AI Explained
Plain explanations of trending AI concepts, with live visualizations.
BeNice shares one KV cache across a looped model's recurrences — Cross-recurrence KV sharing — What does it mean?
BeNice lets a looped model keep one shared KV cache instead of one per recurrence, which earlier sharing did not reliably allow.
PKTP thought pruning — Pass@k-optimal pruning of partial reasoning paths — What does it mean?
Keep the partial reasoning paths most likely to give at least one correct answer, not just the best-looking ones.
OpenAttention reads 1% of the KV cache — Count-aware tail proxy — What does it mean?
OpenAttention keeps 99% of dense accuracy at a 1% KV read by folding the unread tail into one count-aware mean token.
MoEZip keeps expert routing stable at 10% KV retention — Routing-aware KV eviction — What does it mean?
MoE routers read the KV cache, so evicting tokens can change which experts run. MoEZip protects them.
CausalCoT paper — Parent-only context isolation — What does it mean?
CausalCoT answers each reasoning step in a fresh context holding only its parents' answers, so the final answer depends on the routed steps.
A fix for softmax and cross-entropy — Softmax Collapse and the precision cliff — What does it mean?
Why softmax cross-entropy rounds a confident model's gradient to zero, and the fix.
OnQ — on-policy distillation with Dynamic KL for 2-bit LLMs — What does it mean?
OnQ trains a 2-bit LLM on its own responses, with a per-token forward and reverse KL mix from the full-precision model.
RCSD — Training speculative drafts from accepted and rejected sequences — What does it mean?
RCSD trains a speculative draft model on target-model accept/reject preferences, not soft-label distillation.
LongZero trains long-context reasoning from zero data — answer-first synthesis — What does it mean?
LongZero writes the answer first, then long documents around it, and trains long-context reasoning with GRPO.
E8-QAT 2-bit training — E8 lattice implicit codebook — What does it mean?
E8-QAT snaps 8-weight groups to the E8 lattice by arithmetic, so 2-bit training needs no codebook search.
Adaptive Multi-Token Prediction — acceptance-survival draft horizon — What does it mean?
A training-free controller picks, before each round, how many tokens an MTP drafter proposes, from predicted acceptance survival.
Cache Engineering for Agents — Drop and reposition vs re-prefill — What does it mean?
Agents can drop old KV entries and renumber the rest, so editing context no longer forces a re-prefill.