Research map / Label-efficient and robust learning / Recognition and long-tail
Debate Helps Weak-to-Strong Generalization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Combines debate-generated arguments from a strong Qwen-14B with a weak-model ensemble (Qwen-7B) to produce better weak labels for weak-to-strong generalization, outperforming finetuning baselines on four NLP classification tasks.
Problem
Weak human or weak model supervision limits alignment of superhuman models; the paper asks whether scalable oversight (debate) and weak-to-strong generalization can be combined so a strong model's debate helps a weak supervisor produce better labels for training a strong student.
Approach
Two instances of a strong pretrained model debate opposing answers; the transcripts are appended to inputs when training a weak model ensemble on ground-truth labels, and the ensemble's averaged predictions serve as weak labels for finetuning the strong student.
Key results
- Debate ensembles reach 76.5% PGR on SciQ, 69.2% on BoolQ, 56.5% on CosmosQA, and 70.0% on AnthropicHH, versus 41.2%, 56.4%, 17.4%, and 35.0% for the best finetune baseline (aux. loss).
- Debate outperforms consultancy and market-making variants on accuracy and PGR across all four tasks.
- Debate ensembles beat single weak models and finetune ensembles on all four tasks (e.g., SciQ PGR 76.5 vs 50.0 single weak model).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use debate-generated context to train the weak supervisor instead of naive weak supervision | weak-to-strong training · SciQ | PGR: 44.1 -> 76.5 | Helped |
| use debate-generated context to train the weak supervisor instead of naive weak supervision | weak-to-strong training · BoolQ | PGR: 56.4 -> 69.2 | Helped |
| use debate-generated context to train the weak supervisor instead of naive weak supervision | weak-to-strong training · CosmosQA | PGR: 17.4 -> 56.5 | Helped |
| use debate-generated context to train the weak supervisor instead of naive weak supervision | weak-to-strong training · AnthropicHH | PGR: 35.0 -> 70.0 | Helped |
| replace debate with consultancy (ablation) | weak-to-strong training · SciQ, BoolQ, CosmosQA, AnthropicHH | PGR: 76.5 vs 44.1; 69.2 vs 46.2; 56.5 vs 34.8; 70.0 vs 25.0 | Hurt |
| replace debate with market-making (ablation) | weak-to-strong training · SciQ, BoolQ, CosmosQA, AnthropicHH | PGR: 76.5 vs 47.1; 69.2 vs 41.0; 56.5 vs 30.4; 70.0 vs 35.0 | Hurt |
| use debate ensembles instead of a single weak model (ablation) | weak-to-strong training · SciQ, BoolQ, CosmosQA, AnthropicHH | PGR: 50.0 -> 76.5; 56.4 -> 69.2; 39.1 -> 56.5; 35.0 -> 70.0 | Helped |
| vary only finetuning seeds while sharing debate transcripts instead of using diverse debate transcripts (ablation) | weak model ensemble · SciQ, BoolQ, CosmosQA, AnthropicHH | PGR: 52.9 vs 76.5; 59.0 vs 69.2; 39.1 vs 56.5; 45.0 vs 70.0 | Hurt |
| increase ensemble cardinality from 3 to 4 members (ablation) | debate ensemble · SciQ and AnthropicHH | test accuracy | Helped |
| increase ensemble cardinality beyond 4 members (ablation) | debate ensemble · SciQ and AnthropicHH | test accuracy | No clear effect |
| increase debate length beyond 3 turns (ablation) | debate ensemble · SciQ and AnthropicHH | test accuracy | Hurt |
| add an auxiliary confidence loss | Finetune strong student · SciQ, BoolQ, CosmosQA, AnthropicHH | PGR: 44.1 -> 41.2; 51.3 -> 56.4; 30.4 -> 17.4; 35.0 -> 35.0 | Mixed |
Surprising findings
- Adding a confidence loss to finetuning gives only modest gains, and on CosmosQA the aux. loss baseline is worse than plain finetuning in PGR (17.4 vs 30.4).
- Increasing debate turns beyond 3 does not improve performance; strong debaters degrade in accuracy as transcripts grow longer.
Limitations
- Strong and weak models differ only in size (7B vs 14B), a small compute gap that may not reflect true superhuman settings.
- The debate-based approach is expensive, requiring two debaters and multi-turn generation, plus ensemble training.
- Evaluated only on four binary classification NLP tasks converted from QA and preference datasets.
Experimental setup
Datasets: SciQ BoolQ CosmosQA AnthropicHH
Benchmarks: SciQ BoolQ CosmosQA AnthropicHH
Metrics: accuracy performance gap recovered (PGR)
Baselines: Finetune Finetune w/ aux. loss Finetune w/ pro. loss Consultancy Market-Making Single weak model Finetune ensembles
Builds on: weak-to-strong generalization debate consultancy market-making auxiliary confidence loss
Closest papers by meaning
- Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision (ICML 2024 · similarity 0.95)
- Strong Empowered and Aligned Weak Mastered Annotation for Weak-to-Strong Generalization (AAAI 2025 · similarity 0.92)
- Super(ficial)-alignment: Strong Models May Deceive Weak Models in Weak-to-Strong Generalization (ICLR 2025 · similarity 0.92)
- Representations Shape Weak-to-Strong Generalization: Theoretical Insights and Empirical Predictions (ICML 2025 · similarity 0.91)
- Weak-to-Strong Generalization under Distribution Shifts (NeurIPS 2025 · similarity 0.91)
- Quantifying the Gain in Weak-to-Strong Generalization (NeurIPS 2024 · similarity 0.90)
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.