AI EVALUATION, IN CONTEXT.BY EQUALITY MACHINE

Five new benchmark pages: agent safety, cloud operations and video

Open five new benchmark pages with original sources, task explanations and reported results, including Argus-Bench and CloudOpsBench.

Explore gateway security, panoramic video, road-work perception, biological model reconstruction and real cloud operations. Start with the task you need to evaluate, then inspect the source and protocol in its dedicated detail page.

Choose the work you need to test

Each link opens a complete benchmark page with bilingual explanations, original publication links and reported results. The metrics describe different tasks; read each protocol before comparing models.

BenchmarkTaskRead with the score
Argus-BenchGateway prompt-injection detectionAll-attack detection and separate false positives
PanoVideoBenchPanoramic video understandingOverall accuracy plus input and training protocol
WorkIngRoad-work perception and planningDaytime NDS; night, simulation and planning separately
Morpheus Agent BenchmarkScientific model reconstructionFour-mode biological score, judge and repetition counts
CloudOpsBenchCloud operationsCompleted-task counts and live trial rates separately

Argus-Bench · original publication · PanoVideoBench · original publication · WorkIng · original publication · Morpheus Agent Benchmark · original publication · CloudOpsBench · original publication

Argus-Bench: Can an agent gateway detect hostile instructions without flagging benign requests?

Four gateway configurations are tested on 204 attacks, with 60 benign controls reported separately. Inspect attack families and false positives together; the Rebuff result covers only its offline heuristic layer.

Argus-Bench · original publication

PanoVideoBench: Can a video model follow an event outside its initial field of view?

A 1,000-question benchmark tests appearance, spatial reasoning, action, tracking and events in panoramic video. The paper compares full-data accuracy across models and input/training variants; its annotation release does not bundle the videos.

PanoVideoBench · original publication

WorkIng: Can perception and planning cope with a highway work zone?

Pairs real road-work recordings with configurable CARLA scenes. The primary series is daytime real-world NDS across 15 training configurations. Night-time, simulated detection and planning outcomes stay in separate tables, and the current repository is a placeholder.

WorkIng · original publication

Morpheus Agent Benchmark: Can a coding agent rebuild a published multicellular model?

Uses 15 published studies and a four-hour job budget. The authors report biological scores judged against paper-specific criteria. Pooled Morpheus-enabled results exclude the Python/web-only baseline; models have different repetition counts and an LLM judge supplies the assessment.

Morpheus Agent Benchmark · original publication

CloudOpsBench: Can an agent repair real cloud infrastructure?

Provides 45 AWS operations tasks with a shell interface and a 40-minute budget. The dated author analysis reports two completed-task counts. A separate table preserves 18 live leaderboard configurations with model, effort, harness and resolution rate; the exact run dates are not published.

CloudOpsBench · original publication

Added today, with original publication dates preserved

November 3 is the date these Radar pages were added and checked. PanoVideoBench and WorkIng papers were submitted on October 30; the Argus dataset and the dated CloudOps analysis were published November 2. Morpheus keeps its benchmark version, code-release version and published experiment settings separate. All numeric values are transcribed from the cited sources; Radar did not rerun these experiments.

Argus-Bench · original publication · PanoVideoBench · original publication · WorkIng · original publication · Morpheus Agent Benchmark · original publication · CloudOpsBench · original publication

References

  1. Argus-Bench · original publication
  2. PanoVideoBench · original publication
  3. WorkIng · original publication
  4. Morpheus Agent Benchmark · original publication
  5. CloudOpsBench · original publication