FlowingFlowing
← Back to all development cases
Development case · blind model judgment pending independent human review; not a formal Benchmark conclusion.
AutocompleteEvidence version preferred77 / 162 · 0f0e76573d585a9d

Causal Inference with Complex Treatments: A Survey

Statistics · 2407.14022v1

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

Causal Inference with Complex Treatments: A Survey

9. Conclusion

The development of causal inference with continuous treatment is much better than that of multi-valued and bundle treatment. This gap is partly due to the solid foundation of IV in binary treatment, which can be easily extended as a solution of the continuous treatment problem.

However, current literature with respect to continuous treatment mainly focus on the potential outcome framework, which heavily rely on the discovering the causal graph structure, namely causal discovery. It is challenging to ensure the identifiability of causal effects in a complex graph structure, and it is tough to specify the causal graph from observational data.

↓ 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

The difficulty of recovering a unique causal graph from observational data is well documented: in many practical settings the available substantive knowledge is insufficient to specify a single causal graph, and data-driven methods instead recover a Markov equivalence class, the collection of graphs sharing the same observed features.

Passages supplied to the Evidence version

Causal Identification under Markov Equivalence ↗

Assessing the magnitude of cause-and-effect relations is one of the central challenges found throughout the empirical sciences. The problem of identification of causal effects is concerned with determining whether a causal effect can be computed from a combination of observational data and substanti…

Read full passage excerpt

Assessing the magnitude of cause-and-effect relations is one of the central challenges found throughout the empirical sciences. The problem of identification of causal effects is concerned with determining whether a causal effect can be computed from a combination of observational data and substantive knowledge about the domain under investigation, which is formally expressed in the form of a causal graph. In many practical settings, however, the knowledge available for the researcher is not strong enough so as to specify a unique causal graph. Another line of investigation attempts to use observational data to learn a qualitative description of the domain called a Markov equivalence class, which is the collection of causal graphs that share the same set of observed features.

No Evidence

Same model and position · no retrieved passages · first output

Not preferred in this comparison

Future work could explore alternative identification strategies that reduce reliance on causal discovery, such as leveraging experimental or quasi-experimental designs for multi-valued and bundle treatments. Extending the theoretical guarantees developed for continuous treatment to these more complex settings would help close the gap between the maturity of continuous-treatment methods and the relative scarcity of tools for multi-valued and bundle treatments.

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.

Output A provides a coherent continuation grounded in the supplied source. It elaborates on the challenge mentioned in the draft—specifying causal graphs from observational data—by directly incorporating the source's point about Markov equivalence classes when unique causal graphs cannot be identified. The language and concepts match the source snippet exactly. Output B, however, pivots to 'future work' on 'experimental or quasi-experimental designs for multi-valued and bundle treatments' and claims this would 'close the gap between the maturity of continuous-treatment methods and the relative scarcity of tools for multi-valued and bundle treatments.' This introduces multiple unsupported claims: (1) that experimental/quasi-experimental designs are viable alternatives for multi-valued and bundle treatments (not mentioned in sources), (2) that theoretical guarantees for continuous treatment can be extended to these settings (not supported), and (3) that there is a 'relative scarcity of tools' for multi-valued and bundle treatments (the draft states development is 'much better' for continuous treatment, but 'scarcity' is a stronger, unsupported characterization, and the proposed extension mechanism is entirely invented). The draft's focus is on the difficulty of causal discovery for continuous treatment, not on prescribing future directions for multi-valued/bundle treatments. Output B's content is a speculative research agenda with no basis in the provided source or draft.

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.