Skip to content

Domain calibration: balance recall across paper domains #93

Description

@dangng2004

Motivation

OpenAIReview's recall is uneven across paper domains. On the full 72-paper benchmark (GPT-5.5 backend), per-domain recall ranges from 91.8% (Econometrics) down to 45.7% (Complexity), with theoretical domains lagging empirical-science domains by a wide margin:

Domain OpenAIReview recall
Econometrics 91.8%
ML 84.7%
Atomic Phys. 72.0%
HEP 71.8%
Genomics 71.3%
Statistics 66.1%
Math 62.1%
Complexity 45.7%

The ~46-point spread suggests the pipeline's prompts, chunking, or section-routing are tuned implicitly toward some domains and underperform on others — particularly theoretical/proof-heavy papers (Math, Complexity).

Goal

Add domain calibration: adjust the review pipeline so per-domain recall is more uniform, without regressing on the strong domains.

Possible directions (open to contributors):

  • Domain-specific prompts or sub-agent templates (e.g. proof-checking emphasis for Math/Complexity).
  • Domain-aware section routing or chunk weighting.
  • Detect domain up-front (from arXiv category, abstract, or content) and dispatch a tailored pipeline.
  • Calibrate the filter step's threshold per domain.

Testing requirement

This change touches the core pipeline, so we need proper per-domain regression testing before merging.

Where to start

  • Pipeline: src/reviewer/method_progressive.py, method_zero_shot.py, method_local.py.
  • Skill prompts: src/reviewer/skill/references/subagent_templates.md, criteria.md.
  • Benchmark runner: benchmarks/perturbation/run_benchmark.py.
  • Evaluation: benchmarks/perturbation/score.py.

Happy to discuss design before implementation — drop ideas in this thread.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or requesthelp wantedExtra attention is needed

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions