Research map / Large language models / Instruction and fine-tuning
Submodular Optimization for Minimal Augmentation in Robust Language Model Alignment
Ching-Chia Kao, Chia-Mu Yu, Chun-Shien Lu, Chu-song Chen
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 submodular log-det selection instead of random selection (ablation) | augmentation selection · Qwen2.5-7B harmful prompts | ASR: 17.6% -> 0.0% | Improved |
| Use submodular log-det selection instead of Greedy Farthest Distance (ablation) | augmentation selection · Qwen2.5-7B harmful prompts | ASR: 4.4% -> 0.0% | Improved |
| Use submodular selection instead of Greedy Farthest on Mistral-7B (ablation) | augmentation selection · Mistral-7B harmful prompts | ASR: 8.9 -> 0.0 | Improved |
| Use submodular selection instead of Greedy Farthest on Llama-3.1-8B (ablation) | augmentation selection · Llama-3.1-8B harmful prompts | ASR: 13.3 -> 4.4 | Improved |
| Increase cyclic augmentation order from n=1 to n=2 (ablation) | cyclic augmentation · Phi-3-mini inverse-alignment fine-tuning | robustness slope magnitude: -0.40 -> -0.21 | Improved |
| Increase cyclic augmentation order from n=1 to n=4 (ablation) | cyclic augmentation · Phi-3-mini inverse-alignment fine-tuning | robustness slope magnitude: -0.40 -> -0.11 | Improved |
| Increase cyclic order n=1 to n=4 on Qwen-2.5-7B (ablation) | cyclic augmentation · Qwen-2.5-7B inverse-alignment fine-tuning | robustness intercept R0: 0.73 -> 0.84 | Improved |
| Add third context to diminishing returns check, B and C vs empty context (ablation) | augmentation A marginal gain · Phi-3-mini | marginal gain of A: 0.25 -> 0.11 | Hurt |
| Use log-det surrogate with sentence embedding features (ablation) | log-det surrogate · ranking quality of subsets | NDCG: 0.9701 | Improved |
| Use gradient embedding features instead of representation difference (ablation) | log-det surrogate · ranking quality of subsets | Spearman: 0.2771 -> 0.0213 | Hurt |
| Sweep sigma from 0.01 to 1000 (ablation) | log-det noise parameter · Qwen2.5-7B harmful prompts | ASR: 7.4% -> 4.4% -> 6.6% | No clear effect |
| Use biased ground set Gbad instead of random ground set Grand1 with LOGDET (ablation) | ground set choice · ground set ablation | ASR: 0.470 -> 0.565 | 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.
- CARE — CARE: Commonsense-Aware Emotional Response Generation with Latent Concepts (AAAI 2021) · better on 4 of 4 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
- Safety Depth in Large Language Models: A Markov Chain Perspective (NeurIPS 2025 · similarity 0.89)
- Align-Pro: A Principled Approach to Prompt Optimization for LLM Alignment (AAAI 2025 · similarity 0.89)
- Fail-Closed Alignment for Large Language Models (NeurIPS 2026 · similarity 0.89)
- Alignment of Large Language Models with Constrained Learning (NeurIPS 2025 · similarity 0.89)
- One-Shot Safety Alignment for Large Language Models via Optimal Dualization (NeurIPS 2024 · similarity 0.89)
- Oracle-Robust Online Alignment for Large Language Models (NeurIPS 2026 · 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.
