Jev Efficiency

A training-free method that studies redundancy in candidate evaluation for JEV-style multimodal decisions. Shared context anchoring, candidate prefix sharing, and decision-guided pruning cut candidate depth by 43.75%–45.83% while retaining 93.66%–97.52% of task scores.

Encodes an image and public context once, runs isolated question suffixes as a batch, and reads candidate probabilities from the backbone's language-model head. Answer-supervised post-training lifts equal-weight macro accuracy from 70.6% to 76.1% across four benchmarks.

Trains a small decision head label-free to recover about 99% of a frozen language model's multiple-choice accuracy using two-thirds of its transformer layers. Its +1T variant reaches 75.0% mean accuracy across eight benchmarks versus 75.7% for the full model, with higher batched vLLM throughput.

A semantic database system that answers filters, joins, classification, and ranking with fast typed decision models and escalates uncertain cases to LLMs. JEVDB-Flash posts the lowest latency on all 21 SemBench queries, and Semantic Bloom Filters remove 87.4% of join candidates on Shelob.

A vision-language navigation framework that replaces multimodal generation with visual compression and typed action selection: Action-Centric Visual Compression builds action representations, Discriminative Action-Semantic Memory keeps them distinct. On R2R-CE it reaches 27.0% SR and 22.4% SPL at 0.65 s per step.

An integration of Jev's decision API into edge service orchestration that extracts four to eight bounded intent fields per request. Across 33 test conditions Jev cuts median decision latency by 22.7-64.5% versus the fastest LLM, and on a live admission path it keeps 0.91-0.95 of requests exact and on time.