AI Explained
Plain explanations of trending AI concepts, with live visualizations.
RL data scheduler hits target perplexity with 44% fewer pretraining steps — RL-learned data mixture vs fixed pretraining blend — What does it mean?
An RL agent adjusts an LLM's pretraining data mixture on the fly instead of using a fixed blend, hitting target perplexity with 44% fewer steps.
InfoKV: entropy-aware KV-cache compression keeps long-context recall — Forward Influence — What does it mean?
InfoKV compresses the KV cache by a token's predictive uncertainty, not attention alone — keeping the unsure tokens that steer distant context.
JetSpec speeds speculative decoding up to 9.64× — Parallel tree drafting — What does it mean?
JetSpec drafts a tree of candidate tokens in one pass, turning a bigger draft budget into longer accepted runs — up to 9.64× faster decode.
WorldKV — Evict-and-reinsert KV memory — What does it mean?
WorldKV evicts KV chunks to camera-indexed storage and reinserts them on revisit, bounding a video world model's cache at ~2x throughput.
MiniMax-M2 ships a 230B open MoE with 40× faster RL training — Forge RL prefix-tree merging — What does it mean?
MiniMax-M2's Forge RL merges rollouts that share an opening into one tree, so the shared work runs once — a reported 40× training speedup.
Grouped Query Experts puts mixture-of-experts routing inside attention — Query-head expert routing — What does it mean?
Grouped Query Experts adds a per-token router that turns on only a few of attention's query heads — keeping every key-value head dense, so the KV cache is unchanged.
Baidu Unlimited OCR holds the KV cache constant for 40+ pages — Reference Sliding Window Attention — What does it mean?
Baidu's Unlimited OCR swaps decoder attention for R-SWA — each token reads the whole document plus only the last 128 outputs — so the KV cache stays constant across 40+ pages.
GLM-5.2 becomes the top open-weights model — Active vs total parameters — What does it mean?
GLM-5.2 lists 744B total but 40B active parameters — two numbers that decode different costs: the memory you hold vs the compute you pay per token.
Taylor-Calibrate cuts hybrid-attention distillation tokens 4.9–9.2× — Taylor-guided gate initialization — What does it mean?
Taylor-Calibrate presets a linear-attention student's gates from the teacher — hitting distillation targets with 4.9–9.2× fewer tokens.
HydraHead fuses full and linear attention per head, not per layer — Head-axis attention hybridization — What does it mean?
HydraHead mixes full and linear attention head-by-head, keeping exact attention only for retrieval-critical heads — a 7:1 split matching a coarser 3:1.
EfficientRollout — Self-speculative decoding with quantized self-drafters — What does it mean?
EfficientRollout speeds RL rollouts by drafting with a quantized copy of the model itself — a self-drafter that tracks the evolving policy for free.
CacheWeaver reorders RAG evidence for prefix-cache reuse — Prefix-cache-aware evidence reordering — What does it mean?
CacheWeaver reorders the retrieved chunks in a RAG prompt so the serving engine reuses its KV prefix cache — cutting median TTFT 20–33% with no measured quality loss.
