Benchmark Radar Day 93: Judge What the Agent Changed, Not What It Said

Day 93 of Benchmark Radar. The scan fetched 1,280 items, published 643, and marked 320 as recommended.

Author: Koutian Wu; GitHub: ktwu01

Benchmark Radar is live. Open the dashboard, star the repository, or read the paper. The Hugging Face dataset provides the data.

V³AST-Bench evaluates coding-agent safety through 212 risk tasks in 18 categories. Its runtime environments inspect the resulting states, records, and artifacts, while paired capability tests distinguish harm avoidance from simple inability to complete a task. It measures what the agent actually changes rather than judging its written response. V³AST-Bench: Verifiable, Valid, and Versatile Safety Evaluation for Coding Agents.

Δτ-Bench tests agents under changing environments with 1,588 validated multi-turn tasks that alter tool ecosystems as services disappear, replacements emerge, and documentation becomes noisy. SkillShift-Bench independently applies controlled context, perception, and execution changes to web agents. A static test cannot show whether an agent discovers new tool paths and revises outdated procedures. Δτ-Bench: Pushing Agents Toward a Moving Capability.

Why this matters

A safety suite should not give the same credit to a model that cannot act and a model that could act but correctly abstains. And when the services under an agent keep changing, adaptation belongs in the score next to initial task completion.

Scan coverage

This is a keyword-filtered discovery feed, not a representative survey of AI research. The archived briefing used 188 of 643 captured evidence records. These findings summarize its cited releases; discovery dates do not necessarily equal original publication dates.

Data and briefing: 2026-10-27 archived snapshot.