UiPath updated its automation platform with new workflow automation and software testing features at FUSION in Las Vegas. Cartographer automates the data collection step of mapping how employees perform manual tasks, taking the process that can take weeks. UiPath will also roll out enhancements including Autonomous Exploration for Test Cloud and a new UiPath Delegate agent for test administration tasks.
Zvi (Don't Worry About the Vase)·3 weeks ago·
22
● 3 sources
Claude Opus 5.5’s system card lays out updated classifier and evaluation results, including claims about its cyber strength and evidence on alignment-risk and R&D capabilities. METR’s preliminary AI R&D estimate is about 1.5X overall acceleration from AI, with roughly a 30% chance of 2X. Anthropic says it will adjust how it tests biology by dropping “helpful-only” versions and otherwise relies on largely the same oversight/safeguards while deploying Opus 5.5 to the public with cyber controls.
Sentry’s team used agent tracing to debug Search Query Assistant eval failures caused by incorrect field syntax in AI-generated Sentry Search queries for custom numerical attributes. The bug was identified during end-to-end testing when a query intended to find events where a custom numeric field was greater than 10 returned no results. After fixing how tool-call field names were placed into the system prompt, production queries returned correct data and local eval runs were instrumented to send trace data to Sentry for easier debugging.
AI agent evaluation has shifted from grading single model outputs to grading complete task execution that includes sequential tool calls, environment state tracking, and recovery after failures. The piece highlights SWE-bench Verified traces where the reported outcome score is based on end-to-end checks while the process is scored step-by-step, including an example E2E result of 1 with step-level scoring of 3/4. Benchmarks and release gates now prioritize tool-use and environment-based verification (e.g., state changes and tests) alongside process traces, rather than treating call text accuracy as sufficient.
MentalHealthBench was introduced as an expert-informed benchmark for evaluating helpful and safe AI responses in realistic mental health conversations. The article does not provide any numeric benchmark, date, or other concrete figure. This changes how developers can measure and compare AI response quality for mental health dialogue, though the piece itself contains only a brief description.
Nokia’s applied research team open-sourced AnyJev, a Python library that converts an open LLM into a typed multiple-choice/yes-no/bin decision model without training. On Qwen3-8B using BANKING77 (20-way, 300 test items), the option order-flip rate dropped from 0.230 to 0.073 with L0 and reached 0.077 with L1. The library adds cyclic-shift and calibration layers (plus optional temperature scaling) to improve probability calibration and reduce sensitivity to option ordering when serving models via PyPI backends like Hugging Face and vLLM.
Snorkel AI raised funding as an AI training data provider, disclosing a Series E round led by Insight and S32. The company raised $350 million at a $3.5B valuation. It will use the money to hire engineers, invest in AI safety, and develop open-source model evaluation benchmarks while expanding its data-as-a-service offerings for supervised and reinforcement learning.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
