Question 1

What is the core primitive here?

Accepted Answer

Order a small window of candidates for a query. A generative LLM answers that by decoding a permutation; a Jev-style decision model answers it with one forward pass over k candidate paths — zero decoded tokens and a calibrated distribution instead of a bare ordering.

Question 2

What does it plug into?

Accepted Answer

Two published rerankers: BlitzRank's tournament graph, which decides the cheapest set of comparisons to order a pool, and Reranker-Guided Search, which picks which documents in a corpus are worth spending reranker budget on. Both call the same JevPredictor.decide() with one question encoding, ordering rule and accounting.

Question 3

What are the headline numbers?

Accepted Answer

BlitzRank on T2Ranking dev BM25 top-100 pools (100 queries, k=8): nDCG@10 0.6309 after training versus 0.2162 for the generative listwise baseline on the same 0.6B weights, at 33.5 vs 790.1 ms per match and 0 vs 50,044 decoded tokens. Swapping Jev into RGS gave 8.6–9.4x speedups on 30 queries.

Question 4

How was the model trained?

Accepted Answer

Qwen3-0.6B decision heads on the T2Ranking train split: BM25 top-100 candidates with 4-level qrels became 54,264 Choice questions, with the checkpoint selected on a train-split holdout. Evaluation is only on the dev split — real BM25 pools with graded labels, a disjoint qid space never read during training.

Question 5

How much does training change the model?

Accepted Answer

Materially in this setup: before training the tournament scores 0.0881 nDCG@10, and 0.6309 after. A single forward pass over the whole pool without the tournament reaches 0.6256 but costs 550.1 ms per match, which is why the README keeps the tournament.

Question 6

Is the comparison to the generative baseline fair?

Accepted Answer

The README describes the same 0.6B weights decoding a RankGPT prompt locally, no external API called anywhere, BlitzRank's own selectors driven through its CompareOracle, and RGS following Algorithm 1 of the paper including its Reranker@k budget accounting.

Question 7

What are the limits of these results?

Accepted Answer

They are single-benchmark numbers on one dataset family with a 0.6B backbone, reported by the project itself; our copy of the README was truncated inside the quick-start block, so no install or training commands are repeated here.

JevRanker

Highlights

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community