Cover image

The Creative Gap: AI Can Generate Options, but Humans Still Change the Rules

A new perspective on AI creativity separates rapid generation from the harder capabilities of evaluation, redirection, and changing the creative problem itself.

August 21, 2026 · 8 min · Zelina
Cover image

The Player Slot Matters: Why Entity Structure Changes Soccer Action Spotting

A soccer action-spotting study shows that preserving player identity through the model can matter more than simply adding features or temporal processing.

August 21, 2026 · 8 min · Zelina
Cover image

Same Proposition, Different Stance: Grammar as a Model-Risk Variable

Controlled rewrites show that LLM judgments can move with linguistic form, making prompt structure a robustness variable rather than a cosmetic choice.

August 20, 2026 · 7 min · Zelina
Cover image

Two Efficient Attentions, One Denominator Problem

ELSAA shows that combining sparse and low-rank attention requires correcting how their separately normalized outputs are scaled, not merely adding two efficient branches.

August 20, 2026 · 7 min · Zelina
Cover image

When the Test Window Changes the Problem

A time-series model can win or lose because the evaluation window suppresses the zero-occurrence behavior that operations actually depend on.

August 20, 2026 · 7 min · Zelina
Cover image

Higher Pass Rate, More Broken Tasks: The Regression Tax in Agent Skill Libraries

Skill libraries can raise average agent performance while breaking workflows that already worked; paired evaluation reveals how large that reliability cost can be.

August 19, 2026 · 7 min · Zelina
Cover image

Important, but Not Direct: When Time-Series Attribution Misstates Model Dependencies

Why large attribution scores in forecasting models can reflect mediated autocorrelation or off-manifold sensitivity rather than a direct model dependency.

August 19, 2026 · 8 min · Zelina
Cover image

The Trace Has the Answer, Not the Alternative: Agentic-DPO for Offline Agent Training

Agentic-DPO shows how expert traces can supervise the mistakes an agent is likely to make, without requiring full online rollouts during training.

August 19, 2026 · 8 min · Zelina
Cover image

Hide the Worker, Keep the Geometry: What SynthSite Changes About Privacy-Aware Safety Video

SynthSite shows why safety-video anonymization should be judged by preserved task geometry and human-grounded hazard accuracy, not visual concealment or baseline-model consistency alone.

August 18, 2026 · 7 min · Zelina
Cover image

The Reviewer Was Right. The Workflow Still Failed.

Multi-agent oversight improves only when valid critique changes the work that actually moves forward.

August 18, 2026 · 7 min · Zelina