AgenticBBO-Bench
Harbor: pendingA cross-domain benchmark evaluating LLM agents on five black-box optimization tasks (synthetic, HPO, DB tuning, chip design, molecular) under a unified finite-budget protocol.
- Samples
- N/A
- Released
- 2026
- Scores
- 7
PersonaCUABench
PersonaCUABench: Benchmarking Computer-Use Agents Serving Diverse Personas
Evaluates computer-use agents on 74 tasks across 11 domains with 44 diverse personas to measure task completion and user satisfaction beyond fixed specifications.
How to run
ltb · personacuabench
$ npm i -g lettokenburn-cli$ ltb bench info personacuabench
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
9 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Claude Opus 5.5 | 18.6 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| GPT-6 Astra | 17.8 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| MiniMax-M3 | 16.1 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Muse Glimmer 30B | 15.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Qwen3.8 27B | 15.0 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Kimi K3 | 14.5 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| DeepSeek V4.1 Flash | 14.1 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Grok-4.6 | 13.8 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Gemini 3.8 Flash | 13.5 | 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 runsLoading community runs…
About
PersonaCUABench comprises 74 tasks in 11 domains, evaluated against 44 personas grouped into four types. A persona agent rewrites instructions and answers queries, isolating the effect of user identity on agent performance.
Scoring includes task completion against original requirements and user satisfaction ratings. Results show a 36% drop in completion relative to original requests, highlighting gaps between task solving and serving user constraints.