FlowingFlowing
← Back to all development cases
Development case · blind model judgment pending independent human review; not a formal Benchmark conclusion.
AutocompleteEvidence version preferred62 / 162 · 8c7639e77f98f748

Review of Recent Advances in Gaussian Process Regression Methods

Statistics · 2409.08112v1

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

Review of Recent Advances in Gaussian Process Regression Methods

1 Introduction

…2 ] . GP is a stochastic process which defines a distribution over possible functions that fit a set of points. This method often produces a fine and precise trade-off between fitting and smoothing the data. However, like most machine learning methods, big data can pose challenges to GP methods since it contains much information. Several scalable GP prediction methods have been developed to improve scalability without compromising prediction quality. This paper will review the latest developments in the global GP approximation methods, which approximate the whole probability density function.

…in covariance matrix [ 5 , 6 ] . Instead of the cube computation cost, most global approximation GP algorithms scale linearly in the training set size. They can offer a close result with the full GP(standard GP without any approximation or factorisation) with enough inducing points [ 7 ] . However, the representation of the global patterns (long-term spatial correlations) often excludes local patterns due to their limited global inducing set [ 1 ] . Having appealing properties, GPs have been extensively developed, improved and applied to solve a wide range of artificial intelligence problems.

↓ 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

Local approximation methods offer an alternative route to scalability by partitioning the training data into subsets and training individual GP experts on each, which allows them to capture non-stationary local patterns that a single global inducing set may miss. However, this local focus comes at a cost, as naive local approximations can produce discontinuities at the boundaries of sub-regions and risk over-fitting within some local regions.

Passages supplied to the Evidence version

Understanding and Comparing Scalable Gaussian Process Regression for Big Data ↗

Different from the sparse approximations which capture long-term spatial correlations based on the global inducing set, the local approximations focus on subspace learning via local experts, thus capturing local patterns (non-stationary features). As an illustration, we apply the pure local GPs to t…

Read full passage excerpt

Different from the sparse approximations which capture long-term spatial correlations based on the global inducing set, the local approximations focus on subspace learning via local experts, thus capturing local patterns (non-stationary features). As an illustration, we apply the pure local GPs to the toy example ( 25 ) in Fig. 4 . Particularly, we partition the 120 training points into M = 10 M=10 local subsets and train individual GP experts. It is found that the naive local approximation captures local patterns at the cost of (i) producing discontinuity on the boundaries of sub-regions, and (ii) risking over-fitting in some local regions.

No Evidence

Same model and position · no retrieved passages · first output

Not preferred in this comparison

This review focuses on global approximation methods, which summarise the full training set through a reduced set of inducing variables and thereby retain a probabilistic representation of the predictive distribution.

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 = No Evidence; Output B = Evidence.

Both outputs produce coherent academic paragraphs, but B is substantially more useful for this autocomplete task. The draft explicitly states 'This paper will review the latest developments in the global GP approximation methods' and then the blank follows. The paragraph immediately after the blank discusses 'global approximation GP algorithms' and their limitations regarding 'local patterns.' Output B directly bridges this transition by introducing local approximation methods as an alternative, using exact source-supported details about partitioning data, local experts, non-stationary patterns, boundary discontinuities, and over-fitting risks—all directly from the supplied source snippet. This creates a smooth logical flow: global methods (draft) → local alternatives (B) → return to global methods with their limitations (following paragraph). Output A merely restates what the draft already explicitly says ('This paper will review... global approximation methods') and adds generic, unsourced claims about 'retain[ing] a probabilistic representation' that do not advance the narrative or bridge to the next paragraph's discussion of local pattern limitations. B requires less editing to integrate meaningfully and preserves the draft's scope while adding source-grounded content that explains why the subsequent paragraph mentions global methods' failure to capture local patterns.

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.