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
ServeLearnBench
ServeLearnBench: How Well Can Agents Self-Improve from Serving Experience?
A benchmark with 7,718 tasks across retail, banking, and sales to evaluate agent self-improvement from serving experience in evolving environments.
How to run
ltb · servelearnbench
$ npm i -g lettokenburn-cli$ ltb bench info servelearnbench
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 | |||
|---|---|---|---|---|---|---|
| Kimi K3 | 51.1 | skillopt | Independent | medium | Oct 6, 2026 | arxiv.org |
| GLM-5.3 | 39.5 | skillopt | Independent | medium | Oct 6, 2026 | arxiv.org |
| GLM-5.3 Flash | 35.4 | skillopt | Independent | medium | Oct 6, 2026 | arxiv.org |
| DeepSeek V4.1 Flash | 33.2 | skillopt | Independent | medium | Oct 6, 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
ServeLearnBench comprises 53 environment windows and 7,718 tasks (4,508 serving, 3,210 test) across retail support, banking, and sales. It uses an evolving-environment streaming dataset format where agents must infer latent policies from interaction feedback.
Scoring uses hidden-case reward (accuracy). Key findings show a substantial gap between capability and learning, high adaptation costs, and exploration as a bottleneck for effective continual improvement.