MUX: метод непрерывного рассуждения LLM через мультиплексированные токены

Researchers published the MUX (multiplexed tokens) method on arXiv in July 2026 — a way to carry out language model reasoning in a continuous latent space, where a single latent token losslessly encodes an entire chunk of the reasoning chain instead of just one word. According to the authors, MUX outperformed strong latent reasoning baselines across 32 evaluation settings on four language models at once.

Why ordinary reasoning is slow

Classic reasoning models waste computation: each step of the reasoning chain conveys only one subword, and part of every step goes toward verbalizing the thought rather than the actual computation. The authors of MUX call this a computational "bottleneck": the model reasons in natural language, so the throughput of a single step is limited by the size of one subword token — and over long chains this translates into wasted computation.

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 this superposition is reversible: the original span of subwords can be fully recovered (demultiplexing). The key idea is to make the compression reversible by construction rather than approximate: this way the model doesn't "cut corners" and preserves the full original reasoning logic.

*Method — MUX, arXiv preprint 2607.18264, publication type "new" (July 2026)

*Mechanism — one latent token equals a weighted sum of several discrete reasoning subwords

*Reversibility — lossless superposition, the subword span is fully recoverable

*Weighting — position-dependent, for example geometric decay of weights

*Bonus — parallel exploration of several solution branches in tasks that require search

The authors prove that simple position-dependent weights — for example, suitable geometric decay — provide 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 is MUX than alternatives

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

A separate result is the ability for parallel search: multiplexed reasoning can simultaneously explore several solution branches in tasks that require searching through options.

"Lossless superposition as a local training objective is a sufficient

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

What this means

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

Frequently Asked Questions

What are multiplexed tokens?

These are continuous latent tokens, each of which encodes a weighted sum of several subwords of the reasoning chain at once. By construction, the superposition is reversible — the original span of subwords can be fully recovered, so no information is lost.

How does MUX differ from ordinary chain-of-thought reasoning?

Ordinary chain-of-thought conveys one subword per step and reasons in text. MUX compresses an entire chunk of reasoning into one latent token and computes in a continuous space — this increases throughput and, according to the authors, surpasses latent baselines in 32 test settings on four models.