Cover image

When Better Perplexity Hides Less Variety

SOL reveals distributional failures that perplexity and token entropy can miss, offering evaluation teams a new way to compare generated text.

November 4, 2026 · 8 min · Zelina
Cover image

A Successful Run Is Not a Reproduced Result

PaperBenchX shows why scientific agents must regenerate and validate evidence, not merely execute specialized software.

November 3, 2026 · 6 min · Zelina
Cover image

Different Rubrics, Nearly the Same Scoring Lever

A mechanistic audit of 25 language models shows why distinguishable scoring criteria may still lack independent control over an AI evaluator's decisions.

November 3, 2026 · 8 min · Zelina
Cover image

One Harness, Many Failure Modes

How cross-model failure evidence can improve shared AI agent harnesses, and when the development-cost advantage holds.

November 3, 2026 · 8 min · Zelina
Cover image

Predicting More Can Teach an Agent Less

Decision-guided world modeling improves language-agent training by selecting which environmental information deserves prediction supervision.

November 3, 2026 · 8 min · Zelina
Cover image

The Door Opened. The Task Isn't Done.

MultINav shows why robots must verify changes in their environment and distinguish successful physical interactions from completed navigation tasks.

November 3, 2026 · 8 min · Zelina
Cover image

When a Risk Label Hides the Evidence

A suicide-risk AI study shows how complementary factor verification can improve explanations while upstream classification errors can suppress relevant evidence.

November 3, 2026 · 7 min · Zelina
Cover image

When More Geometry Makes the Answer Worse

GeoReform shows how improving the structured information supplied to a multimodal model can improve geometric reasoning without changing the solver.

November 3, 2026 · 8 min · Zelina
Cover image

A Picture Is Not a Permission Boundary

How rendering untrusted text as images can reduce prompt injection—and why model behavior, transcription fidelity, and costs determine whether the defense is deployable.

November 2, 2026 · 8 min · Zelina
Cover image

From Agent Traces to Steering Signals—Without Fine-Tuning

REVEAL shows how successful and failed frontier-agent runs can produce activation controls for smaller models, with promising benchmark gains but untested generalization.

November 2, 2026 · 7 min · Zelina