Cover image

Relevant Is Not Authorized: Put Identity Before Agent Memory Retrieval

Bio-MemArt shows how biometric authorization can restrict access to persistent KV-cache memory before semantic retrieval while preserving the low-prefill advantage of KV reuse.

October 1, 2026 · 7 min · Zelina
Cover image

The First Scribble Does Most of the Work

Interactive PET/CT segmentation gains are sharply front-loaded, suggesting that workflow design should optimize the first correction before adding more review rounds.

October 1, 2026 · 6 min · Zelina
Cover image

The Last Layer Is Not the Last Word on Speech Quality

CAL-MOS shows why speech-quality systems should treat representation depth and layer fusion as deployment choices rather than fixed defaults.

October 1, 2026 · 7 min · Zelina
Cover image

Where You Pause Changes What You Forget

Masked Boundary Pause reframes pause tokens as a fine-tuning control for improving specialization while reducing avoidable loss of pretrained behavior.

October 1, 2026 · 6 min · Zelina
Cover image

Before the Test Comes the Question: The LLM Formulation Gap in Analytics

StatFormBench shows why analytics copilots need to validate the problem, variables, and their roles before they automate statistical execution.

September 30, 2026 · 8 min · Zelina
Cover image

Don’t Make Every Camera Remember the Whole Show

A Dual-Transformer study suggests that multi-camera recommendation improves when recent editing history is modeled separately from the camera choices being evaluated.

September 30, 2026 · 7 min · Zelina
Cover image

Free Will Without Randomness: A Practical Test for AI Agency

Christian List’s framework turns artificial free will from a metaphysical puzzle into three separable questions about agency, alternatives, and control.

September 30, 2026 · 8 min · Zelina
Cover image

Grammar Before Language: What Multilingual LLMs Decide First

Mechanistic evidence suggests multilingual LLMs can commit to target grammar before choosing the target language’s surface tokens, giving evaluation teams a more precise way to diagnose translation failures.

September 30, 2026 · 7 min · Zelina
Cover image

One Token, More Than One Memory

MoME shows how language models can add context-sensitive sparse memory capacity while keeping token lookup cheap and runtime overhead bounded.

September 30, 2026 · 8 min · Zelina
Cover image

When Grounding Becomes the Attack Surface: RAG Under Poisoned Evidence

A controlled RAG poisoning experiment shows why retrieved-document integrity, source independence, and abstention monitoring belong inside production reliability controls.

September 30, 2026 · 7 min · Zelina