Single-pass eligibility scoring via pointer-head multiple-choice inference (no decoding)
Convert eligibility determination into a multiple-choice scoring problem and use a pointer head to assign scores to candidate natural-language options in a single forward pass. This avoids autoregressive verdict generation by directly producing option logits for a fixed set of answers.
Discovered 2026-11-10 · Sources checked 2026-11-10
Why it matters
In settings that require scoring many patient–trial pairs under privacy constraints, single-pass scoring can reduce the need for sequential decoding while still producing a ranked or selected eligibility verdict from natural-language options.
What changed
Instead of generating an eligibility verdict token-by-token, the method uses a LoRA-adapted Qwen model with a pointer head to score predefined answer options in one pass, explicitly attributing the efficiency advantage to avoiding decoding.
Where it helps
- First-pass prioritization in clinical trial screening
- Eligibility or classification tasks with predefined natural-language answer options
- Privacy-restricted environments where external API calls are undesirable
- Large-scale candidate filtering where decoding latency is costly
- Local inference setups using adapter-based fine-tuning
Evidence & reproduction
Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.
Limitations & open questions
Claims and evaluation are reported on synthetic patients and implementation-specific timing; clinical use is not validated, and measured performance is author-reported rather than independently verified here.
Related World Models papers
Sources
- TrialJev: Open-Weight Single Pass Eligibility Scoring for Patient Trial Screening ↗
arXiv · primary · Checked 2026-11-10
