AI Explained
Plain explanations of trending AI concepts, with live visualizations.
PACE — Influence-path confinement vs content vetting — What does it mean?
PACE checks each tool call just before it runs: propose a cut of paths from untrusted or secret data, and check effects against the request.
OverAct — Proactive over-authorization vs request-scoped tool use — What does it mean?
Agents call more private-data tools than a request needs; SelfAudit makes each call cite the request and drops the rest.
AutoCompact — Learned compaction timing — What does it mean?
AutoCompact learns to compact an agent's context when a task stage ends, not when the window fills, and lifts SWE-bench Verified 9.2 points.
AWS Strands Decider 2B — Pointer-head decision models — What does it mean?
Strands Decider 2B replaces the LM head with a pointer head, so a 2B model always picks one supplied option and scores each.
Causal Memory Policy paper — Randomized memory exposure with inverse-propensity weighting — What does it mean?
Reserve 2 of 6 context slots for memories on a known schedule, and an agent can finally measure which memories help.
ActionGuard — Reviewer isolation from poisoned skill text — What does it mean?
ActionGuard approves each tool call against the user's request, and its reviewer never reads the possibly poisoned skill.
Planning-as-Routing — The plan declaration–execution gap — What does it mean?
A plan in a ReAct prompt often goes unfollowed; routing the declared mode to an enforcing executor fixes that. Mode choice stays hard.
Share-Borne AI Virus — Memory-hopping prompt injection — What does it mean?
One poisoned file, saved to memory and copied into new files, spread across agents that never talk to each other.
UserProxyBench audits simulated users — Simulated-user fidelity — What does it mean?
Swap only the simulated user and a frozen agent's benchmark score moves 15 points. UserProxyBench grades the user too.
TokenCast — Segment-level token forecasting — What does it mean?
An agent re-reads its whole context on every call, so TokenCast forecasts spend by segment: direct cost plus growth later calls pay for.
Failure-Transparent Agents — Structured evidence contracts — What does it mean?
After a tool fails, make the agent fill STATUS, EVIDENCE, LIMITATION and NEXT ACTION; false success claims were 0.8%, not 22.8%.
EffectMatch checks what an agent's tool call actually changed — Effect-based authorization — What does it mean?
An approved tool call can succeed and still leave a change nobody approved. EffectMatch checks the real changes before commit.