惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

B
Blog
V
Vulnerabilities – Threatpost
Apple Machine Learning Research
Apple Machine Learning Research
V
V2EX
博客园 - 叶小钗
阮一峰的网络日志
阮一峰的网络日志
人人都是产品经理
人人都是产品经理
Latest news
Latest news
博客园 - 三生石上(FineUI控件)
美团技术团队
aimingoo的专栏
aimingoo的专栏
Google Online Security Blog
Google Online Security Blog
Security Archives - TechRepublic
Security Archives - TechRepublic
T
Threatpost
Y
Y Combinator Blog
T
Tailwind CSS Blog
CTFtime.org: upcoming CTF events
CTFtime.org: upcoming CTF events
A
Arctic Wolf
C
Cyber Attacks, Cyber Crime and Cyber Security
小众软件
小众软件
Recent Commits to openclaw:main
Recent Commits to openclaw:main
T
Tenable Blog
W
WeLiveSecurity
L
LINUX DO - 热门话题
D
Docker
Cyberwarzone
Cyberwarzone
量子位
A
About on SuperTechFans
The Last Watchdog
The Last Watchdog
雷峰网
雷峰网
C
CERT Recently Published Vulnerability Notes
P
Palo Alto Networks Blog
The Hacker News
The Hacker News
Blog — PlanetScale
Blog — PlanetScale
P
Proofpoint News Feed
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
F
Full Disclosure
The Cloudflare Blog
T
The Blog of Author Tim Ferriss
T
The Exploit Database - CXSecurity.com
Engineering at Meta
Engineering at Meta
O
OpenAI News
Hacker News - Newest:
Hacker News - Newest: "LLM"
Scott Helme
Scott Helme
IT之家
IT之家
S
Secure Thoughts
MongoDB | Blog
MongoDB | Blog
L
Lohrmann on Cybersecurity
博客园 - 司徒正美
Google DeepMind News
Google DeepMind News

Hacker News - Newest: "LLM"

GitHub - lechmazur/position_bias: A benchmark for testing whether LLM judges keep the same preference when two lightly edited versions of the same story are shown in opposite orders. Flex routing (EU and EFTA) Dark Factories: Retooling for LLM Velocity Ask HN: What would be the impact of a LLM output injection attack? GitHub - AronDaron/dataset-generator: No-code desktop app for generating high-quality synthetic datasets to fine-tune LLMs — plan-then-execute pipeline, LLM-as-judge, HuggingFace upload. GitHub - Oaklight/llm-rosetta: Production-ready LLM API translation layer for Python — bidirectional conversion between OpenAI, Anthropic & Google formats via hub-and-spoke IR. Optional API gateway. Streaming & non-streaming. Zero core deps. Contributions welcome! GitHub - browser-use/browser-harness: Self-healing browser harness that enables LLMs to complete any task. GitHub - moeen-mahmud/remen: Remen turns thoughts into something you can return to Analyzing 156 LLM Launch Posts on Hacker News ChatGPT vs Gemini vs Claude: The Best LLM Subscription You Should Buy GitHub - salaamalykum/quran-semantic-search: High-density RAG Semantic Search Engine & Quran Corpus (GEO/SEO Architecture) GitHub - NVIDIA/TensorRT-LLM: TensorRT LLM provides users with an easy-to-use Python API to define Large Language Models (LLMs) and supports state-of-the-art optimizations to perform inference efficiently on NVIDIA GPUs. TensorRT LLM also contains components to create Python and C++ runtimes that orchestrate the inference execution in a performant way. The State of LLM Bug Bounties in 2026 Operational Readiness Criteria for Tool-Using LLM Agents Meshcore: Architecture for a Decentralized P2P LLM Inference Network How an LLM becomes more coherent as we train it GitHub - seetrex-ai/laimark GitHub - Jossifresben/BibCrit: AI-assited biblical textual criticism GitHub - wastedcode/memex: File system based wiki, maintained by Claude 99helpers.com GitHub - cliver-project/AITrigram GitHub - unbody-io/adapt: A self-evolving memory layer for AI agents. GitHub - hb20007/awesome-gen-ai-fails: A list of incidents where reliance on generative AI and LLMs resulted in harm to companies, individuals, or society GitHub - nevenkordic/localmind: Run any local LLM with persistent memory and context. CLI agent over Ollama with SQLite-backed hybrid recall. No cloud. Ask HN: What are the machine requirements for a LLM like Llama-3.1-8B? Faster LLM Inference via Sequential Monte Carlo grpo explained: group relative policy optimization for llm finetuning - cgft Stop comparing price per million tokens: the hidden LLM API costs · TensorZero Andrej Karpathy's LLM Wiki Is a Bad Idea GitHub - GG-QandV/mnemostroma: Offline RAM-first cognitive leer/coprocessor for AI agents and robotics. Solves "Context Abandonment" with 20-80ms latency using a dual-thread biomimetic memory architecture (ONNX + SQLite WAL). mempalace/agent at agent · skorotkiewicz/mempalace GitHub - Nyquest-ai/nyquest-rust-fullstack-pub: Nyquest — Semantic Compression Proxy for LLMs. 350+ rules, local LLM stage, 15-75% token savings. Full Rust stack. GitHub - TheoV823/mneme: Enforce architectural decisions in AI-assisted development. GitHub - klemenvod/TokenBrawl: A 1v1 Bomberman-style game where two LLM agents play autonomously against each other. No human plays — you watch the AIs fight. Each agent receives a text description of the board state, reasons about it, and outputs a move as JSON. The game engine executes it. Introducing the Common AI Provider: LLM and AI Agent Support for Apache Airflow Power Circuit AI: Designing Power Electronic Circuits for Motor Drives with Generative Artificial Intelligence Ask HN: How to program with IDE and LLM on CPU locally? Show HN: Agent-cache – Multi-tier LLM/tool/session caching for Valkey and Redis Bonsai 1-bit WebGPU - a Hugging Face Space by webml-community The LLM Fallacy: Misattribution in AI-Assisted Cognitive Workflows Ask HN: Simple tooling for local LLM code critique without IDE integration? Can a General LLM Diagnose a DICOM Slice? A 10-Case Public Benchmark Charts-of-Thought: Enhancing LLM Visualization Literacy (PDF, 2026) GitHub - Mesh-LLM/mesh-llm: Distributed AI/LLM for the people. Share compute privately or publicly to power your agents and chat. GitHub - seamus-brady/springdrift: A persistent runtime for long-lived LLM agents Writing an LLM from scratch, part 32k -- Interventions: training a better model locally with gradient accumulation Ask HN: Which LLM model and agentic CLI are you using for local development? GitHub - wayneColt/modelcascade: Route local. Escalate smart. Never overspend. Open-source multi-model cascade routing for autonomous agents. LLM pricing is 100x harder than you think GitHub - asakin/llm-primer: Pre-warmed Claude Code sessions in tmux. No startup wait. GitHub - EggerMarc/chat-rs: A multi-provider LLM framework for Rust. GitHub - SynapseKit/SynapseKit: Minimal, async-first Python framework for production LLM apps- 2 hard deps, no magic, no SaaS. A Claude Skill that Makes LLM Paragraphs More Bearable Does Gas Town 'steal' usage from users' LLM credits & paid services to improve itself? What's Claude Code Actually Doing? Open the Black Box with the Arthur Engine Milla Jovovich's New Open Source LLM Memory App and the Dark Code Problem Your intuition of LLM token usage might be wrong Show HN: Bloomberg Terminal for LLM ops – free and open source GitHub - 0xchamin/mcptube: Transform YouTube videos into a compounding knowledge base with transcripts, vision analysis, and agentic search. Works as an MCP server for Claude, Copilot & more. Show HN: Open KB: Open LLM Knowledge Base Your LLM is a compiler, not a runtime GitHub - sapountzis/Unslop: A Web Feed That Deserves You crates.io: Rust Package Registry Beyond Karpathy's LLM-Wiki: The Necessity of Cognitive Governance GitHub - amitshekhariitbhu/llm-internals: Learn LLM internals step by step - from tokenization to attention to inference optimization. GitHub - parallem-ai/parallem: An expressive library for running agents with the Batch API. GitHub - stfurkan/pi-llm LLM-Wiki Show HN: Formal – Formal verification for AI-generated code using Lean 4 LRTS – Regression testing for LLM prompts (open source, local-first) LLM Wiki Skill: Build a Second Brain with Claude Code and Obsidian I built an LLM Wiki and RAG solution: here's a demo for a security KB The biggest advance in AI since the LLM Predict-Rlm: The LLM Runtime That Lets Models Write Their Own Control Flow the-synthetic-library/the-synthetic-mind at main · joshferrer1/the-synthetic-library GitHub - yisding/reviewwiggum GitHub - Donnyb369/mcp-spine: Context Minifier & State Guard — Local-first MCP middleware proxy GitHub - Beledarian/wgpu-llm: A from-scratch LLM inference engine that uses wgpu (the cross-platform WebGPU implementation) to dispatch WGSL compute shaders for every math operation a Transformer needs. No CUDA. No Python. No massive framework dependencies. Just Rust, raw shaders, and your GPU. GitHub - anitiue/Hindsight: An experience-driven self-improvement framework for LLM agents — 基于经验的 LLM Agent 自我改进框架 GitHub - stef41/lmscan: 🔍 Detect AI-generated text and fingerprint which LLM wrote it. Open-source GPTZero alternative. Zero dependencies, works offline. GitHub - alainnothere/AmdPerformanceTesting: Amd Performance Testing Ask HN: Is a purely Markdown-based CRM a terrible idea? Optimized for LLM agents Context Engineering - LLM Memory and Retrieval for AI Agents | Weaviate little_helper_tui/letter.md at main · sleepyeldrazi/little_helper_tui GitHub - EvanZhouDev/umr: The Unified Model Registry for all your local AI apps. GitHub - JordanCT/VigIA-Orchestrator Your Agent Is Mine: Measuring Malicious Intermediary Attacks on the LLM Supply Chain A Taxonomy of RL Environments for LLM Agents Llama LLM Network Feture GitHub - genedeng-ca/ai-mac-migration: AI-powered Mac-to-Mac migration tool - replace Apple Migration Assistant with intelligent, selective transfer using local LLMs GitHub - lunargate-ai/gateway: High-performance self-hosted AI gateway (OpenAI-compatible) with routing, retries, and streaming GitHub - AuthBits/webmcp: A lightweight, prompt-driven MCP web research server for high-quality LLM powered information extraction. Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering Springdrift: An Auditable Persistent Runtime for LLM Agents with Case-Based Memory, Normative Safety, and Ambient Self-Perception High-Stakes Personalization: Rethinking LLM Customization for Individual Investor Decision-Making From Static Templates to Dynamic Runtime Graphs: A Survey of Workflow Optimization for LLM Agents HUOZIIME: An On-Device LLM-enhanced Input Method for Deep Personalization TIDE: Token-Informed Depth Execution for Per-Token Early Exit in LLM Inference Characterizing WebGPU Dispatch Overhead for LLM Inference Across Four GPU Vendors, Three Backends, and Three Browsers LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users
JetSpec: Breaking the Scaling Ceiling of Speculative Decoding with Parallel Tree Drafting
Lanxiang Hu, Zhaoxiang Feng, Yulun Wu, Haoran Yuan, Yujie Zhao, · 2026-06-26 · via Hacker News - Newest: "LLM"

TL;DR: Speculative decoding hits a scaling ceiling: a larger draft budget helps only while acceptance stays high and drafting stays cheap. Prior draft heads face a dilemma: autoregressive drafters condition on each path but pay with tree depth, while block-diffusion drafters draft in one pass but score branches independently, creating plausible yet mutually inconsistent trees. JetSpec trains a causal parallel draft head over fused hidden states from a frozen target model, so candidate-tree scores follow the target’s own autoregressive factorization. The frozen target then verifies the full tree in one forward pass, losslessly. On Qwen3-8B, greedy decoding with budget 256, JetSpec reaches 9.64x on MATH-500 and 4.58x on open-ended chat, and these gains carry into real single-stream serving on JetSpec’s own engine with an average of around 1000 TPS throughput on MATH-500 using a single B200 GPU.

speedup_comparison

Figure 1: End-to-end decoding speedup over standard autoregressive decoding on H100 GPUs across math, coding, and chat benchmarks. DFlash denotes the original block-parallel drafting method, DDTree is tree-based variant of DFlash, and JetSpec denotes our method.

Background

Modern LLM serving is still bottlenecked by autoregressive decoding: each token depends on the previous one, so generation is inherently sequential. Speculative decoding accelerates this process by drafting multiple future tokens and verifying them with the target model, but its speedup is controlled by two factors: (1) how many drafted tokens are accepted, and (2) how cheaply those tokens are drafted. Increasing the draft budget only helps when acceptance stays high and cumulative draft overhead stays low.

Existing head-based speculative decoding methods expose a core trade-off. Autoregressive draft heads such as Medusa and EAGLE-style methods preserve the target’s factorization and can produce faithful continuations, but drafting grows with tree depth. Parallel block-diffusion heads can draft a whole block in one pass, but positions are scored independently, so deeper branches can drift away from what the target would actually generate. Retrieval-based drafters avoid learned heads, but depend on lexical overlap or repeated text.

MethodDrafting StyleCausal Draft PathTree QualityDraft CostSpeedup
AR baselineNoneN/AN/AN/A1x
AR draft headsmultiple-pass sequential😃💰💰💰3~4x
Block-diffusion headsone-pass block draft😐💰3~6x
JetSpecone-pass causal tree draft😃💰4~10x

Table 1: Qualitative comparison of speculative decoding families. PLACEHOLDER optional rendered table.

This leads to the central question behind JetSpec: can we draft an entire speculative tree in one parallel pass while still scoring branches according to the target model’s causal, autoregressive factorization? JetSpec answers yes by combining causal parallel tree drafting with one-pass verification by the frozen target model.

At a high level, JetSpec targets both sides of the speculative-decoding bottleneck:

  • Low drafting cost: generate many tree nodes in one draft-head forward pass.
  • High acceptance: condition every node on its branch prefix, not just on its absolute future position.
  • Lossless verification: let the frozen target verify the tree and commit only the prefix it agrees with.

JetSpec trains a lightweight causal parallel draft head on top of a frozen target LLM. The head reuses rich multi-layer hidden features from the target, so drafting remains cheap, but it applies a tree-causal attention mask across draft slots. Each tree node can attend to the original prefix and its own ancestors, but not to unrelated sibling branches or descendants. As a result, all nodes are computed in parallel while every branch still follows an autoregressive-like dependency structure.

At inference time, the frozen target verifies every node in the speculative tree in a single forward pass under a tree-causal attention mask. The acceptance rule follows speculative decoding and commits the longest prefix accepted, so JetSpec preserves the target model’s exact output distribution under the same sampling rule. In other words, JetSpec improves speed without changing what the target would generate.

architecture-overview

Figure 2: JetSpec design overview. JetSpec extracts fused hidden features from the frozen target model and conditions a causal-parallel draft head to generate high-quality candidate trees in one forward pass.

Training the Head

Only the draft head is trained; the target model stays frozen. This lets JetSpec attach to a production model without changing its weights. Training samples anchor positions in target-aligned sequences, builds future-token blocks, and supervises the head against the target’s own next-token distributions.

We train with a causal mask over selected anchor positions, with each anchor expanded into a 16-token draft block. Each draft position attends to preceding positions within its block, enabling a single parallel pass to produce an internally consistent draft tree. In ablations, forward-KL style distillation slightly outperfroms SFT and is preferred because it preserves probability mass over multiple plausible continuations, while mode-seeking reverse-KL is less suitable for budgeted tree drafting.

block_and_anchors training_supervision

Figure 3: (left) Block-wise supervision where anchor tokens carry no loss. (right) Draft positions are supervised against the target distribution with a causal attention mask.

Tree Drafting and Verification

Both JetSpec and diffusion-style draft heads can spend the same draft budget; the difference is the quality of the tree that budget buys. A block-diffusion head produces all positions from one shared hidden state with no causal mask between depths, so a depth-2 distribution may never condition on the token chosen at depth 1. When two independently likely tokens cannot follow each other, the surrogate can still promote their composition to the top of the tree. JetSpec’s causal mask anchors each position to its branch prefix, so the verifier sees branches the target is more likely to accept.

jetspec-treewalk

Figure 4: One JetSpec drafting round. The draft head keeps multiple continuations in one pass; the frozen target verifies the whole tree and accepts the longest matching prefix.

Experiments

Results

JetSpec is evaluated on math, coding, and open-ended chat benchmarks with Qwen3-8B and Qwen3-30B-A3B under both greedy and sampling settings. The headline pattern is budget scaling: as the draft budget grows, JetSpec keeps converting extra candidate nodes into longer accepted prefixes, while prior heads saturate earlier.

At high token budget regime (>64 tokens), JetSpec leads DDTree (a tree-based variant of DFlash, our implementation follows this paper) and DFlash on every benchmark. The full grid below reports end-to-end speedup (Spd) and mean accepted length (τ) on Qwen3-8B over a selection of tasks.

MethodBudgetMATH-500 Spd (τ)AIME25 Spd (τ)HumanEval Spd (τ)MBPP Spd (τ)LCB Spd (τ)MT-Bench Spd (τ)
EAGLE-3642.36 (4.13)2.35 (4.04)2.49 (4.26)2.22 (3.81)2.09 (3.62)2.19 (3.88)
DDTree646.51 (7.16)6.40 (6.96)5.08 (5.57)4.99 (5.49)5.47 (6.06)3.74 (4.51)
DDTree1288.27 (9.19)7.93 (8.66)5.93 (6.52)5.70 (6.28)6.42 (7.25)4.12 (5.14)
DDTree2568.78 (9.81)8.33 (9.24)6.31 (6.96)6.09 (6.70)6.75 (7.72)4.26 (5.41)
JetSpec646.76 (7.42)6.47 (7.00)5.53 (6.06)5.34 (5.88)5.95 (6.59)3.97 (4.77)
JetSpec1288.93 (9.95)8.26 (9.10)6.66 (7.28)6.31 (6.95)7.29 (8.21)4.37 (5.52)
JetSpec2569.64 (10.76)8.78 (9.82)7.12 (7.78)6.73 (7.43)7.67 (8.79)4.58 (5.94)

Table 2: Comparison between JetSpec and other SD methods on Qwen3-8B, mean accepted length (in parentheses) and end-to-end speedup on a single H100 GPU are reported.

Takeaways:

  • Generalizability: largest gains appear on reasoning-heavy math and coding tasks, in consistent with our training data choice. JetSpec also generalizes with >4x speedup on open-ended conversational tasks.
  • Non-greedy sampling: gains shrink but remain consistent, showing the causal-tree benefit is not limited to deterministic decoding.
  • Budget scaling: larger tree budgets help JetSpec more reliably because the draft tree stays branch-conditioned.

Training with and without Causality

Bidirectional diffusion heads can recover some branch quality by training with an exponential loss weighting (an exponential loss weighting factor γ is applied in a causal order from left to right), but only near a tuned setting: they peak at a single weighting and collapse at extremes. JetSpec’s causal head holds strong speedups across the whole range and does not require this tuning.

MATH-500 speedup by loss weightingγ=0γ=3γ=7γ=15
Causal (JetSpec)8.298.508.408.41
Diffusion head5.468.168.366.17

Table 4: MATH-500 end-to-end speedup by loss weighting γ. PLACEHOLDER optional rendered table/figure.

Causality and Tree Quality

In this experiment we compare tree quality with and without causality enforced. The gap is measured the drafter’s log-probability difference, in nats (natural-log units), between its top-ranked branch and the target’s preferred continuation. A small gap means the tree contains the branch the target is more likely to accept and therefore a higher quality. On MATH-500, without loss weighting the block-diffusion head is miscalibrated, accepting a mean 4.84 tokens per round against the causal head’s 9.46.

Rank-1 branch faithfulness, 50 MATH-500 prompts, no loss weightingCausalDiffusion
Faithful rank-1, <+5 nats42%6%
Extreme gap, ≥+80 nats0%26%
Mean accepted length9.464.84

Table 3: Across 50 MATH-500 prompts, in comparison with block diffusion head, the causal head better keeps its top-ranked branch aligned with the target.

Limitations

While the speedups reported are end-to-end and lossless, they depend on GPU compute, memory bandwidth, batching, and the chosen tree budget. In addition to the algorithm knobs and hardware specifications, wall-clock serving speedups run below pure algorithmic accepted-length ratios because real serving includes kernel-launch, KV cache management, scheduling and other host-side overhead. The best draft budget is a trade-off: a larger tree raises accepted length but also increases per-round verification cost, so the optimal point depends on the model and hardware.

JetSpec Serving Optimization and vLLM Integration

To provide an inference path for real serving rather than only offline algorithmic evaluation, JetSpec includes a serving implementation based on our vLLM fork with JetSpec support and organizes draft candidates as a speculative tree, stores parent indices along with node depths, and passes this metadata into target verification. The target then verifies all speculative nodes in one forward pass under a tree-attention mask, and the acceptance rule remains lossless by construction.

We further implement paged FlashAttention kernels in both Triton and NVIDIA CuTe DSL that apply the tree mask directly inside the attention computation, without materializing a dense per-request mask. Larger budgets reduce verification rounds at low to moderate load, but can become less efficient under heavier batching because each round carries more verification and memory pressure.

jetspec_engine_demo

Figure 5: JetSpec deployed on vLLM using a single B200 with Qwen3-8B, MATH-500, batch 1, budget 127, providing a throughput at around 1000 TPS on MATH-500.

Get Started

For more details, please see the JetSpec paper and codebase:

Because JetSpec keeps the target model frozen and verifies under a lossless acceptance rule (speculative decoding), with a plug-and-play causal parallel drafting head, it can be attached to an existing model as a speculative decoding path while preserving the target output distribution.

Citation

@misc{hu2026jetspec,
 title  = {JetSpec: Breaking the Scaling Ceiling of Speculative Decoding with Parallel Tree Drafting},
 author = {Hu, Lanxiang and Feng, Zhaoxiang and Wu, Yulun and Yuan, Haoran and Zhao, Yujie and
           Qian, Yu-Yang and Wang, Bojun and Zhao, Peng and Jiang, Daxin and Zhu, Yibo and Rosing, Tajana and Zhang, Hao},
 year   = {2026},
 eprint = {2606.18394},
 archivePrefix = {arXiv},
 primaryClass = {cs.CL},
 url    = {https://arxiv.org/abs/2606.18394}
}