Researchers propose spreading inference-time sampling across multiple LoRA policies
An arXiv paper argues that allocating a fixed sample budget across several learned adapters can beat repeated sampling from one policy for reasoning tasks.
In plain words
When a large language model answers hard questions, a common trick is to generate several candidate answers and pick the best one. This paper suggests generating those candidates from several small model add-ons, called LoRA adapters, instead of repeatedly from the same model setup. The authors report this improves pass-at-k results on reasoning benchmarks.
What was announced
The paper, posted on arXiv, addresses inference-time scaling, in which a model's output quality is improved by sampling multiple responses and scoring their collective utility. The authors note that most existing methods draw these samples repeatedly from a single policy. Their alternative, which they call stratified policies, trains k separate policies, each contributing one response to a k-sample objective. The population is implemented as separate LoRA adapters on a shared language model backbone and optimized for pass@k, a metric that rewards producing at least one correct answer among k attempts.
To motivate the design, the authors analyze the linear setting and show that a population of LoRAs can reach zero best-of-k approximation error in cases where every individual LoRA still has strictly positive prediction error. Empirically, they report that stratified policies consistently improve pass@k over a single-policy baseline across multiple reasoning tasks. The work is presented as a study of how distributing samples across complementary policies can use a sample budget more effectively than repeated i.i.d. sampling from one policy. The claims rest on the authors' own experiments and analysis; the abstract does not name the tasks, model sizes, or baselines in detail, and no independent evaluation is cited.
What it does not show
- Evidence is limited to the paper's abstract; it does not specify which reasoning tasks, model sizes, or baseline configurations were used, and no independent confirmation of the results is available.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.