Researchers have developed Structured Thoughts, a framework that reorganizes how large language models reason by splitting their thinking into alternating "try" and "outcome" blocks.

The approach addresses a key inefficiency in current AI systems: verbose reasoning traces that consume excessive memory. Try blocks capture exploratory scratch work, while outcome blocks contain distilled conclusions from each reasoning step.

The team constructed their dataset by segmenting existing reasoning traces and prompting an LLM to summarize each step. Fine-tuning foundation models on this reformatted data produced models that naturally adopt the structured reasoning style.

Performance and Memory Gains

Models trained with Structured Thoughts achieved performance improvements of up to 8.08% on reasoning benchmarks compared to standard supervised fine-tuning approaches.

The explicit structure enables aggressive context pruning. After each try-outcome pair, the exploratory try block can be discarded while retaining the essential conclusions in memory.

A proof-of-concept implementation demonstrated an average of 85% memory savings across mathematical tasks, though this came with an 8.67% performance trade-off.

Technical Implementation

The framework organizes reasoning into clearly delineated blocks using XML-style tags. This structure makes it straightforward to identify which portions of the reasoning trace can be safely removed without losing critical information.

The approach differs from existing chain-of-thought methods by explicitly separating exploratory work from final conclusions, rather than treating all reasoning steps equally.

The research team included authors from multiple institutions, with Zain Sarwar as the lead researcher. The work was submitted to arXiv's Computation and Language section in July 2026.

The framework could prove particularly valuable for applications requiring long reasoning chains while operating under memory constraints, such as mobile AI assistants or edge computing scenarios.