AI Paper Atlas

Research map / Large language models / Alignment and preferences

Safety Game: Inference-Time Alignment of Black-Box LLMs via Constrained Optimization

Tuan Nguyen, Long Tran-Thanh

ICML 2026 · Large language models › Alignment and preferences · approach: Steering and correcting LLM behavior · 0 citations (OpenAlex)

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
replace best per-candidate threshold filter with aggregate-risk LP selector on identical Llama-Guard scores (ablation)Safety Game · SORRY-Bench (Llama-Guard-3-8B scorer)prompts with safe response selected: 338/440 -> 419/440Improved
use Safety Game selector instead of Consensus-Game-style rankingLlama-2/3 models · SafetyBenchaccuracy: LLaMA-2-7B 45.71 (G) -> 54.79Improved
use Safety Game selector with a weak self-probe scorerLlama-3.2-1B · SafetyBenchaccuracy: 44.50 (G) -> 39.00Hurt
use Safety Game selector with a weak self-probe scorerLlama-3.2-1B · SORRY-Benchexpected human fulfillment rate: 0.220 (ER-G) -> 0.415Hurt
use Safety Game selector instead of best baselineLLaMA-2-7B · SORRY-Benchexpected human fulfillment rate: 0.272 (SC) -> 0.051Improved
use Safety Game selector instead of G/MIGPT-OSS-20B · SORRY-Benchexpected human fulfillment rate: 0.009 (G) -> 0.223Hurt
use Safety Game selector instead of DLLaMA-2-13B · HHHaccuracy: 75.1 (D) -> 69.2Hurt
use Safety Game selector instead of best baselineLlama-3.1-8B · HHHaccuracy: 65.6 (MI) -> 67.9Improved
use Safety Game selector instead of best baselineGPT-OSS-20B · TruthfulQABLEU-Acc: 52.1 (ER-D) -> 54.7Improved
use Safety Game selector instead of SCLLaMA-2-13B · TruthfulQABLEU-Acc: 51.1 (SC) -> 50.9No clear effect
aggregate-risk LP versus filter-then-rank baseline with same candidates and scores (ablation)Safety Game · SORRY-BenchHFR, safe-count, refusal qualityImproved

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