AI Explained
Plain explanations of trending AI concepts, with live visualizations.
STRACE extracts causal root causes from noisy agent traces — Causal trace localization — What does it mean?
STRACE shrinks a failed agent's huge run log down to the few steps that actually caused the failure — causal localization, not length-based trimming — lifting a verifier agent 42.5% → 58.5%.
Microsoft open-sources Flint for agent-generated visualizations — Semantic intermediate representation — What does it mean?
Microsoft's Flint has the agent write a short chart spec; a compiler fills in the brittle low-level details — for more reliable agent-drawn charts.
EdgeBench measures scaling laws for agents learning in the wild — Environment-learning scaling law — What does it mean?
EdgeBench watches agents keep learning on the job and fits one clean curve to how they improve: a log-sigmoid law with R²=0.998.
Proactive memory agent counters long-horizon state decay — Memory-grounded reminder injection — What does it mean?
A second agent watches a long task, keeps a memory bank, and re-injects a buried instruction at the right moment — lifting pass@1 without touching the action agent.
OpenAI finds about 30% of SWE-Bench Pro tasks broken — Benchmark task validity — What does it mean?
OpenAI audited SWE-Bench Pro and found ~34% of tasks broken — a benchmark's pass rate only means something if its tasks are valid.
Mistral releases Robostral Navigate for single-camera robot nav — Pixel-waypoint action space — What does it mean?
Robostral Navigate steers a robot from one RGB camera by predicting a pixel in the image to move toward — no depth sensor or LiDAR, and it still hits 76.6% on R2R-CE.
Multi-agent attacks slip past per-agent monitors (FakeLab) — The fragmentation effect — What does it mean?
Split a malicious goal across a team of agents and each trace looks harmless, so a monitor that checks agents one at a time can miss it.
AWS maps MCP tool-design tradeoffs for agents — Tool design as context engineering — What does it mean?
AWS reframes MCP tool design as context engineering — tighten schemas, trim responses, lazy-load detail, and at the limit hide the whole workflow behind one endpoint.
Agentic Botnets exploits hallucinated repo and skill names — Adversarial hallucination squatting — What does it mean?
Attackers pre-register the fake repo and skill names an LLM hallucinates, then hide malicious prompts there — up to 85% repo / 100% skill hallucination rates.
ToolFailBench separates skip, ignore, and fabricate failures — Tool-use failure taxonomy — What does it mean?
ToolFailBench sorts agent tool-use failures into four modes, and adds trap tasks to catch tools that shouldn't have been used.
Vera tests LLM-agent safety with executable evidence checks — Evidence-grounded verification — What does it mean?
Vera judges an AI agent's safety from the real evidence it leaves in a sandbox — files, tool calls, commands — not the model's own say-so.
LLM-as-a-Verifier scales agent feedback with logit-score expectations — Verification as a scaling axis — What does it mean?
Reading the verifier's whole score distribution — not one label — turns a coarse pass/fail into a continuous score you can scale.
