TLDR: SLIM (Subtrajectory-Level Elimination for More Effective Reasoning) is a novel framework designed to improve the complex reasoning capabilities of Large Language Models (LLMs). It works by dissecting an LLM’s reasoning process into ‘subtrajectories’ and applying a ‘5+2’ framework to identify and eliminate suboptimal or inefficient parts. This includes five criteria for assessing subtrajectory quality and two steps for determining if a suboptimal subtrajectory can be safely removed. Coupled with a smart sampling algorithm that prioritizes high-quality, diverse reasoning paths, SLIM enables more effective supervised fine-tuning. Experimental results show that SLIM significantly boosts LLM accuracy on mathematical benchmarks with less training data, reduces the occurrence of suboptimal reasoning steps, and mitigates ‘underthinking’ by encouraging deeper exploration within each reasoning path.
Large Language Models (LLMs) have made incredible strides in tackling complex reasoning tasks, often by generating detailed ‘reasoning trajectories’ where they explore, reflect, and self-verify before reaching a conclusion. However, a new study reveals that not all parts of these trajectories are equally beneficial. In fact, some components can actually hinder performance.
Researchers from Huawei Technologies Co., Ltd. have introduced a novel framework called SLIM: Subtrajectory-Level Elimination for More Effective Reasoning. This approach aims to enhance the effectiveness of LLMs by systematically identifying and removing these suboptimal parts from their reasoning processes.
The core idea behind SLIM is to break down a complex reasoning trajectory into smaller, individual units called ‘subtrajectories.’ The team developed a ‘5+2’ framework to evaluate and refine these subtrajectories. The ‘5’ refers to five human-established criteria used to identify suboptimal subtrajectories:
Five Criteria for Quality Subtrajectories
- Effort: A subtrajectory should not just propose a method but also demonstrate its relevance and apply it in detail.
- Effectiveness: It should attempt to solve the problem efficiently, simplifying steps, advancing the process, or clarifying limitations.
- Coherence: Each step must be logically connected, with no leaps in reasoning.
- Preliminary Conclusion: Before moving on, a subtrajectory should draw a conclusion, whether it’s a final answer, intermediate finding, or an evaluation of the approach.
- Valid Verification: It should avoid redundant self-verification of statements already checked.
If a subtrajectory fails any of these five criteria, it’s flagged as suboptimal. But not all suboptimal parts can be simply removed. This is where the ‘+2’ part of the framework comes in:
Also Read:
- Unlocking Advanced Reasoning in LLMs: The Structured Solution Template Approach
- Balancing Brainpower: A New Approach to Efficient LLM Reasoning with Compact Chain-of-Thought
Two Steps for Elimination
- Independence: The framework assesses if a suboptimal subtrajectory is independent of subsequent content. If its removal would break the flow or understanding of later steps, it must be retained.
- Elimination: Only if a subtrajectory is both suboptimal and independent can it be eliminated.
Beyond just cleaning up individual reasoning paths, SLIM also introduces a sophisticated sampling algorithm. This algorithm selects high-quality Question-Answer (QA) pairs for supervised fine-tuning (SFT) of LLMs. It assigns a quality score to each thinking process, giving more weight to longer suboptimal subtrajectories, ensuring that significant inefficiencies are penalized more heavily. The algorithm also considers the distribution of subtrajectory counts to prevent over-favoring simpler reasoning paths, thus maintaining the model’s exploratory ability.
The experimental results are compelling. When fine-tuning Qwen2.5-Math-7B, the SLIM method achieved an average accuracy of 58.92% on challenging benchmarks like AIME24, AIME25, AMC24, and MATH500, using only two-thirds of the training data. This surpassed the 58.06% accuracy achieved with the entire dataset without SLIM. Furthermore, the method reduced the number of suboptimal subtrajectories during inference by 25.9%.
SLIM also effectively addresses the “underthinking” phenomenon, where LLMs frequently switch between reasoning strategies without fully exploring each one. Models fine-tuned with SLIM-curated data showed a significant decrease in the average number of subtrajectories and an increase in the average number of tokens per subtrajectory, indicating a deeper and more focused reasoning process. This leads to more efficient thinking and better performance across various inference token limits.
This research highlights that the quality of reasoning trajectories generated by LLMs, especially those trained with Reinforcement Learning, can be inconsistent. By providing a systematic way to refine these trajectories and curate high-quality training data, SLIM offers a powerful tool for developing more accurate and efficient LLMs. The full details of this work can be found in their paper: SLIM: Subtrajectory-Level Elimination for More Effective Reasoning.
