JointCoreBench
A benchmark with 6,912 exact conditions across nine interaction families and six workflow domains to evaluate behavioral and logical memory sufficiency in LLM agents.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info jointcorebench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
4 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Llama–3.3–70B NF4 | 0.6 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Gemma-4-12B-it | 0.6 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Qwen2.5 14B Instruct | 0.5 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Mistral–3.1–24B NF4 | 0.4 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
JointCoreBench features natural-language decisions with exact sufficient-set oracles, covering redundancy and conjunction interactions. It separates search error from model error using controlled tasks.
Scoring measures minimum-cost recovery and active-memory savings. The baseline JointCore achieves 100% recovery on exact conditions and reduces subset evaluations by 62.9% compared to singleton deletion.