Dual-Track CoT Boosts Small LM Reasoning Under Token Budgets
A new research project, Dual-Track CoT, proposes a budget-aware stepwise guidance method to improve multi-step reasoning in Small Language Models (SLMs) of 7–8B parameters. Existing techniques like self-consistency, Tree-of-Thoughts, and critique-revise loops enhance performance but consume excessive tokens and lack fine-grained control. This work investigates whether process supervision and simple test-time controls—such as token budgets and rejection of redundant steps—can substitute for model scale or large sampling counts. The approach is both scientifically probing and practically relevant for on-device, low-latency, or cost-constrained deployments. The paper is available on arXiv under ID 2604.25039.
Key facts
- Dual-Track CoT targets SLMs of 7–8B parameters
- Existing methods like self-consistency and Tree-of-Thoughts are token-inefficient
- Project explores process supervision and token budgets as substitutes for model scale
- Practical applications include on-device and low-latency settings
- Paper published on arXiv with ID 2604.25039
Entities
Institutions
- arXiv