AI Explained
Plain explanations of trending AI concepts, with live visualizations.
TeamGym paper — Distributed pivotal information in agent teams — What does it mean?
TeamGym finds agent teams lose 22.9 points when needed facts are split across members, and 19.9 more when they must find them.
EnvSched paper — Reward-variance rollout scheduling — What does it mean?
EnvSched picks agent-RL rollouts by predicted reward variance plus environment coverage, so each batch teaches more than random sampling.
Probing agents before they leak shared credentials — Activation probes vs text monitors — What does it mean?
A linear probe on activations flags the call that will leak a secret before it is written; a trained text classifier nearly ties it.
Apple SynthAPI — Tool diversity before task count — What does it mean?
In the paper's tests, more apps and database variety beat more tasks at large budgets; the fitted law held at 16x more compute.
KARMA freezes weights and code — Regression-gated atomic edits — What does it mean?
KARMA edits one agent rule at a time and keeps the edit only if validation gain reaches 0.01; otherwise it restores the old snapshot.
Checkpoint agent context to use 6–7× fewer input tokens — Agent-callable context reset — What does it mean?
An agent that writes down verified facts, then clears its own context, beats re-reading the full history.
TAO cuts agent-serving TTFT by up to 36.3% — Joint session admission and placement — What does it mean?
TAO schedules agent sessions by tool-return readiness and decayed cache credit, solving admission and placement jointly.
Evaluate every tool-call prefix — prefix executability — What does it mean?
Score tool-using agents by how long their calls stay valid, not only by whether the final task succeeds.
GUIFlow — Cross-step KV reuse for GUI agents — What does it mean?
GUIFlow reuses cached KV states across GUI-agent steps and updates only selected states.
The Physical Cost of Context — Nominal tokens vs physical prefill — What does it mean?
Prompt-token savings can raise real prefill work if they edit cached history.
CacheLint paper — Prompt-cache preconditions — What does it mean?
CacheLint checks the conditions a prompt-cache call needs; 90% of real bugs fail silently.
Compact-Bench — measuring compaction damage with paired forks — What does it mean?
Compact-Bench forks an agent at a compaction point: full history vs summary only, same repo, same tests. Summaries damaged 5–64% of repairs.