Research map / Large language models / AI and society
Scalable Oversight via Partitioned Human Supervision
Ren Yin, Takashi Ishida, Masashi Sugiyama
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine ordinary and complementary labels with inverse-variance weights (IVW) instead of ordinary only | accuracy estimator · MMLU-Pro, MedQA, GPQA, MATH | gap to oracle / variance | Improved |
| use IVW instead of fixed 0.5 weighting (ablation) | accuracy estimator · five LLM benchmarks | within-run std: MATH within-run std 2.52 -> 2.37 | Improved |
| use complementary labels alone with the same count as ordinary labels | accuracy estimator · MMLU-Pro | std across seeds: 1.73 -> 12.49 | Hurt |
| increase complementary labels to variance-matched size (Comp-Var) | accuracy estimator · MMLU-Pro | std across seeds: 12.49 -> 2.15 | Improved |
| use ML instead of IVW (ablation) | accuracy estimator · MMLU-Pro, MATH, EDINET | accuracy estimate: 77.94 vs 77.97 (MMLU-Pro) | No clear effect |
| use complementary-only estimator on 16-class low-accuracy task | accuracy estimator · EDINET-Bench | std across seeds: Ord 5.59 vs Comp-(K-1) 8.12 | Hurt |
| extend EDINET-Bench from 496 to 3,269 samples | accuracy estimator · EDINET-Bench | std across seeds: Ord 5.59 -> 2.46 | Improved |
| use optimal IVW weights when nc is below no or above 6 no (ablation) | accuracy estimator · MedQA-USMLE | variance | Improved |
| use optimal IVW weights when nc about 3 no (ablation) | accuracy estimator · MedQA-USMLE | variance | No clear effect |
| use complementary-label fitness in agent search (ADAS, AFlow) | ADAS, AFlow · GPQA, MATH-MC, Medical Abstracts | accuracy | Improved |
Closest papers by meaning
- Debate Helps Weak-to-Strong Generalization (AAAI 2025 · similarity 0.88)
- Human Expertise in Algorithmic Prediction (NeurIPS 2024 · similarity 0.85)
- On scalable oversight with weak LLMs judging strong LLMs (NeurIPS 2024 · similarity 0.85)
- Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision (ICML 2024 · similarity 0.84)
- Easy-to-Hard Generalization: Scalable Alignment Beyond Human Supervision (NeurIPS 2024 · similarity 0.84)
- A No Free Lunch Theorem for Human-AI Collaboration (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
