Tencent releases DSpine draft model for Qwen3.8-27B
The five-layer checkpoint uses block size 16 and an 8,192-token window and is trending at number 30.
Stories per day · 30 days
3 RSS
Relatedinference10qwen4llm-inference3multi-token-prediction3agentic-ai2drafting2draft-models2dspark2hybrid-linear-attention2language-models2long-horizon-agents2mixture-of-experts2reasoning2reinforcement-learning2adversarial-preference1adversarial-reasoning1agentic-reasoning1automark1
49 stories, newest first · one card per story
Stories tagged #speculative-decoding
D3D enables asynchronous speculative decoding
D3D drafts during verification without target hidden states for 5.71x speedup.
Speculative Listening speeds speech models using idle compute
It achieves 2.6-2.8x speedup on speak chunks and reduces delay by 73%.
D3D enables asynchronous speculative decoding
Standalone diffusion drafter removes sequential dependency for 5x LLM inference speedup.
Adversarial Speculative Reasoning Improves Long Horizon Agents
Training draft models to reject at critical decision points improves agent accuracy by training 0.06% to 0.28% of parameters on agentic coding benchmarks.
Adversarial speculative decoding improves long-horizon agent reasoning
Training draft models to be rejected at critical points raises end-to-end success on complex agentic coding benchmarks without changing target model output.
SpecOPD branch-parallel distillation accelerates speculative decoding
SpecOPD parallelizes on-policy distillation across branches for up to 41.45% higher mean acceptance length and 3.56times generation speedup.
SpecPilot raises speculative decoding speedup from 4.46x to 10.87x
SpecPilot is an agentic loop that discovers algorithms and optimizations for speculative decoding across deployments.
TextSeal2 watermarks LLM output under speculative decoding without distortion
At equal diversity and a 10^-3 false-positive rate, it detects 8 to 22 points more 100-token answers than TextSeal.
Polymathic AI and Princeton release FlashMap for faster LLM decoding
The headline describes FlashMap as a flow-map drafter claimed to speed decoding by 6.14 times.
1
ADLER limits cloud LLM messages to the public view and answer
Speculative decoding sent a planted secret in rejected drafts; audited ADLER runs did not expose it beyond the answer.
Soft-MTP drops a noisy residual in multi-token prediction training
From 167M to 1.2B parameters, Soft-MTP lowers next-token perplexity and speeds speculative decoding without new parameters.
SPOT predicts commit order to accelerate diffusion language decoding
On SDAR-8B, SPOT cuts target passes by up to 6.8× and reaches up to 3.4× standard throughput.
NgramRL speeds RL rollouts using drafts from recent history
Warm-cache runs commit 3.60–7.77 tokens per forward, and a prefetcher raises Spider throughput 1.38x.
Speculative diffusion paper maps joint speed-quality tradeoffs
Authors report that jointly varying diffusion effort and acceptance leniency beats single-step diffusion drafting on the speed-quality frontier.
RCSD raises speculative-decoding acceptance length 35% over HASS
Draft models trained on accepted and rejected tokens report 1.35 times faster inference without a quality loss.
Adaptive multi-token prediction selects draft horizons online
On Llama and Qwen pairings, it was up to 4.1x faster than standard decoding and beat a fixed window in 8 of 10 settings.
Specter speeds single-GPU MoE inference via speculative expert offloading
A quantized on-device draft model predicts future experts, cutting time per output token by 1.50x on average.
ProVe proposer-verifier decoding accelerates language models
ProVe reaches 2.7x throughput using a 3B verifier at only 2.6 accuracy points below autoregressive baselines.
Lafa scales continuous flow language models to 8B
Trained on fewer than 30B tokens, its self-speculative decoding reaches 2.54 times autoregressive throughput at comparable quality.
MiniMax releases MiniMax-M3.1-Flash, a 428B multimodal MoE
About 23B parameters activate per token; inputs include text, image, and video, with a 1 million token context window.
2
ECHO folds multi-token prediction into the backbone itself
Dynamic Depth Sampling draws training depth with decaying probabilities, at half the overhead of training every depth.
Lossy speculative decoding inflates completion length
At the most relaxed settings tested, models accepted 1.34× tokens per round but wrote 1.54× longer completions.
DLite drops draft-model history for speculative decoding on agentic workloads
Authors report 2-6 acceptance length and gains of up to 70% over DSpark on agentic and long-context tests.