AI Explained
Plain explanations of trending AI concepts, with live visualizations.
MCP-Injector — Protocol-aware fault injection — What does it mean?
Faults from a permanent-heavy MCP profile cut agent task score 60%; a 120B model was no more resilient than a 20B one.
HarnessMM cuts agent tokens up to 87.93% — Guarded decision reuse — What does it mean?
HarnessMM lets an agent harness run already-decided actions without calling the model again.
Train agents to respect token and tool-call budgets — Overspend penalties — What does it mean?
How reward design teaches one agent to stay inside a token, API-call or step budget: hard gates, fixed fines and learned overspend prices.
AfterDeleteBench — deleted memory, surviving summary — What does it mean?
Deleting an agent's memory record can leave a secret answerable through the summary built from it.
SWE-DIVE coding-agent benchmark — Dynamic user simulation — What does it mean?
SWE-DIVE tests coding agents against a simulated user who talks back, not one frozen issue prompt.
MHG-Mem multimodal agent memory — Provenance hypergraph vs flat text chunks — What does it mean?
MHG-Mem keeps agent memory as a provenance hypergraph, so identity, change, and source evidence stay linked instead of flattened into chunks.
EpiSEM-Bench — Matched safe/unsafe pairs and evidence recovery — What does it mean?
EpiSEM-Bench hides true safety facts from an agent and scores whether it acts unsafely, refuses, or goes and gets the fact.
AgenticEvolve lets agents test ideas during program search — Persistent experiment agents — What does it mean?
Island agents test ideas with their own scripts before spending a scarce formal evaluation.
Document color cues lost before agents read them — Formatting-aware ingestion vs plain-text extraction — What does it mean?
Plain-text converters drop colour and highlight; keeping them as tags lifts requirement F1 from 0.075 to 0.690.
Transferable tool-selection paper — Plan, Retrieve, Select with a trained 8B model — What does it mean?
A trained Qwen3-8B plans, retrieves and selects tools in large corpora, and beats a prompted 402B model on ToolBench G3.
Agent Behavior Maps — Centered action log-likelihoods — What does it mean?
Place agents on a map using action log-likelihoods on shared recorded runs, then predict success and similar behavior without new rollouts.
Cohere Rerank 5 — Context Focus calibrated threshold — What does it mean?
Rerank 5 scores spans inside documents; a calibrated threshold replaces fixed top-k.