From Pareto to Preference: Personalized Test-Time Scaling via Amortized Agentic Policy Discovery
PersonTTS reuses prior search to find test-time controllers matching each user's accuracy, latency, and cost goals.
The paper frames Personalized Test-Time Scaling as finding executable controllers that jointly satisfy user-specific accuracy, latency, and inference-cost requirements. PersonTTS amortizes that search by reusing prior experience through requirement-matched controller initialization and source-distilled procedural guidance, while still evaluating every candidate on the target profile. On AIME and HMMT it substantially outperforms strong test-time scaling baselines in joint requirement satisfaction for unseen user profiles and held-out problems, and cross-user reuse improves policy quality while cutting discovery-agent time and cost under a fixed evaluation budget.
- Personalized TTS seeks controllers meeting accuracy, latency, and cost together.
- PersonTTS reuses prior search via requirement-matched controller initialization.
- Source-distilled procedural guidance cuts repeated policy-discovery cost.
- Every candidate is still evaluated on the target user profile.
- AIME and HMMT results beat strong TTS baselines on unseen profiles.
From the original article
1 min read at huggingface.co
Test-time scaling (TTS) improves the reasoning capabilities of large language models by allocating additional inference computation. Existing approaches to improving TTS efficiency largely optimize accuracy against one resource dimension at a time, advancing either the accuracy--cost or accuracy--latency Pareto frontier.
By Xinglin Wang, Zishen Liu, Tong Zheng · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.