AI Explained
Plain explanations of trending AI concepts, with live visualizations.
D3D diffusion drafter — Asynchronous speculative decoding — What does it mean?
A diffusion drafter without target hidden states drafts during verification, reaching 5.00× and 5.71× average speedups.
vLLM vs SGLang benchmark gap — Chunked-prefill confound — What does it mean?
Matching one chunked-prefill setting removed 78% of the vLLM-SGLang capacity gap, so default benchmarks overstate what the cache does.
Signed MCP tools can still leak — Germline execution — What does it mean?
Fork a data-free copy of each MCP tool per call, handing it only what its destination may see, so a signed Trojan has nothing to leak.
Targeted guidance in agent RL — Failure-gated self-distillation — What does it mean?
Distill judge feedback only on failing agent turns, and keep RL rewards, so agents learn new behavior without regression.
VeriHarm-Bench — Harmful success in coding agents — What does it mean?
Coding agents can pass their tests and still break things; VeriHarm adds a hidden harm test to measure it.
Frontier LLMs edit weights directly — Weight-space programming — What does it mean?
An LLM edited 131,072 weights of a small model, with no training, and lifted its MATH-500 score from 7.9% to 48.1%.
ARLoop — Attention residuals over recurrent depth — What does it mean?
ARLoop lets each looped layer attend over earlier depth states, and a short window of recent blocks is enough.
LatentTriad screens malicious agent skills from one prefill — Prefill-state skill screening — What does it mean?
Flags malicious agent skills from one prefill's hidden states, with no output tokens, tool calls or skill code run.
A Circuit for In-Context Learning — gated associative lookup — What does it mean?
Few-shot learning in 3–4B transformers traced to key-value label slots, reader heads, a repeat gate and brake heads.
Strategist evolves reusable inference strategies at 39% lower cost — Inference strategies as executable code — What does it mean?
Strategist saves inference strategies as reusable code, so each new task builds on past ones and costs 39% less.
LARK formally verifies LLM-written GPU kernels — Segment-wise Lean equivalence proofs — What does it mean?
LARK has LLM agents prove in Lean that a GPU kernel equals its PyTorch reference, one segment at a time.
DepthGraft rebuilds deep-layer KV from shallow and middle layers — Cross-layer KV reconstruction — What does it mean?
DepthGraft rebuilds the last third of layers' KV from earlier layers, cutting the cache by about a third.