Skip to content

ScholarEval 9-Dimension Scoring Guide

ScholarEval is this skill's internal name for a reviewer-style scoring rubric. The dimensions below follow the criteria used in OpenReview / NeurIPS / ICLR reviewer forms — they are not derived from arXiv:2510.16234, whose "ScholarEval" framework scores research ideas on two dimensions only (soundness + contribution). Only Literature Grounding is inspired by that paper's literature-grounding idea. Do not cite arXiv:2510.16234 as the source of the full rubric.

Overview

The rubric scores academic papers across 8 dimensions plus a computed overall (9 total) using a hybrid script + LLM approach. Script-based dimensions use automated issue detection; LLM-based dimensions require reading the full paper.

Dimension weights are defined once in scripts/scholar_eval.py (SCHOLAR_EVAL_DIMENSIONS); the percentages below mirror that source.

Dimensions

1. Soundness (18%, Script)

What it measures: Technical correctness, logical validity, well-supported claims.

Scoring criteria:

ScoreDescription
9-10Flawless logic, all claims rigorously supported
7-8Sound methodology with minor gaps
5-6Some logical issues or unsupported claims
3-4Significant methodological concerns
1-2Fundamental logical errors

Script source: Deductions from LOGIC module issues.

2. Clarity (13%, Script)

What it measures: Writing quality, readability, organization.

Scoring criteria:

ScoreDescription
9-10Exceptionally clear, well-organized, precise language
7-8Clear writing with minor grammar/style issues
5-6Readable but has notable style problems
3-4Difficult to follow, significant language issues
1-2Incomprehensible or severely disorganized

Script source: Deductions from GRAMMAR, SENTENCES, FORMAT, DEAI modules.

3. Presentation (8%, Script)

What it measures: Visual quality, figure/table clarity, reference integrity.

Scoring criteria:

ScoreDescription
9-10Professional figures, perfect references, clean layout
7-8Good presentation with minor issues
5-6Some figure/reference problems
3-4Poor visual quality, broken references
1-2Missing figures, numerous broken references

Script source: Deductions from FIGURES, VISUAL, REFERENCES modules.

4. Novelty (13%, LLM)

What it measures: Originality of contributions, distinction from prior work.

LLM evaluation prompt (used in SKILL.md):

Evaluate the novelty of this paper. Consider:

  • How different is the approach from existing methods?
  • Are the contributions genuinely new or incremental?
  • Does the paper clearly articulate what is novel?

Scoring criteria:

ScoreDescription
9-10Paradigm-shifting contribution
7-8Significant novelty, clearly beyond prior art
5-6Moderate novelty, some overlap with existing work
3-4Mostly incremental, limited differentiation
1-2No novelty, rehashes known approaches

5. Significance (13%, LLM)

What it measures: Potential impact on the field, community benefit.

Scoring criteria:

ScoreDescription
9-10Likely to reshape the field
7-8High impact, addresses important problem
5-6Moderate impact, useful contribution
3-4Limited impact, niche application
1-2Negligible impact

6. Reproducibility (8%, Mixed)

What it measures: Can others reproduce the results?

Script component: Checks for methodology-related issues in LOGIC module. LLM component: Evaluates experimental description completeness, code/data availability.

Scoring criteria:

ScoreDescription
9-10Complete code, data, and detailed methodology
7-8Sufficient detail for reproduction with some effort
5-6Key details present but gaps exist
3-4Difficult to reproduce, missing key information
1-2Impossible to reproduce from the paper alone

7. Ethics (5%, LLM)

What it measures: Ethical considerations, conflicts of interest, data privacy.

Scoring criteria:

ScoreDescription
9-10Thorough ethics discussion, no concerns
7-8Adequate ethics consideration
5-6Some ethical aspects overlooked
3-4Notable ethical concerns unaddressed
1-2Serious ethical violations

8. Literature Grounding (12%, Mixed) -- NEW in v3.0

What it measures: How well the paper is grounded in and positioned within existing literature.

Script component: Automated literature search coverage analysis via literature_compare.py. LLM component: Domain Reviewer's qualitative literature assessment.

Requires: --literature-search flag for script component.

Scoring criteria:

ScoreDescription
9-10Comprehensive coverage, thematic organization, clear gap derivation
7-8Solid coverage with minor gaps, adequate organization
5-6Reasonable coverage but notable gaps, weak gap identification
3-4Incomplete coverage, enumerated rather than thematic
1-2Minimal or absent literature review

See references/LITERATURE_GROUNDING_GUIDE.md for detailed scoring rubric.

9. Overall (10%, Computed)

Weighted average of all available dimension scores, normalized for missing values.

Publication Readiness Scale

ScoreLabelMeaning
9.0+Strong AcceptReady for top venue submission
8.0-8.9AcceptPublication ready with confidence
7.0-7.9Minor RevisionsReady after addressing minor issues
6.0-6.9Major RevisionsSignificant improvements needed
5.0-5.9Significant ReworkSubstantial revision required
<5.0Not ReadyNot suitable for submission

Deduction Rules (Script Dimensions)

Issues detected by automated checks reduce scores from a base of 10:

SeverityDeduction
Critical-2.5
Major-1.25
Minor-0.5

Minimum score is 1.0 (floor).

LLM Evaluation JSON Format

When using --llm-json, provide a file with this structure:

json
{
  "novelty": {
    "score": 7.5,
    "evidence": "The paper introduces a novel attention mechanism..."
  },
  "significance": {
    "score": 8.0,
    "evidence": "Addresses a critical gap in real-time processing..."
  },
  "reproducibility_llm": {
    "score": 6.5,
    "evidence": "Code is promised but not yet released..."
  },
  "ethics": {
    "score": 9.0,
    "evidence": "No ethical concerns identified..."
  }
}

Relationship to NeurIPS 4-Dimension Scoring

ScholarEval is complementary to the existing 4-dimension NeurIPS scoring:

  • NeurIPS: Quality, Clarity, Significance, Originality (1-6 scale)
  • ScholarEval: 8 scoring dimensions + computed overall (1-10 scale)

Both systems run independently and appear in the report side by side. Enable ScholarEval with the --scholar-eval flag.

Released under the MIT License.