Same Answer, Different Reason: What a Reasoning Model Says When It Follows a Planted Hint
I put wrong answers in front of 200 exam questions and read what a small model’s reasoning said when it fell for them.
The post Same Answer, Different Reason: What a Reasoning Model Says When It Follows a Planted Hint appeared first on Towards Data Science.
Could On-Policy Self-Distillation Revolutionize Post-Training?
Teaching foundation models new skills without breaking them is hard. A new reinforcement learning-based approach might have the power to change that.
The post Could On-Policy Self-Distillation Revolutionize Post-Training? appeared first on Towards Data Science.
Decision Models in Action: Build a Hybrid LLM Workflow
A hands-on guide to building browser automation by combining decision models with LLMs.
The post Decision Models in Action: Build a Hybrid LLM Workflow appeared first on Towards Data Science.
Grid puzzles that people solve almost perfectly show where language and vision models still break
The post Spatial Reasoning in AI Models, Taken Apart appeared first on Towards Data Science.
I Built a Pipeline to Track the Job Market I'm Trying to Break Into.
Building an incremental pipeline with dbt snapshots, and the API that quietly lied to me along the way.
The post I Built a Pipeline to Track the Job Market I'm Trying to Break Into. appeared first on Towards Data Science.
Synthetic-to-Real Transfer with Frozen DINOv3 Features
Training models on procedurally generated datasets for card detection and segmentation
The post Synthetic-to-Real Transfer with Frozen DINOv3 Features appeared first on Towards Data Science.
Skill or MCP Server? Ask What Stops the LLM From Disobeying
Towards Spec-Driven Test Automation, Part 4
The post Skill or MCP Server? Ask What Stops the LLM From Disobeying appeared first on Towards Data Science.
What's in the Box? Distinguishing Acceptable Traces from Unacceptable Ones
Turn scattered requirements into deployable agents
The post What's in the Box? Distinguishing Acceptable Traces from Unacceptable Ones appeared first on Towards Data Science.
A million-token output window changes what you can hand a model in one go, and the benchmarks and the bill show who it actually suits.
The post Welcome to Google Gemini 4 Argon appeared first on Towards Data Science.
Can Conformal Prediction Take You from Accurate to Reliable Machine Learning?
ML model adoption doesn't stop with acceptance of accuracy; it starts there.
The post Can Conformal Prediction Take You from Accurate to Reliable Machine Learning? appeared first on Towards Data Science.
Ten quantization guides, and not one threshold-dependent metric between them — what that gap hides, and the two numbers that close it.
The post The Number Your Quantization Gate Doesn't Have appeared first on Towards Data Science.
When EEG Models See the Same People in Training and Testing
How participant-aware evaluation changed performance across three machine-learning models
The post When EEG Models See the Same People in Training and Testing appeared first on Towards Data Science.
Do You Still Own the Code Your AI Wrote? I Built CodeOwnership to Measure It
I rated my confidence 4 out of 5, then couldn’t find a planted bug in 8 minutes. Here is the tool, the run, and its limits.
The post Do You Still Own the Code Your AI Wrote? I Built CodeOwnership to Measure It appeared first on Towards Data Science.
I Built a Claim-Checker for Writing With Codex. It Passed a Wrong Claim.
The tool found the right passage but missed the mismatch, and the audit shows where rule-based checks break down.
The post I Built a Claim-Checker for Writing With Codex. It Passed a Wrong Claim. appeared first on Towards Data Science.
How Many Guesses Does It Take to Crack an AI's "Random" Number?
I read the exact answer distribution from the logits of open models, base and instruction-tuned, and turned it into bits and guesses-to-crack.
The post How Many Guesses Does It Take to Crack an AI's "Random" Number? appeared first on Towards Data Science.
I Tested Four Ways to Give an AI Assistant a Memory. The Obvious One Was the Worst Value.
Pasting in the whole chat history cost 20 to 31 times more than looking up notes, and it did no better on the tasks.
The post I Tested Four Ways to Give an AI Assistant a Memory. The Obvious One Was the Worst Value. appeared first on Towards Data Science.
A Scaffold at the Gate: Building AI Applications That Can Evolve
A technical reflection on why data scientists need software engineering principles when AI agents write code.
The post A Scaffold at the Gate: Building AI Applications That Can Evolve appeared first on Towards Data Science.
Coding Agents Can Solve the Task — But How Much Work Did They Waste?
Real repositories, seeded agent decisions: what 50,400 runs show about wasted turns, patch churn, and which harness rules help
The post Coding Agents Can Solve the Task — But How Much Work Did They Waste? appeared first on Towards Data Science.
We Specified an AI Agent in 2016. We Called It Analytics.
Re-reading my own IBM patent on real-time concert feedback as what it actually was: an agentic loop with multimodal perception, memory-grounded inference, and a human in the loop. What it got right, and what a decade of hindsight teaches agent builders today.
The post We Specified an AI Agent in 201...
I Built a Trap for AI Image Judges. They Walked Into a Different One.
Three vision AI models ignored the pretty image, yet GPT-4o invented answers and Claude kept answering B.
The post I Built a Trap for AI Image Judges. They Walked Into a Different One. appeared first on Towards Data Science.
Add architectural guardrails around how agents read untrusted sources and use them in the workflow.
The post How Can AI Agents Read Untrusted Sources Safely? appeared first on Towards Data Science.
Can TypeSafe's Jev Make AI Agents Safer Without Another LLM?
How a decision-first model can catch risky tool calls before they turn into real-world actions
The post Can TypeSafe's Jev Make AI Agents Safer Without Another LLM? appeared first on Towards Data Science.