CIFQA: Deterministic Tool-Grounded Multi-Agent LLMs for Calculation-Intensive Financial Queries
CIFQA tackles the classic LLM failure mode in finance: generating answers that look right but are numerically wrong when multi-step calculations and rule constraints are involved. It splits work across specialized agents for interpretation, parameter extraction, and computation planning, then relies on deterministic Python tools for exact rate lookup and logic, hitting 95.54% accuracy on calculation-intensive fixed-deposit questions.
Trump EPA wants to let data centers hide their air pollution
The Verge reports the EPA plans to scrap a rule requiring public notice and comment when industrial sites apply for air permits, a change that could let data centers move faster without alerting or involving nearby residents. Advocates warn the rollback would reduce transparency at the exact moment local communities most need forewarning about construction and emissions impacts.
Agentic Pipelines for Explaining ICU Mortality Predictions Show Fewer Safety-Relevant Pitfalls Than Standalone LLMs
This feasibility study tests whether LLM-based explanations can be made more clinically grounded by decomposing the workflow into agentic steps that separate data interpretation from guideline checking and final explanation writing. In an eICU Demo setting, a four-step pipeline avoided explicit outcome leakage seen in the standalone LLM and improved guideline grounding, value specificity, and overall plausibility on overlapping explanation subsets.
OpenAI and Thailand’s MHESI launch an accelerator for next-generation AI startups
OpenAI and Thailand’s Ministry of Higher Education, Science, Research and Innovation (MHESI) are starting an eight-week program aimed at turning early AI prototypes into products that are ready for real-world adoption. The accelerator targets 10 health, wellness, and education startups and emphasizes trustworthiness as companies move from experimentation to deployment.
The Accuracy-Efficiency Paradox: Why More-Accurate On-Device Forecasting Can Cost Net Energy
The paper argues that higher-accuracy forecasting models can backfire in edge deployments because inference compute consumes energy and battery aging compounds the losses over time. It introduces a Total Cost of Ownership framing that treats battery degradation as a form of energy dissipation, showing that precision gains are sometimes outweighed by the system-level net energy deficit.
Anthropic was illegally blacklisted by the Trump administration, court rules
A judge ruled that the Pentagon’s blacklisting of Anthropic earlier this year was unconstitutional, concluding the government’s national-security justification wasn’t a blank check to punish the company. The Verge frames the decision as a significant win for the AI lab after months of litigation over allegations of unlawful retaliation.