AI Paper Atlas

Research map / Large language models / Alignment and preferences

Steerable Adversarial Scenario Generation through Test-Time Preference Alignment

Tong Nie, Yuewen Mei, Yihong Tang, Junlin He, Jie Sun, Haotian Shi, Wei Ma, Jian Sun

ICLR 2026 · Large language models › Alignment and preferences · approach: Reward model design and preference models

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 changedWhereMeasuredEffect
Add map-compliance as hard feasibility precondition in HGPO instead of removing it (ablation)SAGE HGPO · Map feasibility during fine-tuningmap feasibility rate: nearly 90% feasibility vs collapse without map constraintImproved
Use hard-constraint decoupling rather than weighted map penalty (ablation)SAGE HGPO · Map feasibility and rewardfeasibility and rewardImproved
Use group-based pairs (HGPO) instead of single winner-loser DPO (ablation)SAGE · Fine-tuning efficiencysample efficiency, training loss stabilityImproved
Mix expert weights rather than logits or trajectories at test time (ablation)SAGE · Pareto front adversariality vs realismrealism at given adversarialityImproved
Increase w_adv from 0.0 to 1.0 (ablation)SAGE · WOMD vs Replay policyAttack success rate: 16.26% -> 76.15%Improved
Increase w_adv from 0.0 to 1.0 (ablation)SAGE · WOMD vs Replay policyBehavioral realism penalty (lower better): 0.332 -> 1.429Hurt
Train ego with SAGE adversaries instead of CAT adversariesTD3 ego agent · Closed-loop adversarial envroute completion: 0.58 -> 0.69Improved
Train ego with SAGE adversaries instead of no adversaryTD3 ego agent · Closed-loop adversarial envcollision rate: 0.44 -> 0.31Improved
Train ego with SAGE adversaries on out-of-distribution adversary test setsTD3 ego agent · Cross-eval CAT and Rule scenarioscollision rate: 0.307 vs CAT-trained 0.298 on CAT scenariosMixed

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy