yogsoth-ai Artifact Detection Tactic
Systematically probe benchmarks for annotation artifacts, dataset shortcuts, and spurious correlations that allow models to achieve high scores without the intended capability.
Stages
Stage 1: Hypothesis-Only Baseline Test
Search literature for evidence that partial-input baselines achieve unexpectedly high performance:
- Hypothesis-only baselines (NLI without premise)
- Question-only baselines (QA without context)
- Label-word frequency baselines
- Majority-class and surface-pattern baselines
Search queries: "[benchmark] annotation artifacts", "[benchmark] hypothesis only", "[benchmark] spurious correlations", "[benchmark] dataset bias"
If published partial-input results exist, record performance gap between partial and full input. Gap < 10 points above random indicates severe artifacts.
Stage 2: Contrast Set Construction
Identify whether contrast sets or adversarial evaluations exist:
- Search for "[benchmark] contrast sets", "[benchmark] adversarial examples"
- Check if CheckList-style behavioral tests have been applied
- Look for counterfactual data augmentation studies
Record performance drops on contrast sets. Drops > 20 points indicate reliance on surface patterns.
Stage 3: Format Manipulation Probes
Search for evidence of format sensitivity:
- Prompt template sensitivity studies
- Label name/ordering effects
- Verbalization effects in classification
- Input length correlations with labels
Record whether minor format changes cause disproportionate score changes.
Stage 4: Conclusion Synthesis
Aggregate evidence into artifact severity assessment:
| Severity | Criteria |
|---|---|
| Critical | Partial-input baseline within 5 points of full model |
| High | Contrast set drop >20 points OR format sensitivity >10 points |
| Medium | Known artifacts documented but partial mitigations exist |
| Low | Minor artifacts, full-input still required for high performance |
| None | No evidence of artifacts (may indicate insufficient probing) |
Output
artifact_report:
benchmark: string
overall_severity: critical|high|medium|low|none
partial_input_baselines:
- input_type: string # e.g., "hypothesis only"
performance: float
full_model_performance: float
gap: float
source: string
contrast_set_results:
- contrast_set: string
original_performance: float
contrast_performance: float
drop: float
source: string
format_sensitivity:
- manipulation: string
score_range: string
source: string
shortcuts_identified:
- shortcut: string
mechanism: string
exploitability: high|medium|low
evidence_completeness: thorough|partial|minimal
Yield Report
| Metric | Minimum |
|---|---|
| Literature sources checked | 5 |
| Artifact categories probed | 3 |
| Evidence items collected | 4 |
| Severity classification produced | 1 |