PapersMulti-source11月3日 14:20

MIT 研究：编辑 0.016% 权重让 Qwen3.5-0.8B 的 MATH-500 成绩从 7.9% 升至 48.1%

Latent.Space@latentspacepod

Post· 8 sources

Why we picked it

Qwen3.5-0.8B 做数学题经常卡在死循环里烧 32K token，MIT 找到出问题的两层，只改 0.016% 权重就把 MATH-500 从 7.9 分拉到 48.1 分。

MIT 的一项研究中，GPT-6 Astra 分析了 Qwen3.5-0.8B 在数学题上耗尽 32K token 仍无答案的原因，将问题定位到两层循环。研究团队随后编辑了该模型 0.016% 的权重。在采样条件固定的情况下，Qwen3.5-0.8B 在 MATH-500 上的成绩从 7.9% 提升到 48.1%。

MIT 研究：编辑 0.016% 权重让 Qwen3.5-0.8B 的 MATH-500 成绩从 7.9% 升至 48.1% — Image · Latent.Space

Original · Latent.Space

Qwen3.5-0.8B often spent 32K tokens on a math problem without reaching an answer. In an MIT study, GPT-6 Astra traced the loops to two layers and edited 0.016% of Qwen’s weights. With sampling held fixed, its MATH-500 score rose from 7.9% to 48.1%. latent.space/p/ainews-llms-… 💬 0 🔄 0 ❤️ 4 👀 617 📊 1 ⚡

rohanpaul_ai11-02 20:17原文
shao__meng00:45原文
The Rundown AI11-02 20:17原文
IT之家02:00原文
kimmonismus11-01 10:51原文
Jerry Liu11-02 19:44原文
marktechpost11-02 07:28原文

#Qwen3.5 #MATH-500 #模型微调 #推理模型

Keep reading

Recent stories with the same tag or section

Models BuildersMulti-source04:19

Chinese contentMIT 研究用 GPT-6 Astra 直接改权重提升 Qwen3.5-0.8B 数学成绩

不训练也能改模型：Astra 找到让 Qwen3.5-0.8B 复读循环的信号并从权重里减掉，0.016% 的改动让 MATH-500 从 7.9% 涨到 48.1%。

#GPT-6 Astra #Qwen3.5 #MATH-500 #MIT

Event

rohanpaul_ai@rohanpaul_ai7 sourcesSource

Models BuildersOfficial17:07

Chinese contentqwenGPU：把 4 通道 SIMT GPU 编译进 Qwen3.5-2B 权重

阿里 Qwen 团队把一个小 GPU 直接编进 2B 模型权重里，模型自己写 CUDA，玩法挺硬核，喜欢底层架构的可以看看。

#Qwen3.5 #qwenGPU #CUDA #SIMT

Event

官方一手marktechpost@Michal Sutter3 sourcesSource

Papers Builders12:35

Chinese contentarXiv 研究：给训练好的模型加循环层只有小模型才划算

想省显存给模型加循环层？这篇论文实测了 Qwen3 从 0.8B 到 9B 的效果，结论是小模型才划算，大模型纯亏推理算力。

#Qwen3 #Qwen3.5 #推理模型 #开源模型

Event

arXiv cs.LG@Rhea Sanjay Sukthanker, Pashmina Cameron2 sourcesSource

Models BuildersMulti-sourceSelected09:07

SenseNova Loop-9B-Preview open-source model matches 27B rivals

Parameter-sharing loops let a 9B model punch at 27B weight class, which cuts inference cost for teams running complex agent pipelines on limited hardware. Developers building coding or office-automation agents can download the weights from Hugging Face and benchmark them against Qwen3.5-27B directly. The approach also signals a broader shift in Chinese labs toward compute-depth efficiency over raw parameter scaling.

#SenseNova #SenseTime #open-source models #Looped Transformer

Event

IT之家4 sourcesSource

Models BuildersMulti-sourceSelected21:18

Xiaomi open-sources 0.8B OCR model Xiaomi-OCR-0

A sub-1B model reaching near-top OmniDocBench scores means document parsing and invoice extraction can now run on modest hardware rather than large multimodal models. Teams building document pipelines can deploy it locally, pull weights from Hugging Face or GitHub, and use the JSON and Markdown outputs directly in automation workflows. Against rivals like PaddleOCR's larger document models, it trades scale for efficiency while matching them on benchmarks.

#Xiaomi #Xiaomi-OCR-0 #open-source models #OCR

Event

IT之家3 sourcesSource

Models BuildersSelected14:19

Qwenputer Embeds a RISC-V CPU Inside Qwen3.5-4B Weights

Encoding a real instruction set inside model weights points to a future where LLMs handle deterministic computation natively instead of routing to a code interpreter, trading speed and memory for reliability. Developers working on agents or tool-free reasoning can try the open-source weights on a single 24GB GPU, though the multi-token-per-instruction design makes it impractical for latency-sensitive use today.

#Qwen3.5 #Qwen #RISC-V #open-source models

Event

IT之家2 sourcesSource