Research map / Large language models / Alignment and preferences
Contrastive Weak-to-Strong Generalization
Houcheng Jiang, Junfeng Fang, Jiaxin Wu, Tianyu Zhang, Chen Gao, Xiang Wang, Xiangnan He, Yang Deng
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 |
|---|---|---|---|
| add training-time contrast to ConG-S sampling to form ConG (ablation) | ConG-S (w->s) · Qwen2.5-7B weak-to-strong | Avg. of AlpacaEval2 and Arena-Hard win rates: 48.0 -> 52.7 | Improved |
| add training-time contrast to ConG-S sampling to form ConG (ablation) | ConG-S (w->s) · Llama3-8B weak-to-strong | Avg. of AlpacaEval2 and Arena-Hard win rates: 36.7 -> 41.6 | Improved |
| use ConG instead of SimPO in self-alignment | SimPO · Qwen2.5-3B self-alignment | Avg. of AlpacaEval2 and Arena-Hard win rates: 41.5 -> 45.5 | Improved |
| use ConG instead of SimPO in self-alignment | SimPO · Llama3.2-3B self-alignment | Avg. of AlpacaEval2 and Arena-Hard win rates: 33.5 -> 34.6 | Improved |
| use ConG instead of WSPO for weak-to-strong | WSPO · Qwen2.5-7B weak-to-strong | Avg. of AlpacaEval2 and Arena-Hard win rates: 25.6 -> 52.7 | Improved |
| use ConG instead of Weak SFT (post) | Weak SFT (post) · Llama3-8B weak-to-strong | Avg. of AlpacaEval2 and Arena-Hard win rates: 27.2 -> 41.6 | Improved |
| use ConG instead of WeakTeacher (weak model outputs) | WeakTeacher · Qwen2.5-7B weak-to-strong | Avg. of AlpacaEval2 and Arena-Hard win rates: 43.5 -> 52.7 | Improved |
| use cross-family weak model (Llama3.2-3B) instead of same-family | ConG · Qwen2.5-7B strong model | Avg. of AlpacaEval2 and Arena-Hard win rates: 52.7 -> 51.7 | No clear effect |
| use GPT-5 judge instead of GPT-4 | ConG · Qwen2.5-7B weak-to-strong | Avg. of AlpacaEval2 and Arena-Hard win rates: 52.7 -> 47.3 | No clear effect |
| apply ConG instead of Base model | Qwen2.5-7B-Instruct · GSM8K | accuracy: 80.44 -> 80.21 | No clear effect |
| apply ConG instead of Base model | Llama3-8B-Instruct · GSM8K | accuracy: 75.28 -> 75.66 | No clear effect |
| apply Weak SFT instead of Base model | Qwen2.5-7B-Instruct · GSM8K | accuracy: 80.44 -> 75.28 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SimPO — SimPO: Simple Preference Optimization with a Reference-Free Reward (NeurIPS 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Quantifying the Gain in Weak-to-Strong Generalization (NeurIPS 2024 · similarity 0.91)
- Weak-to-Strong Search: Align Large Language Models via Searching over Small Language Models (NeurIPS 2024 · similarity 0.91)
- A transfer learning framework for weak to strong generalization (ICLR 2025 · similarity 0.90)
- Weak-to-Strong Preference Optimization: Stealing Reward from Weak Aligned Model (ICLR 2025 · similarity 0.89)
- Escaping Collapse: The Strength of Weak Data for Large Language Model Training (NeurIPS 2025 · similarity 0.88)
- Bayesian WeakS-to-Strong from Text Classification to Generation (ICLR 2025 · similarity 0.88)
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.
