Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Multi-Stream LLMs: Unblocking Language Models with Parallel Streams of Thoughts, Inputs and Outputs

About

The continued improvements in language model capability have unlocked their widespread use as drivers of autonomous agents, for example in coding or computer use applications. However, the core of these systems has not changed much since early instruction-tuned models like ChatGPT. Even advanced AI agents function on message exchange formats, successively exchanging messages with users, systems, with itself (i.e. chain-of-thought) and tools in a single stream of computation. This bottleneck to a single stream in chat models leads to a number of limitations: the agent cannot act (generate output) while reading, and in reverse, cannot react to new information while writing. Similarly, the agent cannot act while thinking and cannot think while reading or acting on information. In this work, we show that models can be unblocked by switching from instruction-tuning for sequential message formats to instruction-tuning for multiple, parallel streams of computation, splitting each role into a separate stream. Every forward pass of the language model then simultaneously reads from multiple input streams and generates tokens in multiple output streams, all of which causally depend on earlier timesteps. We argue that this data-driven change remedies a number of usability limitations as outlined above, improves model efficiency through parallelization, improves model security through better separation of concerns and can further improve model monitorability.

Guinan Su, Yanwu Yang, Xueyan Li, Jonas Geiping• 2026

Related benchmarks

TaskDatasetResultRank
Logical reasoningLogicNLI
Accuracy63.55
23
Question AnsweringSQuAD
Accuracy (Acc)74.74
12
Mathematical ReasoningMATH 500
Accuracy64
6
Mathematical ReasoningMATH500
Accuracy64
6
Logical reasoningProofWriter
Acc91.8
6
Question AnsweringPubMedQA
Accuracy73.5
6
Mathematical ReasoningGSM8K
Accuracy89.51
6
Indirect Attack SecurityStruQ OOD
Indirect ASR57.84
4
Instruction FollowingPrompt-L
Instruction Following Rate49.72
4
Instruction FollowingInst-L
Instruction Following Score60.19
4
Showing 10 of 17 rows

Other info

GitHub

Follow for update