Cover image

A Pulse Wave Is Not a Patient History

PPG-LM demonstrates strong waveform-language understanding while exposing the limits of inferring patient history and clinical events from physiological monitoring.

October 27, 2026 · 10 min · Zelina
Cover image

A Repair Log Is Not an Explanation

THEMIS shows how recording AI-generated code repairs makes workflow failures inspectable, without establishing that the recorded explanations are correct.

October 27, 2026 · 9 min · Zelina
Cover image

Asking Better Does Not Mean Specifying Better

A 341-task benchmark shows why coding agents must be evaluated separately on requirement elicitation, specification accuracy, and implementation.

October 27, 2026 · 8 min · Zelina
Cover image

The Return on Remembering: Reusing UAV Channel Models

LiLib shows how compact model libraries can help UAV communication systems recover from recurring environmental changes without repeatedly relearning familiar conditions.

October 27, 2026 · 9 min · Zelina
Cover image

The Search Stack Changes the Answer

A multilingual browsing benchmark shows why model choice, retrieval integration, and execution reliability must be evaluated together.

October 27, 2026 · 8 min · Zelina
Cover image

The Smallest Client Needs More Than a Better Backbone

Federated fine-tuning experiments show how backbone choice, aggregation, personalization, and evaluation determine whether data-poor clients benefit.

October 27, 2026 · 9 min · Zelina
Cover image

Translation Has More Than One Definition of Correct

Index-Translate shows why multilingual coverage, instruction compliance, controlled dubbing, and long-document consistency require different training and evaluation strategies.

October 27, 2026 · 9 min · Zelina
Cover image

Code Translation Is Not a Round Trip

ORCA shows why AI-generated data science migrations fail across library directions and pipeline interfaces, and what teams should validate before deployment.

October 26, 2026 · 8 min · Zelina
Cover image

Relevance Is Not Verification: Choosing an ESG RAG Model

A seven-model ESG benchmark shows why relevant, source-grounded AI answers still require factual validation before business use.

October 26, 2026 · 8 min · Zelina
Cover image

The Pedal Override Is Feedback, Not a Command

How driver accelerator overrides can teach adaptive cruise control to anticipate individual overtaking preferences without confusing prediction with safe vehicle control.

October 26, 2026 · 8 min · Zelina