MUX: Continuous LLM Reasoning Through Multiplexed Tokens

In July 2026, researchers published MUX (multiplexed tokens) on arXiv, a method for running language-model reasoning in a continuous latent space, where a single latent token losslessly encodes an entire fragment of the reasoning chain instead of one word. According to the authors, MUX outperformed strong latent reasoning baselines across 32 evaluation settings on four language models.

Why conventional reasoning is slow

Classic reasoning models waste computation: each step in the chain of thought passes along only a single subword, and some steps go to verbalizing the thought rather than doing the actual computation. The MUX authors call this a computational bottleneck: the model reasons in natural language, so the bandwidth of each step is capped by the size of one subword token, and on long chains this adds up to wasted compute.

What multiplexed tokens are

MUX distills a discrete reasoning chain into continuous latent tokens. Each such token is trained to represent a weighted linear superposition (multiplexing) of an entire span of subwords, and that superposition is invertible: the original span of subwords can be fully recovered (demultiplexing). The key idea is to make the compression reversible by construction rather than approximate, so the model does not "cut corners" and keeps all of the original reasoning logic.

• Method: MUX, arXiv preprint arXiv:2607.18264, publication type "new" (July 2026)
• Mechanism: one latent token equals a weighted sum of several discrete reasoning subwords
• Invertibility: the superposition is lossless, and the span of subwords can be fully recovered
• Weighting: position-dependent, for example geometric decay of the weights
• Bonus: parallel exploration of multiple solution branches on tasks that require search

The authors prove that simple position-dependent weights, such as a suitable geometric decay, yield lossless multiplexing. This in turn prevents "latent collapse": a degenerate shortcut behavior in which the model collapses the latent representation and stops actually reasoning.

How much better MUX is than the alternatives

MUX outperformed strong latent reasoning baselines in 32 experimental settings spanning four different language models, according to the authors' abstract on arXiv. That means an advantage not over a single model but over a whole class of latent reasoning methods, across the full grid of 32 configurations. In addition, ablation and probing analyses showed that the learned latent tokens encode correct, interpretable reasoning rather than random noise.

Another result is the ability to search in parallel: multiplexed reasoning can explore several solution branches at once on tasks that require trying multiple options.

"Lossless superposition as a local training objective is a sufficient

condition for strong and efficient continuous reasoning in latent space," the MUX abstract on arXiv states.

What this means

Latent reasoning is an attempt to remove the "language tax" from reasoning models: computing in a compact continuous space instead of long text chains. If MUX's results reproduce on larger models, it points toward cheaper and faster reasoning without loss of accuracy or interpretability, and therefore toward reasoning AI that spends less compute on the same task.

Frequently asked questions

What are multiplexed tokens?

They are continuous latent tokens, each of which encodes a weighted sum of several subwords from a reasoning chain. By construction, the superposition is invertible: the original span of subwords can be fully recovered, so no information is lost.

How does MUX differ from ordinary chain of thought?

Ordinary chain-of-thought passes one subword per step and reasons in text. MUX compresses an entire fragment of reasoning into a single latent token and computes in a continuous space. This raises throughput and, according to the authors, outperforms latent baselines in 32 test settings on four models.