From Pareto to Preference: Personalized Test-Time Scaling via Amortized Agentic Policy Discovery
Xinglin Wang, Zishen Liu, Tong Zheng, Shaoxiong Feng, Peiwen Yuan, Yiwei Li, Jiayi Shi, Yueqi Zhang and 4 others
Beijing Institute of Technology · Xiaohongshu Inc
Published October 7, 2026 · on the feed October 8, 2026 · ▲ 16 · arXiv 2610.09684
The abstract
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. Yet user requirements are multidimensional: users may specify accuracy, latency, and inference-cost requirements jointly, and different requirements can favor different controllers. We formulate Personalized Test-Time Scaling as discovering executable controllers that maximize the joint satisfaction rate of user-specific requirements. To reduce the overhead of repeated policy discovery for new user profiles, we propose PersonTTS, an amortized agentic policy-discovery framework that reuses prior search experience through requirement-matched controller initialization and source-distilled procedural guidance, while retaining target-profile evaluation for every candidate. Experiments on AIME and HMMT show that PersonTTS substantially outperforms strong TTS baselines in joint requirement satisfaction on unseen user profiles and held-out problems. Under the same candidate-evaluation budget, cross-user experience reuse further improves policy quality while substantially reducing discovery-agent time and cost.
The Brief
2:00A short podcast summary of the paper. Written and voiced by AI.
Transcript
From Pareto to Preference. Personalized Test-Time Scaling via Amortized Agentic Policy Discovery. By Xinglin Wang and eleven others, from the Beijing Institute of Technology and Xiaohongshu. Published on the seventh of October.
Test-time scaling improves how well large language models reason by spending more computation at inference. Work on making it efficient has mostly optimised accuracy against one resource at a time, pushing out either the accuracy and cost frontier or the accuracy and latency frontier. Users do not think along one axis. They may set accuracy, latency and inference cost requirements together, and different combinations of those requirements can favour different controllers.
The paper reframes the goal. Personalized Test-Time Scaling is defined as discovering executable controllers that maximise the joint satisfaction rate of a particular user's requirements. Success is no longer a point on a curve, but whether all of one user's conditions are met at once.
Discovering a policy for every new user profile from scratch would be expensive, so the proposed framework, PersonTTS, amortises the search. It is an agentic policy-discovery framework that reuses prior search experience in two ways: controllers are initialised by matching against earlier requirements, and procedural guidance is distilled from the source search. What it does not skip is evaluation. Every candidate is still evaluated against the target profile.
Experiments use AIME and HMMT. On unseen user profiles and held-out problems, PersonTTS substantially outperforms strong test-time scaling baselines in joint requirement satisfaction. Under the same candidate-evaluation budget, reusing experience across users further improves policy quality, while substantially cutting the time and cost of the discovery agent.
The move is from a frontier to a person. A Pareto curve describes what is achievable, but a user arrives with a specific mix of demands, and PersonTTS treats meeting that mix as the objective, while making each new search cheaper by learning from the searches before it.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free