Research map / Large language models / Alignment and preferences
When Labelers Stay Silent: The Power of Ties in Cost-Effective Preference Learning
Jiaqi Lv, Zihan Zhang, Chengjing Yang, Shiyu Xia, Ning Xu, Xin Geng
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 |
|---|---|---|---|
| filter out tie pairs from GPT-4 golden labels (ablation) | DPO · UltraFeedback, Llama-3-8B | Gold Reward: 6.48 -> 6.78 | Improved |
| filter out tie pairs from GPT-4 golden labels (ablation) | DPO · UltraFeedback, Mistral-7B | Gold Reward: 4.40 -> 5.13 | Improved |
| filter out ties from Llama-3B relabeled data (ablation) | DPO · UltraFeedback, Llama-3-8B | Gold Reward: 7.94 -> 8.06 | Improved |
| model ties with symmetric tie loss instead of forced choice (ablation) | DPO with Llama-3B labeler · UltraFeedback, Llama-3-8B | Gold Reward: 7.94 -> 8.20 | Improved |
| use Llama-3B silent-aware labels instead of GPT-4 forced labels | DPO · UltraFeedback, Llama-3-8B | Gold Reward: 6.48 -> 8.20 | Improved |
| add quality tags (Both-Good/Both-Bad) with directional tie loss and corrected risk (ablation) | SymTie · UltraFeedback, Llama-3B labeler delta_M=1, Llama-3-8B | Gold Reward: 8.20 -> 9.24 | Improved |
| add deconstruct-then-KTO on quality-tagged ties (ablation) | SymTie · UltraFeedback, Llama-3B labeler delta_M=1, Llama-3-8B | Gold Reward: 8.20 -> 8.81 | Improved |
| use SymTie instead of TODO tie-aware baseline | tie handling · UltraFeedback, Llama-3B labeler delta_M=1, Llama-3-8B | Gold Reward: 8.28 -> 8.20 | No clear effect |
| increase resolution threshold delta_M (more ties) (ablation) | silent-aware methods · UltraFeedback, Llama-3B/8B labelers | Gold Reward: e.g. Llama-8B SymTie 8.23 (delta_M=1) -> 10.28 (delta_M=5) | Improved |
| increase tie ratio to 80% percentile threshold (ablation) | SymTie and C-DirTie with Qwen labelers · Llama-3-8B policy | Gold Score | Hurt |
| apply tie modeling to human-labeled data (ChatArena golden) | Forced DPO · ChatArena, Llama-3-8B | Gold Reward: 6.73 -> 9.99 | Improved |
| silent-aware objectives on Mistral-7B with ChatArena (ablation) | Filtered baseline · ChatArena, Mistral-7B, Llama-3B delta_M=5 | Gold Reward: Forced 7.15, Filtered 5.06, SymTie 7.29, C-DirTie 6.52 | Mixed |
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.
- TODO — TODO: Enhancing LLM Alignment with Ternary Preferences (ICLR 2025) · 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
- Learning Where It Matters: Geometric Anchoring for Robust Preference Alignment (NeurIPS 2026 · similarity 0.89)
- Geometric-Averaged Preference Optimization for Soft Preference Labels (NeurIPS 2024 · similarity 0.88)
- Pairwise Calibrated Rewards for Pluralistic Alignment (NeurIPS 2025 · similarity 0.88)
- On Extending Direct Preference Optimization to Accommodate Ties (NeurIPS 2025 · similarity 0.88)
- Distributionally Robust Listwise Preference Optimization (NeurIPS 2026 · similarity 0.88)
- Sail into the Headwind: Alignment via Robust Rewards and Dynamic Labels against Reward Hacking (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.
