Think-at-Hard: Selective Latent Iterations to Improve Reasoning Language Models

The paper introduces Think-at-Hard (TaH), a looped transformer designed to enhance reasoning in Large Language Models (LLMs) by selectively iterating on token predictions. TaH employs a lightweight neural decider to trigger additional iterations only for tokens predicted to be incorrect, utilizing depth-aware Low-Rank Adaptation (LoRA) and a duo-causal attention mechanism to enable efficient cross-iteration information flow. Experimental results demonstrate that TaH achieves performance improvements of 3.8-4.4% over always-iterate baselines while skipping iterations on 93% of tokens, highlighting its potential for optimizing reasoning tasks in LLM applications.