SuperThoughts: Reasoning Tokens in Superposition

A new method called SuperThoughts compresses pairs of consecutive chain-of-thought tokens into single latent representations, reducing reasoning length in large language models by roughly 20 to 30 percent while preserving accuracy, according to research posted to arXiv on June 11, 2026 [1]. The framework targets the computational cost of long chain-of-thought reasoning, where models generate sequential tokens to solve complex problems. SuperThoughts instead processes and generates tokens in pairs, halving the number of forward passes required during inference [2]. A lightweight Multi-Token Prediction module decodes two discrete tokens per step: one from the main model backbone and one from the MTP module [3]. This design preserves token-level cross-entropy supervision throughout training, an element earlier latent-reasoning approaches often lacked, which contributed to training instability and difficulty scaling to long-horizon tasks [4]. The researchers fine-tuned three variants of Qwen2.5-Math — at 1.5 billion, 7 billion, and 14 billion parameters — and evaluated performance on MATH500, AMC, OlympiadBench, and GPQA-Diamond [1]. Across these benchmarks, SuperThoughts achieved a CoT length reduction of approximately 20 to 30 percent. Accuracy degradation was limited to 1 to 2 points on most tasks [2]. A confidence-based adaptive mechanism triggers a fallback to standard decoding when the model’s certainty is low, which the authors credit for maintaining performance [3]. The work enters a field where latent chain-of-thought reasoning has drawn attention but also skepticism. A separate study presented at the LIT Workshop at ICLR 2026 examined whether language models genuinely exploit superposition — the ability to hold multiple candidate solutions in a single representation — when given continuous reasoning tokens. That analysis, covering the Soft Thinking and Coconut methods, found that superposition collapsed within the first few layers and that models processed soft tokens nearly identically to discrete ones [5]. SuperThoughts differs by retaining discrete token supervision and by operating on compressed pairs rather than fully continuous latent states. Neural networks, the substrate for these models, consist of layers of artificial neurons that transform inputs through weighted connections, with deep networks employing multiple hidden layers to learn hierarchical representations [6]. Transformer architectures, which underpin large language models, use attention mechanisms to model long-range dependencies in data [8]. The computational demands of training and running such networks have driven interest in efficiency methods like SuperThoughts.