FlowingFlowing
← Back to all development cases
Development case · blind model judgment pending independent human review; not a formal Benchmark conclusion.
AutocompleteNo-Evidence version preferred15 / 162 · 2efbfb8491689b52

A Unified View of Deep Learning for Reaction and Retrosynthesis Prediction: Current Status and Future Challenges

Quantitative biology · 2306.15890v1

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

A Unified View of Deep Learning for Reaction and Retrosynthesis Prediction: Current Status and Future Challenges

4 Limitations and Challenges > 4.3 Limitations in Evaluation

…ugh, which only contains 50k backward reactions. Considering many recent approaches only have slight numerical differences in top-k accuracies, the current small-scale dataset is not adequate for testing model capability. Second, the current dataset will bias edit predictions and leaving group selections. Most backward reactions only have one single edit while very few have multiple edits, which results in poor prediction accuracy in multiple-edit cases. Additionally, leaving group distribution is very imbalanced, which makes graph-based models tend to select a few frequently occurring atoms.

…aluations since most of the reaction types in the testing set are already covered in the training set. Therefore, harder dataset split like scaffold splits and time splits should be taken into account for future evaluations of reaction predictions. Scaffold split is testing whether the model can generalize well under out-of-distribution settings, in which the training data distribution is very different from testing data distribution. Time split is splitting reactions in the order of discovery time. This split aims at testing whether the model can truly discover new chemical reactions or not.

↓ 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: No-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

Not preferred in this comparison

These limitations are not unique to retrosynthesis prediction; benchmarks of temporal distribution shift in other domains likewise report substantial performance drops when evaluation moves from in-distribution to out-of-distribution data, underscoring the need for evaluation protocols that reflect realistic deployment conditions.

Passages supplied to the Evidence version

Wild-Time: A Benchmark of in-the-Wild Distribution Shift over Time

To address this gap, we curate Wild-Time, a benchmark of 5 datasets that reflect temporal distribution shifts arising in a variety of real-world applications, including patient prognosis and news classification. On these datasets, we systematically benchmark 13 prior approaches, including methods in…

Read full passage excerpt

To address this gap, we curate Wild-Time, a benchmark of 5 datasets that reflect temporal distribution shifts arising in a variety of real-world applications, including patient prognosis and news classification. On these datasets, we systematically benchmark 13 prior approaches, including methods in domain generalization, continual learning, self-supervised learning, and ensemble learning. We use two evaluation strategies: evaluation with a fixed time split (Eval-Fix) and evaluation with a data stream (Eval-Stream). Eval-Fix, our primary evaluation strategy, aims to provide a simple evaluation protocol, while Eval-Stream is more realistic for certain real-world applications. Under both evaluation strategies, we observe an average performance drop of 20% from in-distribution to out-of-distribution data.

No Evidence

Same model and position · no retrieved passages · first output

Preferred by blind review

Third, evaluation protocols often rely on random splits, which can overestimate performance because test reactions may share scaffolds or reaction types with the training set. Adopting distribution-shift-aware splits, such as scaffold or time splits, would provide a more realistic assessment of generalization to unseen chemistry.

Blind model judgment

The blinded model review preferred the no-Evidence version. Its original rationale appears below.

Read the original model judgment

Output A continues the manuscript's argument coherently: it introduces a third limitation (evaluation protocols with random splits) that parallels the draft's 'First...Second...' structure, then proposes scaffold/time splits as solutions—directly flowing from the draft's final sentence about 'harder dataset split like scaffold splits and time splits.' The content is grounded in the draft's own discussion of distribution shifts and evaluation methodology. Output B introduces a cross-domain comparison to 'benchmarks of temporal distribution shift in other domains' that report 'substantial performance drops.' The source (Wild-Time) does discuss temporal distribution shifts and 20% performance drops, but the source is about patient prognosis and news classification—not chemistry or retrosynthesis. The claim that these limitations are 'not unique to retrosynthesis prediction' and that other domains 'likewise report substantial performance drops' constitutes an unsupported cross-domain generalization. The source does not establish that retrosynthesis prediction shares these patterns, nor does the draft mention other domains. The 'substantial performance drops' phrasing also loosely echoes the source's 20% figure without proper grounding. This is an unsupported claim that misleads by implying the source validates a chemistry-domain parallel.

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.