can-jev-bayes
Study of Jev's choices in multi-armed bandits, testing whether Bayesian instructions and supplied statistics improve its exploration.
Overview
can-jev-bayes studies how well Jev makes sequential decisions under uncertainty and whether Bayesian methods can help. It starts with stationary Bernoulli multi-armed bandits with 2 to 15 arms, where an agent must balance exploring uncertain options against exploiting known good ones, and compares Jev's choices against well-understood Bayesian and, where feasible, Bayes-optimal strategies.
Ten experiments plus a batching validation are complete on jev-1.13.0, totalling 598,080 evaluation and control questions across fixed-state probes and 1,560 independent online environments. Reported findings include that Jev could use external decision calculations effectively, with zero observed local decision loss over 4,000 decisions when exact action values were supplied, and that instructions explaining the value of exploration changed behaviour without reliably producing optimal planning.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which arm should be pulled next? | arm options (2-15) |
How it works
- Present bandit state and history to Jev
- Optionally add Bayesian instructions or computed statistics
- Jev picks an arm
- Measure regret against Bayesian references
Strengths
- Large, systematic experiment set
- Clear mathematical baselines
- Separates instructions from supplied computations
Caveats
- Specific to bandit tasks and jev-1.13.0
At a glance
- Type
- Project
- Status
- research
- Runtime
- experiment harness
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
How well can Jev make sequential decisions under uncertainty, and how can Bayesian methods help it learn and act more effectively?
logicrw/awesome-jev-projects
Frequently asked questions
Can Jev explore well by itself?
Instructions helped some choices but did not reliably produce optimal planning.
Does supplying Bayesian values help?
Yes, with exact action values supplied it had zero observed local decision loss over 4,000 decisions.
How large is the study?
598,080 questions across probes and 1,560 online environments.
Related projects
Experiment testing whether a local Qwen read via label log-probabilities can replace Jev on synthetic AIOps alert decisions.
ProjectresearchPythonLocal Jev-style decision model on Qwen3.5-4B with four stress tests that show where one-pass decisions break.
ProjectresearchPythonLogitTrail (formerly Mini Jev): local workbench that reads typed Choice, Noul and Score decisions from a frozen LLM's logits.
ProjectresearchExperimental harness for typed JSON decisions from DiffusionGemma by picking the likeliest allowed tokens from final logits.
BenchmarkresearchPython- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython Research decoder that treats Jev as a language model, assembling replies from repeated choices over phrases, words and letters.
ProjectresearchPython