Entity · paper

Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers

paperactiveprovisionalbridging-the-gap-between-latent-and-explicit-reasoning-with-looped-transformers-4a17bada·1 events·first seen 2d ago

Aliases: Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers

Co-occurring entities

Chain-of-Thought Reasoning LOTUS

More like this (12)

Fixed-Point Reasoners: Stable and Adaptive Deep Looped Transformers Latent Reasoning with Normalizing Flows Training-Free Looped Transformers Reasoning Language Models Learning to Reason by Analogy via Retrieval-Augmented Reinforcement Fine-Tuning The Riddle Riddle: Testing Flexible Reasoning in Large Language Models and Humans Long-context Reasoning Benchmarks latent reasoning Dynamic Short Convolutions Improve Transformers Beyond the Commitment Boundary: Probing Epiphenomenal Chain-of-Thought in Large Reasoning Models Reasoning over Grammar: Can Synthetic Linguistic Reasoning Traces Enhance Low-Resource Machine Translation?Graph Transformer

Recent events (1)

6arXiv · cs.CL·2d ago·source ↗

LOTUS: Looped Transformers bridge latent and explicit chain-of-thought reasoning at 3B scale

Researchers introduce LOTUS (Looped Transformers with parallel supervision on latents), a latent chain-of-thought method that processes reasoning steps in hidden states rather than decoded tokens. LOTUS is claimed to be the first latent-CoT approach to match explicit CoT performance at the 3B parameter scale, while reducing thought-phase latency by 2.5x–6.9x. The method uses a looped (recurrent-depth) Transformer backbone with parallel cross-entropy supervision on gold CoT-step tokens at each latent position, and the latent space is shown to be interpretable by projecting through the base LM head to recover reasoning steps.

Evaluation and Benchmarking Inference Economics Chain-of-Thought Reasoning LOTUS Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers