CommunityJev Practices & Patterns
FastJEV: Understanding Redundancy for Compact JEV Inference (arXiv 2610.11379)
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.
- Format
- Article
- Published by
- Community
- Last verified
- Oct 2026
Highlights
- Identifies repeated matching token histories, duplicate inference states, and full backbone execution as redundancy sources.
- Shared context anchoring reuses recurrent initial states and omits unused final recurrent caches.
- Candidate prefix sharing retains intermediate states needed by subsequent branches.
- Decision-guided pruning uses relative score changes measured on a small unlabeled set.
- Reduces candidate depth 43.75% to 45.83% while retaining 93.66% to 97.52% of task scores on average.
Watch out
arXiv preprint submitted October 8, 2026 (cs.CV) by Jie Ma and seven co-authors. Results are self-reported across OmniJev models and six evaluation sets; the abstract notes candidate prefix sharing can reduce repeated computation while increasing latency, and names no code release.
Questions about this entry
Short answers drawn from the linked project or page.
Even though the common context is encoded once, candidate evaluation can repeat matching token histories, duplicate inference states, and execute the full backbone, according to the paper.
Shared context anchoring reuses recurrent initial states and omits unused final recurrent caches; candidate prefix sharing retains intermediate states for later branches; decision-guided pruning then uses relative score changes measured on a small unlabeled set.
At the selected pruning budgets the complete method reduces candidate depth by 43.75% to 45.83%.
Across six evaluation sets and three OmniJev model sizes, it retains 93.66% to 97.52% of the original task scores on average.
No. The method retains full context encoding and all candidates without additional training, according to the paper.
They show how candidate overlap and branching structure affect the execution cost of history reuse, and the implementation finds candidate prefix sharing can reduce repeated computation while increasing latency.
The findings motivate designing sharing granularity and execution schedules together, since sharing computation requires preserving states for later branches.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: A one-token logprob wrapper that also classifies webcam frames.
allanrbo151 pointsA single function Jev-like wrapper for LLMs, including vision modelsRead the thread on Hacker NewsMore Jev resources like this
Studies a fast learned actor in closed-loop Doom that hands control to a reasoning vision-language model only when a gate opens. On 900 held-out questions, deferring the least confident 30% of decisions gains +0.13 with doomLaya's option order, but no closed-loop variant reaches the exit.
Jev Practices & PatternsArticle#community#research#arxiv
Community
dexmac221/doomgemmaPython
dexmac221/doomgemmaPythonReference-bound Visual Jev rewards verify whether requested attributes, actions, and relations hold for specified reference subjects in multi-subject image generation. Training a Qwen3.5-4B verifier as a GRPO reward raises a GPT-5.4 composite score from 41.78 to 52.50 on a 897-task MICo-Bench subset.
Jev Practices & PatternsArticle#community#research#arxiv
Community
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.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
A theoretical study of recalibrating many related yes/no questions that each have few labels, motivated by decision models such as Jev. It derives minimax compound-risk rates and shows pooled isotonic fitting with one per-question correction achieves them, while tuning questions separately loses a factor.
Jev Practices & PatternsArticle#community#research#arxiv
Community
A test of Jev on reflection tests, one-shot matrix games, ALFWorld, and robot control that finds a sharp boundary: it succeeds when the right option can be judged from what the input describes, but fails when the right option depends on predicting something not in the input.
Jev Practices & PatternsArticle#community#research#arxiv
Community
A position and case-study paper on using System One decision models like Jev and the open-source Laya as decision layers in LLM-driven penetration-testing harnesses. It names four decision points, runs an exploratory 13-vulnerability NeuroSploit comparison, and proposes Rave, an unevaluated model.
Jev Practices & PatternsArticle#community#research#arxiv
Community
From the Jev guides
Original write-ups on what this entry is for.
Training recipes, looped architectures, and language ports: what the open side of the typed-decision class shipped this month, and how to pick by runtime and calibration instead of the leaderboard.
Jev Models & Reimplementations#open-models#training#local
Read article
More than eighty preprints now evaluate, attack, and reproduce the typed-decision class. The broadest board, the coherence gap, the rejection bottleneck, and the attacks that still work.
Jev Practices & Patterns#research#evaluation#calibration
Read article
A Choice can only pick from its options. How rejection options, candidate coverage, and escalation bands keep a narrow taxonomy honest under inputs it never anticipated.
Jev Practices & Patterns#rejection#confidence#patterns
Read article
From the Jev community
Posts from builders shipping with Jev right now.