Research map / Large language models / Alignment and preferences
AURA: Affordance-Understanding and Risk-aware Alignment Technique for Large Language Models
Sayantan Adak, Pratyush Chatterjee, Somnath Banerjee, Rima Hazra, Somak Aditya, Animesh Mukherjee
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 self-critique to PRM-guided selection (best-of-8) (ablation) | AURA · SituationAfford safety rate | safety rate: 0.34 -> 0.52 (@2 vs w/o self-critique, Qwen2.5-7B) | Improved |
| increase number of sampled trajectories from 2 to 8 (ablation) | AURA · SituationAfford safety rate | safety rate: 0.52 -> 0.71 (Qwen2.5-7B) | Improved |
| add AURA reward-guided decoding | instruction-tuned 7-9B LLMs · SituationAfford safety rate | safety rate: 0.28 -> 0.71 (Qwen2.5-7B) | Improved |
| add self-critique only (no PRM selection) (ablation) | base LLMs · SituationAfford safety rate | safety rate: 0.28 -> 0.34 (Qwen2.5-7B) | Improved |
| add full AURA pipeline | instruction-tuned LLMs · CoSafe / STREAM | ASR: 0.15 -> 0.08 (CoSafe, Qwen2.5-7B) | Improved |
| add self-critique on top of PRM (ablation) | AURA · SituationAfford ASR | ASR: 0.33 -> 0.22 (Qwen2.5-7B) | Improved |
| train AffordRanker as step-level PRM vs existing PRMs | step-level classification · SituationAfford step-level | F1: 0.68 -> 0.88 (balanced coherence vs PathFinder-PRM-7B) | Improved |
| apply AURA at inference | Llama-3.1-8B · MMLU / ARC-Challenge | accuracy: 65.6-66.2 -> 65.2-65.9 (MMLU) | No clear effect |
Closest papers by meaning
- Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons (NeurIPS 2025 · similarity 0.89)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.89)
- Towards Safe Reasoning in Large Reasoning Models via Corrective Intervention (ICLR 2026 · similarity 0.89)
- AlphaAlign: Incentivizing Safety Alignment with Extremely Simplified Reinforcement Learning (ICLR 2026 · similarity 0.88)
- Pragma-VL: Towards a Pragmatic Arbitration of Safety and Helpfulness in MLLMs (ICLR 2026 · similarity 0.88)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · 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.
