

Same writing position ▌

Matched autocomplete at the same draft position

02 · TWO MATCHED INPUTS

SHARED BY BOTH

Manuscript context, model, task and prompt

A · With Evidence

Retrieved paper passages supplied

B · Without Evidence

No retrieved passages supplied

LLM

Same model and version

A → first output

B → first output

Blind judge agent

First outputs are anonymized as X and Y

Continuation review: accuracy, fit to the writing task and usability
Returns: X preferred / tie / Y preferred / both unusable

Order check: X / Y → Y / X

HOW DOES BLIND JUDGING WORK?

① Anonymize both outputs
The judge sees the same draft and both first outputs without knowing which received Evidence.

② Compare and swap order
The judge applies task-specific criteria in X/Y and then Y/X order.

③ Review disagreements
A third pass resolves disagreements. Incomplete reviews remain in the denominator.

How are autocomplete positions stratified?

Before seeing generation outcomes, we check whether a retrieved passage contains a specific proposition that directly supports the next writing move. Those positions appear in the left opportunity group; the rest are ordinary positions on the right. We select a balanced sample from admitted papers in each field. The 50/50 split is experimental, not a measure of how often either type occurs in writing.

AUTOCOMPLETE · 110

Biology, statistics and astrophysics

55 positions on each side; statistics uses the ten-paper rerun.

AUTOCOMPLETE · 52

Psychology and climate science

26 positions on each side; psychology includes nine papers and climate science four.

How are the table percentages calculated?

Across five fields, 51 of 81 left-column positions preferred the Evidence version. Ties, pairs where both versions were unusable and incomplete reviews remain in the denominator.

51Evidence version preferred
÷
81All positions in this group
=
63%Evidence preference in this group

Source contribution is a separate review: 16 of 22 Evidence wins entered into source review directly used retrieved papers; another 29 wins await review.

These are development-stage model judgments pending independent human review. They are not formal Benchmark conclusions and do not, on their own, establish causality.

Manuscript writing position

Text excerpt · not a PDF page

Research manuscript · excerpt

Deciphering genomic codes using advanced NLP techniques: a scoping review

4 Discussion

Among the studies, most datasets are publicly accessible, with a few studies having limited or request-based access for specific subsets. It fosters inclusivity and sustainable development in integrating genomic data with NLP, enhancing collaboration and progress in this rapidly evolving field.

…lassifier such as XGBoost [ 14 , 28 ] or small neural networks [ 34 , 29 ] on top have minimal requirements. Fine-tuning or continuously pretraining from a publicly available model lies between these extremes [ 25 , 15 ] . In addition, some studies intentionally consider resource constraints in model design and training processes. For example, Roy et al. stopped training at 10,000 steps due to resource limitations and diminishing marginal returns to training [ 17 ] . Furthermore, Wang et al. designed a small architecture (a two-layer transformer) to fit into low-resource environments [ 35 ] .

↓ Next paragraph was generated here

The highlighted text is near the writing position. This is a layout of the evaluation input, not a PDF screenshot.

A matched comparison at the same writing positionJudgment: Tie

Both arms use the same model, task and manuscript context. Only the retrieved passages differ. The first recorded output from each arm is shown below.

Evidence

Same model and position · retrieved passages supplied · first output

These design choices suggest that computational efficiency is not merely a practical constraint but can shape the model architectures adopted in genomic NLP.

Passages supplied to the Evidence version

Medical Concept Normalization in a Low-Resource Setting

In this thesis, I explore the challenges of medical concept normalization in a low-resource setting. Specifically, I investigate the shortcomings of current medical concept normalization methods applied to German lay texts. Since there is no suitable dataset available, a dataset consisting of posts…

Read full passage excerpt

In this thesis, I explore the challenges of medical concept normalization in a low-resource setting. Specifically, I investigate the shortcomings of current medical concept normalization methods applied to German lay texts. Since there is no suitable dataset available, a dataset consisting of posts from a German medical online forum is annotated with concepts from the Unified Medical Language System. The experiments demonstrate that multilingual Transformer-based models are able to outperform string similarity methods. The use of contextual information to improve the normalization of lay mentions is also examined, but led to inferior results.

No Evidence

Same model and position · no retrieved passages · first output

These design choices suggest that computational feasibility, rather than model capacity alone, shapes how genomic NLP pipelines are configured in practice.

Blind model judgment

The blind model judge did not prefer either output.

Read the original model judgment

Order-swapped judges disagreed. Pass 1: Both outputs are coherent continuations grounded in the draft's discussion of resource constraints and model design choices. Output A introduces 'computational efficiency' and claims it 'can shape the model architectures adopted in genomic NLP.' Output B uses 'computational feasibility' and states it 'shapes how genomic NLP pipelines are configured in practice.' Both are reasonable inferences from the draft's examples (Roy et al. stopping training early, Wang et al. designing small architectures). However, Output B is preferable because 'configured in practice' more accurately captures the breadth of design choices described (training stopping points, architecture size, and the earlier mention of fine-tuning vs. classifiers), whereas Output A's 'model architectures' is slightly narrower and doesn't encompass training process decisions like stopping at 10,000 steps. Output B's 'rather than model capacity alone' also better echoes the draft's contrast between minimal-requirement classifiers, fine-tuning, and resource-conscious designs. Pass 2: Both outputs are coherent continuations grounded in the draft's discussion of resource constraints and design choices. Output B is preferable because it more accurately reflects the draft's content. The draft explicitly describes studies that 'intentionally consider resource constraints in model design and training processes' — Roy et al. stopping training early and Wang et al. designing 'a small architecture (a two-layer transformer) to fit into low-resource environments.' Output B's 'shape the model architectures adopted' directly captures this architectural adaptation described in the draft. Output A's 'rather than model capacity alone' introduces a comparison ('model capacity') that is not present or contrasted in the draft; the draft never discusses model capacity as an alternative consideration, making this a slight conceptual addition. Output B's 'not merely a practical constraint but can shape the model architectures' better preserves the draft's meaning that resource constraints actively influence design decisions.

The passages are input material; their presence does not mean an output used them or that they support every claim. Draft and source passages are excerpted; no simulated scores or PDF appear here.