AI Explained
Plain explanations of trending AI concepts, with live visualizations.
vLLM 0.25 makes Model Runner V2 the dense default — Retiring PagedAttention — What does it mean?
vLLM 0.25 removes PagedAttention and makes Model Runner V2 the dense default, and reports full CUDA graphs for decode.
Linearization study isolates cache routing for long-context attention — Training-free linear attention — What does it mean?
A frozen model is converted to linear attention without retraining — sink tokens, a short convolution, and fixed-budget cache routing add back the long-context recall a plain swap loses.
Quantization study finds accuracy can hide behavior drift — Correctness agreement — What does it mean?
A quantized LLM can keep the same accuracy yet change which answers it gets right — correctness agreement measures that hidden drift.
NVIDIA Audex — Unified audio-text token space — What does it mean?
Audex is one text decoder that also hears and speaks: audio shares the token space of text, with little to no drop in text skill.
MAESTRO prunes MoE experts with routing-aware Markov chains — Markov-chain expert pruning — What does it mean?
MAESTRO models MoE routing as a Markov chain, ranks experts by their long-run share of traffic, and prunes the rarely-visited ones — keeping up to 10.61% more performance at 50% compression.
DominoTree speeds speculative decoding with conditional draft trees — Conditional draft-tree scoring — What does it mean?
DominoTree scores speculative-decoding draft trees by how well each token follows the last, not one token at a time — up to 6.6× on Qwen3-4B.
TF-Engram uses SSD-backed phrase memory for train-free LLM recall — Predictive prefetching that hides SSD latency — What does it mean?
TF-Engram stores phrase memories across GPU-DRAM-SSD and prefetches them so a slow SSD read overlaps decode instead of stalling it.
DeLS-Spec adds short-context heads to block drafting — Long-short logit fusion — What does it mean?
DeLS-Spec fuses a short-context head into a block drafter, discounting common words, so a fast block draft finally picks up the local context its parallel guesses skipped.
SIS turns off-policy RL tokens into on-policy updates — Selective Importance Sampling — What does it mean?
Selective Importance Sampling accepts the agreeing tokens as on-policy (ratio 1), taming the variance when RL reuses off-policy rollouts.
Discrete diffusion theory unifies denoisers, scores & bridge predictors — One object, three coordinates — What does it mean?
The three ways to train a diffusion language model — denoiser, score, and bridge — turn out to be one object written in different coordinates.
Direct-OPD transfers weak-model RL gains as log-ratio rewards — Weak-to-strong reward transfer — What does it mean?
Direct-OPD reuses a small model's before/after RL log-ratio as a dense reward to lift a stronger model — no RL on the big model.
LOCOS finds non-literal retrieval heads by scoring logit contribution — Logit-Contribution Scoring — What does it mean?
LOCOS finds the attention heads that answer from meaning, not copied words, by scoring how much each head writes toward the answer token.
