AI Explained

Plain explanations of trending AI concepts, with live visualizations.

LLM

Gated DeltaNet-2 paper — Decoupled channel-wise erase/write gates — What does it mean?

Gated DeltaNet-2 replaces the single scalar gate of prior linear-attention models with two independent per-channel vectors — one for erasing old state, one for writing new state — so each dimension of the recurrent memory can decay at its own rate.

LLM

ACC paper — Tool-output unmasking — What does it mean?

ACC reformats agent trajectories as long-context QA pairs by unmasking tool outputs — Qwen3-30B-A3B gains +18.1 MRCR, matching Qwen3-235B-A22B 8× smaller.

LLM

RELEX paper — Rank-1 RLVR weight-trajectory extrapolation — What does it mean?

RELEX exploits the empirical finding that RLVR fine-tuning weight trajectories are near rank-1 — fit a line through 15% of training steps and extrapolate; matches full RLVR quality from a fraction of the compute.

LLM

OScaR paper — Token Norm Imbalance — What does it mean?

Token Norm Imbalance — a few tokens carry outsized KV norms along the sequence axis. Channel rotation can't flatten them; OScaR can.

LLM

MSSP paper — Scale-stable parameterization beyond muP — What does it mean?

MSSP applies Dynamical Mean Field Theory to MoE training and derives a parameterization that — unlike muP — keeps the optimal learning rate stable as both model width and expert count scale.

LLM

Mix-Quant paper — NVFP4 prefill + BF16 decode — What does it mean?

Mix-Quant quantizes only the prefill phase to NVFP4 and keeps decode in BF16 — up to 3× prefill speedup with task performance largely preserved, because the two phases sit on opposite sides of the roofline.

LLM

PSD paper — Parallel speculative decoding for diffusion LLMs — What does it mean?

Parallel speculative decoding gets up to 5.5× tokens per forward pass on diffusion LLMs by attacking spatial and temporal axes at once.

LLM

Attention Once Is All You Need — Persistent KV cache across queries — What does it mean?

AOIAYN persists the KV cache across queries in a streaming session and advances it as data arrives — prefill leaves the critical path and per-query latency stays constant in context length.

LLM

ZEDA paper — Zero-output expert self-distillation — What does it mean?

ZEDA injects parameter-free zero-output experts into a finished MoE and uses two-stage self-distillation to teach the router to skip ~50% of expert FLOPs at marginal accuracy loss.

LLM

SGLang v0.5.12 — TokenSpeed MLA backend — What does it mean?

SGLang v0.5.12 ships TokenSpeed MLA — a Blackwell attention backend for Multi-head Latent Attention that caches one shared low-rank K/V latent instead of per-head K/V, with TMA bulk-store reporting up to ~12× speedup on the cache-write kernel.

LLM

Spec-decode latency paper — Load-dependent latency model — What does it mean?

Paper decomposes spec-decode latency into load-independent and load-dependent parts via Little's Law — wins shrink as the server saturates.

LLM

RoPE provably fails at long context — Position and token discrimination limits — What does it mean?

Formal proof that RoPE's attention scores converge to random along BOTH the position axis and the token-identity axis as context grows — and the RoPE base parameter only trades one collapse for the other.