Four new benchmarks for collaborative work, coding and laboratory understanding
Explore SlackBench, PersonaCodeBench, ServerBench and BEAKER with bilingual explanations, reported results and original publication links.
Find evaluations for evidence retrieval, personalized coding, efficient LLM serving and laboratory understanding. Open the detail pages to compare tasks, metrics and experimental conditions.
Choose the task you need to evaluate
Each detail page includes task explanations, source-linked results and original publication links. These evaluations answer different questions, so start with the task and then inspect its settings.
| Benchmark | Task |
|---|---|
| SlackBench | Collaborative search across messages and code |
| PersonaCodeBench | Coding with personalized user requirements |
| ServerBench | Designing specialized LLM serving systems |
| BEAKER | Understanding laboratory scenes and procedures |
- SlackBenchModel reports
- PersonaCodeBenchModel reports
- ServerBenchModel reports
- BEAKERModel reports
SlackBench · original paper · PersonaCodeBench · original paper · ServerBench · original paper · BEAKER · original paper
SlackBench: Find the evidence needed to answer a workplace question
SlackBench asks agents to recover context across conversations, code repositories and documents. Its 140 queries include search, abstaining on questions that cannot be answered, targeted summaries and inference from circumstances. The detail page connects the reported results with the agent harness and original dataset.
- SlackBenchModel reports
PersonaCodeBench: Does working software also satisfy this user?
PersonaCodeBench evaluates multi-turn coding sessions in which user requirements are revealed progressively. It checks software correctness and persona-specific requirements separately, with joint success subject to budget and validity conditions. The page explains what the reported joint score requires.
- PersonaCodeBenchModel reports
ServerBench: Can a coding agent deliver an efficient, verified service?
ServerBench lets coding agents optimize complete LLM serving systems. Its objective values tokens that satisfy latency requirements relative to GPU cost. Results depend on the deployment and workload, and independent verification checks the submitted system. Read the deployment conditions alongside each score.
- ServerBenchModel reports
BEAKER: Can a model identify what to do in a laboratory scene?
BEAKER contains 1,000 expert-curated questions spanning eleven image and video task types. It tests action grounding, apparatus, states, safety and procedure reasoning in an offline, single-turn setting. These scores describe answers to laboratory questions; they do not establish reliable physical robot execution.
- BEAKERModel reports
Added today, with original publication dates preserved
November 5 is the date these pages were added and reviewed. The original arXiv records date SlackBench and PersonaCodeBench to November 3, and ServerBench and BEAKER to November 4. Radar transcribes the authors’ reported results and does not claim to have rerun these experiments.
SlackBench · original paper · PersonaCodeBench · original paper · ServerBench · original paper · BEAKER · original paper