DiscoLoop: Looping Discrete Embeddings and Continuous Hidden States for Multi-hop Reasoning

2026-06-30

AI Summary

Standard non-recurrent Transformers struggle to perform multi-hop reasoning within a single forward pass due to a local memory bottleneck that prevents effective reuse of early-layer knowledge in deeper layers. This work proposes DiscoLoop, an architecture that, for the first time, integrates discrete token embeddings and continuous hidden states into a dual-channel recurrent mechanism, enabling memory reuse and multi-hop reasoning in a single forward pass. A key insight is that representation misalignment constitutes the primary performance bottleneck; accordingly, the authors introduce an alignment intervention strategy that requires no additional training. Evaluated on symbolic and synthetic language tasks, DiscoLoop achieves near-perfect accuracy with fewer training steps and demonstrates significantly reduced pretraining loss and improved downstream benchmark performance in real-world settings.

Abstract

Large language models achieve strong performance on many reasoning tasks when allowed to externalize intermediate steps as Chain-of-Thought (CoT). However, many questions require the model to internalize the multi-step reasoning within a single forward pass before generating the answer. We study this challenge through two-hop reasoning, a representative task where the model must compose multiple pieces of parametric knowledge within a single forward pass. Standard non-recurrent Transformers suffer from a depth-local storage problem: facts learned in earlier layers are unavailable where second-hop retrieval happens. We found that Looped Transformers mitigate this issue by reusing the same memory, but still generalize imperfectly. We show that the remaining bottleneck is representational. In the two-hop reasoning task, the first loop often makes the correct bridge entity nearly perfectly decodable, yet the corresponding hidden state remains poorly aligned with the bridge token embedding. Surprisingly, an easy training-free realignment intervention nearly closes the generalization gap. Building upon this insight, we propose DiscoLoop, a looping architecture whose recurrence carries both a discrete embedding channel and a continuous hidden-state channel. DiscoLoop achieves near-perfect accuracy with substantially fewer training steps across symbolic and synthetic-language multi-hop reasoning tasks. When applied to real-world pretraining, DiscoLoop attains lower training loss and stronger benchmark performance than looped-transformer baselines, suggesting that the mixed-channel design transfers to practical language modeling.

Problem

• multi-hop reasoning
• depth-local storage
• internalized reasoning
• looped architectures
• representational bottleneck

Innovation

• DiscoLoop
• multi-hop reasoning
• looped Transformer
• discrete embedding
• continuous hidden state