FlowingFlowing
← Back to all development cases
Development case · blind model judgment pending independent human review; not a formal Benchmark conclusion.
AutocompleteEvidence version preferred75 / 162 · 36e8a2af193b40ef

Interpretable Machine Learning for Weather and Climate Prediction: A Survey

Climate and environmental science · 2403.18864v1

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 there are 81 source-opportunity positions. The original blind review preferred Evidence in 51; a task check moved one empty Evidence continuation to both unusable, leaving 50 in public counts. In the ordinary group, another pair of empty outputs moved from no winner to both unusable. Original verdicts remain visible on case pages.

50Evidence version preferred
÷
81All positions in this group
=
62%Evidence preference in this group

Source contribution is a separate review: 16 of 21 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

Interpretable Machine Learning for Weather and Climate Prediction: A Survey

6 Research Challenges > 6.4 Interpretability for Large Foundataion Models

Explainable AI techniques provide valuable insights into the rationale behind the predictions of complex machine learning models. However, the full potential of interpretability lies in utilizing those insights to further improve model performance. The integration of explainability analysis into model development workflows can enable developers to continuously monitor feature attributions and model behaviors. This allows identifying areas where the model violates known constraints or relies on faulty correlations, guiding iterative refinement steps (Fig. 5 ).

…onable logic and relationships in the model. Once violations of physical consistency are detected through explanations, constraints can be introduced into model training and architectures. This includes adding loss penalty terms based on physical principles, using physics-infused neural networks, or representing conservation properties. Imposing appropriate physical constraints helps prevent models from making unreal predictions while still leveraging the flexibility of machine learning. This will eventually shift deep learning models closer to a hybrid data-driven and physics-based paradigm.

↓ 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: Evidence version preferred

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

Preferred by blind review

A complementary direction is to embed interpretability directly into the model architecture rather than applying it post hoc. For example, defining reduced-order models in proper orthogonal decomposition coordinates constrains predictions to a subspace defined by the governing equations, embedding physics information from high-fidelity simulation snapshots directly into the model representation.

Passages supplied to the Evidence version

Learning physics-based reduced-order models for a single-injector combustion process ↗

This paper presents a physics-based data-driven method to learn predictive reduced-order models (ROMs) from high-fidelity simulations, and illustrates it in the challenging context of a single-injector combustion process. The method combines the perspectives of model reduction and machine learning.…

Read full passage excerpt

This paper presents a physics-based data-driven method to learn predictive reduced-order models (ROMs) from high-fidelity simulations, and illustrates it in the challenging context of a single-injector combustion process. The method combines the perspectives of model reduction and machine learning. Model reduction brings in the physics of the problem, constraining the ROM predictions to lie on a subspace defined by the governing equations. This is achieved by defining the ROM in proper orthogonal decomposition (POD) coordinates, which embed the rich physics information contained in solution snapshots of a high-fidelity computational fluid dynamics (CFD) model. The machine learning perspective brings the flexibility to use transformed physical variables to define the POD basis.

No Evidence

Same model and position · no retrieved passages · first output

Not preferred in this comparison

Beyond physics-based constraints, interpretability can also inform the design of model architectures themselves. Attribution patterns that recur across forecasts may point to interactions or dependencies that the current architecture does not explicitly represent, suggesting structural modifications rather than post hoc corrections.

Judgment and task check

The blinded model review preferred the Evidence version. Its original rationale appears below.

Read the original model judgment

In this rationale, Output A = Evidence; Output B = No Evidence.

Both outputs provide coherent academic paragraphs that continue the manuscript's theme of integrating explainability into model development. However, Output A is superior because it directly leverages the supplied source about physics-based reduced-order models in POD coordinates, creating a concrete, source-grounded continuation that bridges the draft's discussion of physical constraints with the specific methodology from the source. The phrase 'defining reduced-order models in proper orthogonal decomposition coordinates constrains predictions to a subspace defined by the governing equations' directly maps to the source's description. Output B, while topically relevant and plausible, is more generic and speculative—it discusses 'attribution patterns that recur across forecasts' and 'structural modifications' without any source support, and introduces concepts (forecasting, architecture redesign based on patterns) not present in the draft or sources. Output A provides a more substantive, evidence-based continuation that a researcher could insert with less revision, while Output B's claims about what attribution patterns 'may point to' are unsupported extrapolations.

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.