TBNAI
From Pareto to Preference: Personalized Test-Time Scaling via Amortized Agentic Policy Discovery
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 06/10/2026, 20:00:00 UTC
Briefing created: 05/11/2026, 13:08:18 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 06:18:24 UTC.
Briefings
A new framework called PersonTTS enables personalized test-time scaling by discovering executable controllers that satisfy multiple user requirements simultaneously. Experiments on AIME and HMMT datasets demonstrate that PersonTTS significantly outperforms existing baselines in joint requirement satisfaction, while reducing the time and cost of policy discovery through amortized agentic search.
Key points
- PersonTTS is an amortized agentic policy-discovery framework that reuses prior search experience to improve policy quality while reducing discovery-agent time and cost.
- PersonTTS substantially outperforms strong TTS baselines in joint requirement satisfaction on unseen user profiles and held-out problems.
Context
Test-time scaling improves large language model reasoning by allocating additional inference computation. Existing approaches optimize accuracy against one resource dimension at a time, whereas user requirements are multidimensional, favoring different controllers. This research addresses that gap by formulating Personalized Test-Time Scaling to discover executable controllers maximizing joint satisfaction of user-specific requirements.
Limitations
Experiments were conducted on AIME and HMMT datasets. Results are limited to these specific benchmarks and do not generalize to other domains or user profiles without further validation. The study does not provide causal claims about the underlying mechanisms of the discovered controllers, only empirical performance comparisons.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 05/11/2026, 13:08:18 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeieotlwcfnipxash7f2vdovpj3pd4und7irpzhltcenap4wemqxzjy