Research map / Large language models / Alignment and preferences
Discovering Implicit Large Language Model Alignment Objectives
Edward Chen, Sanmi Koyejo, Carlos Guestrin
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 |
|---|---|---|---|
| use full checkpoint trajectory instead of only base and final checkpoints (ablation) | Obj-Disco · HH-RLHF GRPO Llama-8B controlled | Model-Fit | Improved |
| use full checkpoint trajectory instead of only base and final checkpoints (ablation) | Obj-Disco · Misalignment case study | trials capturing misaligned behavior: 1 of 4 -> 3 of 4 | Improved |
| use residual-driven open-ended discovery instead of fixed objective set (Fixed-3) (ablation) | Obj-Disco · Misalignment case study | trials identifying misaligned objective: 23.5% -> 58.8% | Improved |
| open-ended discovery versus Fixed-15 curated set (ablation) | Obj-Disco · Misalignment case study | trials identifying misaligned objective: 44.1% -> 58.8% | Improved |
| add full-context proposer instead of limited context window (ablation) | Zero-Shot · Misalignment case study | trials identifying misaligned objective: 5.9% -> 0.0% | Improved |
| use Obj-Disco instead of Iter-Filter | DIR discovery · Objectives causality user study | selection frequency: 16.7% -> 35.6% | Improved |
| use Obj-Disco objective explanations instead of random samples | Objective Explanations · Objective Explanations user study | identification accuracy: 25.5% -> 39.9% | Improved |
| use Obj-Disco instead of Zero-Shot | DIR discovery · Controlled and open-source RM settings | Model-Fit | Improved |
Closest papers by meaning
- RMB: Comprehensively benchmarking reward models in LLM alignment (ICLR 2025 · similarity 0.92)
- Theoretical Limits of Language Model Alignment (NeurIPS 2026 · similarity 0.91)
- Aligner: Efficient Alignment by Learning to Correct (NeurIPS 2024 · similarity 0.90)
- LLM Alignment--Utility Asymmetry under Semantic-Preserving Transformations (NeurIPS 2026 · similarity 0.89)
- MM-RLHF: The Next Step Forward in Multimodal LLM Alignment (ICML 2025 · similarity 0.89)
- SR-GRPO: Stable Rank as an Intrinsic Geometric Reward for Large Language Model Alignment (NeurIPS 2026 · similarity 0.89)
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.
