TrigReason: A Trigger Mechanism-Based Collaborative Framework for Large and Small Reasoning Models

TrigReason is a trigger mechanism-based collaborative framework for large and small reasoning models. Its core lies in enabling small-model-led collaborative reasoning with on-demand large model intervention through three intelligent triggers. While maintaining accuracy, the framework offloads 1.70-4.79 times more reasoning steps to small models, reducing latency by 43.9% and API costs by 73.3%, providing a new solution for balancing reasoning performance and efficiency.

Efficiency Dilemma of Reasoning Models

Large Reasoning Models (LRMs) such as OpenAI o-series and DeepSeek-R1 perform well in complex tasks (math competitions, programming challenges, etc.), but their autoregressive reasoning mechanism leads to high latency and high API costs, limiting their popularization. Small Reasoning Models (SRMs) are fast and low-cost but have weak capabilities; thus, rational task allocation becomes the key to balancing performance and efficiency.

Three Typical Risks of Small Models

Through experimental analysis, small models face three types of risks in complex reasoning:

1. Path Divergence: Lack of initial strategic planning ability, leading to reasoning deviating from the optimal path;
2. Cognitive Overload: Capacity limitations make it difficult to handle complex steps (e.g., multi-step derivation, constraints);
3. Recovery Incapability: Lack of self-reflection and error correction mechanisms, making it easy to persist on wrong paths. These risks are the premise for designing collaborative strategies.

TrigReason proposes a collaborative framework of selective intervention instead of continuous polling, with the core being to activate large models only when necessary and delegate most steps to small models. The three intelligent triggers correspond to the three types of risks:

Strategic Initiation Trigger

Triggered at the start of reasoning, the large model generates a problem-solving strategy and a framework of key steps to guide the subsequent reasoning of the small model, solving the path divergence problem.

Cognitive Offloading Trigger

Monitors signals of overconfidence from the small model during reasoning (e.g., sudden certainty in answers, skipped steps). When triggered, the current step is handed over to the large model for processing, solving the cognitive overload problem.

Intervention Request Trigger

Triggered when an invalid loop in reasoning is detected (repeated conclusions, lingering on the same choices, etc.), introducing the large model to break the deadlock, solving the recovery incapability problem.

TrigReason achieved the following results in benchmark evaluations of AIME24, AIME25 (math competitions), and GPQA-D (scientific question answering):

1. Accuracy Preservation: Equivalent to or even higher than the full large model, without sacrificing problem-solving quality;
2. Reasoning Step Offloading: Successfully delegated 1.70-4.79 times more steps to small models (the offloading ratio for structured tasks is close to 5 times);
3. Edge-Cloud Scenario Benefits: When small models run locally and large models are called from the cloud, latency is reduced by 43.9% and API costs by 73.3%.

Implementing TrigReason requires addressing three major engineering challenges:

1. Trigger Threshold Tuning: Provides an automatic tuning mechanism based on the validation set, finding optimal parameters through grid search;
2. Context Management: Maintains a unified reasoning state (including steps, conclusions, strategic blueprint), and formats prompts during switching to ensure coherence;
3. Error Recovery Mechanism: Lightweight error detection and backtracking; when the large model identifies previous errors, it rolls back to a checkpoint and re-reasons.

Limitations

1. The trigger design depends on the error patterns of small models; different small models require targeted adjustments;
2. Threshold tuning requires validation data, making zero-shot application to new tasks challenging.

Future Directions

1. Explore learning-based triggers to automatically learn optimal intervention timing;
2. Study multi-small-model collaboration, using different strengths to handle subtasks;
3. Extend the trigger mechanism to multi-modal reasoning scenarios.

TrigReason realizes a collaborative model of "small models as the mainstay, large models as the finishing touch". While maintaining accuracy, it improves efficiency and reduces costs. Its design philosophy reflects that intelligent resource scheduling and model capability enhancement in AI systems can produce synergistic effects. With the enhancement of edge computing and model diversification, such collaborative frameworks will play an important role in practical applications.