Research papers, read to you in five minutes.
Research papers, read to you in five minutes.

Learn why skill histories, permission evidence, and runtime replay solve different memory problems, and how to choose among them without assuming that deleting one record removes its downstream influence.

Agents · Sep 7

Three Ways "Our Agent's Memory Still Works" Can Be Wrong

Three archived papers give agent-memory teams three separate tests: whether an upgraded reader can actually use an inherited store, whether stale memory beats a live authoritative value, and whether copied mutation evidence is tamper-evident under an external anchor.

Audio/Speech · Sep 6

Two Different Failures Look Identical in Your Audio Eval

An audio model gets the answer wrong. Was the waveform ever admissible evidence, or does accuracy hold up even when that admitted audio is removed or swapped? Two recent studies supply separate tests for those two questions — and they need separate scoreboards.

Evaluation · Sep 5

Two Cache Risks: Replay Failures and Hidden Grounding Loss

September 5, 2026 — Two archived studies examine caches as evaluation variables rather than as speed knobs: one shows a serving trajectory that depends on inherited server state, the other shows an offline RAG cache whose correct answers drift away from their evidence at INT4.

Evaluation · Sep 5

Three Independent Reasons a Robot Success Rate Can Mislead You

Three recent benchmarks pull apart what a single task-success number hides: whether a policy can recover from its own mistakes, whether a completed contact action was physically valid, and whether the model assigning the success label is right.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Sign up free

Agents · Sep 8

Agent Memory Controls Must Follow Consequences, Not Labels

Learn why skill histories, permission evidence, and runtime replay solve different memory problems, and how to choose among them without assuming that deleting one record removes its downstream influence.

Agents · Sep 7

Three Ways "Our Agent's Memory Still Works" Can Be Wrong

Three archived papers give agent-memory teams three separate tests: whether an upgraded reader can actually use an inherited store, whether stale memory beats a live authoritative value, and whether copied mutation evidence is tamper-evident under an external anchor.

Audio/Speech · Sep 6

Two Different Failures Look Identical in Your Audio Eval

An audio model gets the answer wrong. Was the waveform ever admissible evidence, or does accuracy hold up even when that admitted audio is removed or swapped? Two recent studies supply separate tests for those two questions — and they need separate scoreboards.

Evaluation · Sep 5

Two Cache Risks: Replay Failures and Hidden Grounding Loss

September 5, 2026 — Two archived studies examine caches as evaluation variables rather than as speed knobs: one shows a serving trajectory that depends on inherited server state, the other shows an offline RAG cache whose correct answers drift away from their evidence at INT4.

Evaluation · Sep 5

Three Independent Reasons a Robot Success Rate Can Mislead You

Three recent benchmarks pull apart what a single task-success number hides: whether a policy can recover from its own mistakes, whether a completed contact action was physically valid, and whether the model assigning the success label is right.

Never fall behind the literature again.

Free account. Follow topics, build your queue.

Browse topics

Research papers, read to you in five minutes.Research papers, read to you in five minutes.

Research papers, read to you in five minutes.
Research papers, read to you in five minutes.