Clairic Hit 98% Accuracy on 72.8% of Questions via Confidence Scores
With a 0.9 confidence cutoff, Cloudflare's System One model handles the easy cases and routes the rest to a bigger, slower LLM like Luna.
The post Clairic Hit 98% Accuracy on 72.8% of Questions via Confidence Scores appeared first on Towards Data Science.
Five Guardrails That Limit the Damage an AI Agent Can Do
AI Agents for Beginners, Part three: strict Pydantic arguments, a tool allowlist, read-only SQLite, a step cap, and confirmed writes, all verified by five pytest tests in 0.68s.
The post Five Guardrails That Limit the Damage an AI Agent Can Do appeared first on Towards Data Science.
Analytics Engineering's Missing Layer Is Business Communication
Why grains, joins, filters, and tests need business translations, and how a four-step framework keeps stakeholders from misreading metric changes.
The post Analytics Engineering's Missing Layer Is Business Communication appeared first on Towards Data Science.
Neurosymbolic Agents Explained, With a Z3 Example in Python
Starting from Air Canada's chatbot case, a tour of Logic-LM, Agent-C, and Amazon's checks shows the solver is sound but the English-to-logic step is not.
The post Neurosymbolic Agents Explained, With a Z3 Example in Python appeared first on Towards Data Science.
When Qini Rewards the Wrong Uplift Model, Use Doubly Robust Scores
On 1 million simulated customers, calibrated Qini scored a Linear DR learner 0.0856 against 0.0826 for the true-CATE ranking that should have won.
The post When Qini Rewards the Wrong Uplift Model, Use Doubly Robust Scores appeared first on Towards Data Science.
Google ADK Agents That Pause, Resume, and Wait Hours for a Human
A synchronous call kept an RSS agent's connection open for 83 seconds; webhooks and 202 receipts free the client while the agent works.
The post Google ADK Agents That Pause, Resume, and Wait Hours for a Human appeared first on Towards Data Science.
After 10 Heads in a Row, Small LLMs Bet Up to 82% on an 11th
Telling eight base and instruction-tuned open models that the coin is fair and the flips are independent trims their streak bias by only a few points.
The post After 10 Heads in a Row, Small LLMs Bet Up to 82% on an 11th appeared first on Towards Data Science.
Local LLMs Handle Textbook RAG but Fail on New Research Papers
Five setups, one blind grading pack: textbook scores ran 12 to 20 of 20, but a paper published this year dropped them to 0 to 9 of 15.
The post Local LLMs Handle Textbook RAG but Fail on New Research Papers appeared first on Towards Data Science.
Computing Effective Dimension for Quantum ML Models in PyTorch
Local and global variants, Fisher estimation, epsilon-ball sampling, and log-space stability, with PennyLane code and a time-series experiment.
The post Computing Effective Dimension for Quantum ML Models in PyTorch appeared first on Towards Data Science.
Data From 100+ AI Agents Shows They Fail by Silent Drift, Not Errors
Quality and token alerts were early signals for 38% of agents and errors for about 1%, so monitoring should start with quality, cost, and latency.
The post Data From 100+ AI Agents Shows They Fail by Silent Drift, Not Errors appeared first on Towards Data Science.
Keyword Search Tied an 8B Embedding Model at Picking Agent Tools
On 200 ToolRet questions and up to 3,794 tools, BM25 ran about 40x faster than BGE-large on just a CPU, and full tool docs made it the best setup tested.
The post Keyword Search Tied an 8B Embedding Model at Picking Agent Tools appeared first on Towards Data Science.
What AI Agents Speed Up in PhD Research, and What They Can't
A PhD student in ML and ecology on why faster code, experiments, and drafts don't fix data with a backstory, and three predictions for how labs will change.
The post What AI Agents Speed Up in PhD Research, and What They Can't appeared first on Towards Data Science.
I Mutation-Tested 50 Python Libraries. Tests in the Young Ones Missed 41% of Planted Faults.
Across 2,312 planted faults the study also surfaced four real library problems, from a Windows data-loss bug to README examples that do not work as written.
The post I Mutation-Tested 50 Python Libraries. Tests in the Young Ones Missed 41% of Planted Faults. appeared first on Towards Data Science.
“Are You Sure?” Flipped Qwen2.5-1.5B’s Right Answer 2 Times in 3
Three small open models, 150 MMLU questions each, five pushback styles: “That’s wrong” did more damage, and naming a wrong option was close to fatal.
The post “Are You Sure?” Flipped Qwen2.5-1.5B’s Right Answer 2 Times in 3 appeared first on Towards Data Science.
A Jev of All Trades: Build and Train a Jev-style Decision Model From Scratch
Trained on six PathMNIST tissues with a small KEEP adapter, it handled added labels without retraining, though a linear classifier hit 88% on familiar ones.
The post A Jev of All Trades: Build and Train a Jev-style Decision Model From Scratch appeared first on Towards Data Science.
Scikit-Learn Matched or Beat Jev's Calibration on BANKING77
Jev needs no training and at most three examples per intent; the open rebuild needs labels. Both still fall into the none-of-the-above trap.
The post Scikit-Learn Matched or Beat Jev's Calibration on BANKING77 appeared first on Towards Data Science.
Copilot Caught 8 of 8 Diff Bugs With Rules, 0 of 2 Outside the Diff
A guard on negative enrollment days looked correct while the function it called overbilled a customer 11 days, and an invariant test would have caught it.
The post Copilot Caught 8 of 8 Diff Bugs With Rules, 0 of 2 Outside the Diff appeared first on Towards Data Science.
Continuous Batching Beats Static and Dynamic by Up to 2.91× in vLLM
For anyone running an inference server: a zero-variance control, five traffic traces, and an engine-vs-scheduler split show where the gains come from.
The post Continuous Batching Beats Static and Dynamic by Up to 2.91× in vLLM appeared first on Towards Data Science.
Studies of short stories, metaphors, and 486 participants point one way, an 800-person experiment points the other, and the author weighs both.
The post Ask AI to Refine Your Ideas, Not Invent Them appeared first on Towards Data Science.
Quantile Regression Caught the House Price OLS Missed by $55K
On 88 house sales, OLS predicted $164K for a $219K home, while a 90% interval from 5th and 95th percentile models spanned $136K to $227K.
The post Quantile Regression Caught the House Price OLS Missed by $55K appeared first on Towards Data Science.
Change One Monty Hall Rule and Five LLMs Still Say Switch
Even with step-by-step reasoning, the models landed near 33% accuracy, and one quoted the host's rule correctly before ignoring it.
The post Change One Monty Hall Rule and Five LLMs Still Say Switch appeared first on Towards Data Science.
Timestral, t0, Tabby, and Morpheus-2.0 Compared on GIFT-Eval
All four time-series foundation models scale, patch, and predict quantiles; the real split is how series share information.
The post Timestral, t0, Tabby, and Morpheus-2.0 Compared on GIFT-Eval appeared first on Towards Data Science.
Build Your First Multi-Tool AI Agent with Ollama and Python
AI Agents for Beginners, Part 2: The agent picks between three database tools, uses each result to choose its next call, and stops at an eight-turn circuit-breaker if it gets stuck.
The post Build Your First Multi-Tool AI Agent with Ollama and Python appeared first on Towards Data Science.
Give Your Python AI Agent Persistent Memory with One JSON File
Three Python classes, an OpenRouter key, and gpt-4.1-mini give the agent a calculator, a stub web search, and conversation recall that survives restarts.
The post Give Your Python AI Agent Persistent Memory with One JSON File appeared first on Towards Data Science.