Research map / Large language models / Alignment and preferences
PALC: Preference Alignment via Logit Calibration
SANGHYUN LEE, Hoh Peter In
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 PALC logit calibration module | LLaMA-7B-SFT · HH-RLHF | win+0.5tie rate vs base (%): 58.17 win+0.5tie; 54.67 win | Improved |
| use PALC rather than static steering vectors | CAA · HH-RLHF | win+0.5tie rate: 77.17 win+0.5tie vs CAA | Improved |
| use PALC rather than test-time optimization | RE-Control · HH-RLHF | win+0.5tie rate: 61.67 win+0.5tie vs RE-Control | Improved |
| use PALC rather than reward-model guided decoding | ARGS · HH-RLHF | win+0.5tie rate: 55.50 win+0.5tie vs ARGS | No clear effect |
| use PALC rather than activation steering | BiPO · HH-RLHF | win+0.5tie rate: 49.00 win+0.5tie vs BiPO | No clear effect |
| use training-time DPO or GenARM instead of PALC | PALC · HH-RLHF | win+0.5tie rate: 41.17 vs DPO; 44.33 vs GenARM | Hurt |
| vary bottleneck dimension from 16 to 1024 (ablation) | PALC · HH-RLHF | win+0.5tie rate vs base (%): 53.7 to 58.2 | No clear effect |
| increase bottleneck dimension to 4096 (ablation) | PALC · HH-RLHF | win+0.5tie rate vs base (%): 58.2 -> 18.3 | Hurt |
| scale calibration with gamma 0.5-3.0 (ablation) | PALC · HH-RLHF | win+0.5tie rate vs base (%): 55.3 to 58.2 | No clear effect |
| raise scaling factor gamma to 10 (ablation) | PALC · HH-RLHF | win+0.5tie rate vs base (%): 58.2 -> 38.7 | Hurt |
| use PALC rather than reward-model guided decoding for latency | GenARM · HH-RLHF | relative latency: 3.17x -> 1.08x | Improved |
| apply PALC to MT-Bench | LLaMA-7B-SFT · MT-Bench (LC win+tie vs base) | length-controlled win+tie rate (%): 61.9 (GenARM 58.7) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- GenARM — GenARM: Reward Guided Generation with Autoregressive Reward Model for Test-Time Alignment (ICLR 2025) · better on 3 of 3 shared results
- ARGS — ARGS: Alignment as Reward-Guided Search (ICLR 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
- On-the-fly Preference Alignment via Principle-Guided Decoding (ICLR 2025 · similarity 0.92)
- Test-Time Preference Optimization: On-the-Fly Alignment via Iterative Textual Feedback (ICML 2025 · similarity 0.92)
- Latent Preference Coding: Aligning Large Language Models via Discrete Latent Codes (ICML 2025 · similarity 0.91)
- Linear Alignment: A Closed-form Solution for Aligning Human Preferences without Tuning and Feedback (ICML 2024 · similarity 0.91)
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.91)
- Panacea: Pareto Alignment via Preference Adaptation for LLMs (NeurIPS 2024 · similarity 0.91)
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.
