Papers8月12日 16:39

AINews：如何窃取推理轨迹

[AINews] How to steal a Reasoning Trace

Official accountLatent Space (swyx)

Blog/Media

Read the original

Why we picked it

想知道怎么从推理模型里偷思维链？这篇讲透了投机解码的另类用法，安全研究者必看。

本期AINews探讨了通过投机解码技术窃取推理模型思维链的方法。文章指出，投机解码不仅能加速推理，还能被用来提取模型的内部推理过程。

Body

这种技术可能对模型安全构成威胁，因为推理轨迹包含敏感信息。文章还讨论了蒸馏与推理轨迹窃取之间的界限。

Original · Latent Space (swyx)

[AINews] How to steal a Reasoning Trace

Speculative Decoding by any other name would distil as sweet

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk. Key entities and technical topics under continuous tracking include 投机解码, 推理轨迹, 模型安全, 蒸馏.

Source attribution: curated from Latent Space (swyx).

#投机解码 #推理轨迹 #模型安全 #蒸馏

Keep reading

Recent stories with the same tag or section

Models BuildersMulti-sourceSelected01:08

Tencent Youtu releases Qwen3.8-27B-DSpine-b16 draft model

Draft-model engineering is quietly becoming a competitive lever: the gain comes from a better-matched draft model rather than the target model itself, so teams running Qwen3 27B can swap this in and roughly double effective throughput with no accuracy trade-off. Inference engineers and startups serving Qwen at scale should benchmark it against their current MTP setup, ideally on Hopper cards where FlashAttention 3 is available. Comparable accelerations typically require paid inference providers; this one is a free HuggingFace download.

#Tencent Youtu #Qwen3 #speculative decoding #HuggingFace

Event

karminski-牙医 (AI工具)@karminski3Source

Models BuildersSelected01:02

Muse Spark Open Hits 433 tok/s Per User on TPU v7 Megakernel

The notable part is that a 816B MoE model now runs end-to-end as one Pallas program on a single TPU host through SGLang and TorchTPU, not through a closed inference stack. Teams building long-context agent products should study the deep dive for fp8 KV and session-aware cache strategies they can replicate. Compared with CUDA-only megakernel efforts, this demonstrates the PyTorch-to-TPU path has matured enough for production serving.

#Muse Spark Open #PyTorch #SGLang #TPU

Official accountPyTorch@PyTorchSource

Papers BuildersMulti-source11:09

Chinese contentD3D：异步投机解码框架，Qwen3-8B 上实现 5 倍加速

复旦团队新出的 D3D，让扩散草稿不用等目标模型的隐藏状态，解码比 DFlash 快一截，GSM8K 上能到 7.48 倍，做推理加速的可以看看。

#D3D #Qwen3 #投机解码 #扩散模型

Event

arXiv cs.LG@Baoyou Chen et al. (7)Source

Papers Builders10:56

Chinese contentSplit-KV：面向 Diffusion Drafter 训练的高效上下文并行方法

训练 diffusion drafter 的人可以看看，Split-KV 在 CP 64 下最高加速 3.67 倍，负载均衡还快三个数量级，代码已开源在 GitHub。

#Split-KV #Diffusion #投机解码 #上下文并行

arXiv cs.AI@Geng Zhang et al. (5)Source

Papers Builders10:52

Chinese contentRECAP：面向多模态投机解码的预测回收方法

免训练不换草稿模型，直接给 Qwen2.5-VL 和 LLaVA 视觉模型推理提速 1.5 倍以上，代码开源可以自己跑。

#RECAP #Qwen2.5-VL #LLaVA #投机解码

arXiv cs.AI@Hengjie Zhu et al. (6)Source

Papers BuildersSelected14:04

Adversarial Speculative Reasoning Boosts Long-Horizon Agent Accuracy

Speculative decoding is usually tuned for speed, so the finding that acceptance choices can silently shift reasoning accuracy across long agent runs is a practical surprise. Developers building multi-step agents can adopt this training trick to improve reliability without touching the target model's weights or outputs. The parameter cost — under 0.3% — makes it feasible even for teams serving open-weight models.

#speculative decoding #agents #reasoning models #open-source models

arXiv cs.AI@Hongyin Luo et al. (8)Source