[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-80138":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":9,"language":10,"languages":9,"totalLinesOfCode":9,"stars":11,"forks":12,"watchers":13,"openIssues":14,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":15,"stars7d":12,"stars30d":16,"stars90d":15,"forks30d":15,"starsTrendScore":17,"compositeScore":18,"rankGlobal":9,"rankLanguage":9,"license":9,"archived":19,"fork":19,"defaultBranch":20,"hasWiki":19,"hasPages":19,"topics":21,"createdAt":9,"pushedAt":9,"updatedAt":22,"readmeContent":23,"aiSummary":24,"trendingCount":15,"starSnapshotCount":15,"syncStatus":14,"lastSyncTime":25,"discoverSource":26},80138,"streaming","seal-rg\u002Fstreaming","seal-rg","Code for the paper Multi-Stream LLMs: Unblocking Language Models with Parallel Streams of Thoughts, Inputs and Outputs",null,"Python",59,4,53,2,0,7,1,41.8,false,"main",[],"2026-06-12 04:01:26","# Multi-Stream LLMs: Unblocking Language Models with Parallel Streams of Thoughts, Inputs and Outputs\n\nCode release for *\"Multi-Stream LLMs: Unblocking Language Models with Parallel\nStreams of Thoughts, Inputs and Outputs\"* (Su, Yang, Li, Geiping; 2026).\n\nThe paper has three experimental sections, each with its own subfolder\nunder this directory. The subfolders are self-contained and each can be\nset up and run independently.\n\n## Layout\n\n```\nsec5_efficiency\u002F      Paper Section 5: Efficiency\n                      Qwen3-1.7B \u002F 4B with 2- or 3-stream interleaved packing.\n                      Trains \"solving-while-reading\" and \"auditing-while-solving\".\n                      Eval: GSM8K, MATH500, LogicNLI, SQuAD, ProofWriter, PubMedQA.\n\nsec6_security\u002F        Paper Section 6: Security\n                      Qwen2.5-7B \u002F Qwen3-4B with interleaved packing.\n                      Trains on multi-stream-reconstructed Alpaca.\n                      Eval: TensorTrust, Gandalf, Purple, RuLES, StruQ-ID\u002FOOD,\n                      NESSiE, IFEval.\n\nsec7_monitorability\u002F  Paper Section 7: Monitorability\n                      Stream-8B (Qwen3-8B) and Stream-27B (Qwen3.5-27B) with\n                      10 cognitive streams. Qwen3.5 uses per-stream\n                      Gated-DeltaNet states.\n                      Eval: AF eval-aware\u002Fsub-vocalization, monitor accuracy\n                      (Meinke\u002FSchoen 6-class), concern sub-vocalization.\n```\n\n## Implementation Notes across subfolders\n\n\n| Aspect              | Sec 5 \u002F Sec 6                            | Sec 7                              |\n| ------------------- | ---------------------------------------- | ---------------------------------- |\n| Backbone            | Qwen2.5 \u002F Qwen3                          | Qwen3 \u002F Qwen3.5 (incl. DeltaNet)   |\n| Dataset format      | `.jsonl`, single-row per timestep        | `.npz` packed shards, 10-channel   |\n| Data construction   | wait-$k$ (MetaMath, HotpotQA, Alpaca)    | synthetic 10-stream tabular        |\n| Number of streams   | 2 (solving-while-reading) \u002F 3 (auditing) | 10 (User, Output, 8 thinking)      |\n| Entry point         | `train\u002Ftrain\u002Ftrain{_qwen3}.py`           | `train\u002Ftrain\u002Ftrain_stream.py`      |\n\nNote: Sec 5 \u002F Sec 6 model classes are named `Qwen2ForMedusa` \u002F\n`Qwen3ForMedusa` and have a `medusa_num_heads` config attribute for historical reasons, but at\ninference time the model functions as described in the paper with complete weight sharing between streams.\n\n## Quick start\n\nPick the section you care about and follow its README:\n\n- **Sec 5 (Efficiency)** — `sec5_efficiency\u002FREADME.md`\n- **Sec 6 (Security)** — `sec6_security\u002FREADME.md`\n- **Sec 7 (Monitorability)** — `sec7_monitorability\u002FREADME.md`\n\n\n## Citation\n\n```bibtex\n@article{su_2026_multi-stream,\n  title={Multi-Stream LLMs: Unblocking Language Models with Parallel Streams of Thoughts, Inputs and Outputs},\n  author={Su, Guinan and Yang, Yanwu and Li, Xueyan and Geiping, Jonas},\n  year={2026}\n}\n```\n","该项目是为论文《多流LLM：通过并行思维、输入和输出流解锁语言模型》提供的代码实现。核心功能包括利用Qwen系列模型，通过并行处理多个数据流来提高效率、安全性和可监控性。技术特点涵盖基于Qwen2.5\u002FQwen3的多流交错打包训练方法，支持从2到10个认知流的配置，并采用Gated-DeltaNet状态进行高级别控制。适用于需要提升大语言模型在解决复杂问题时的实时响应速度、确保信息安全以及增强模型行为透明度的研究或应用场景。","2026-06-11 03:59:24","CREATED_QUERY"]