DRG: A Training-Free Finalization Recovery Method for Reasoning Models Under Strict Token Constraints

Detect-Restart-Gate (DRG) is a training-free method designed to address output quality issues of reasoning models under strict token budgets. By detecting pathological signals (repetition, excessive length, stagnation) during reasoning, it triggers a retry mechanism and intelligently gates answer selection, significantly improving accuracy in mathematical reasoning tasks. This method was released by AnonymousAuthor0211 on GitHub on May 24, 2026 (Project link: https://github.com/AnonymousAuthor0211/detect-restart-gate).

In recent years, LLMs (such as DeepSeek-R1, Qwen3, Ministral) have demonstrated strong reasoning capabilities through chain-of-thought, but verbose outputs often lead to 'unfinished' results in scenarios with limited token budgets. Traditional solutions like supervised fine-tuning require significant resources, while self-consistency methods have high reasoning costs—both have shortcomings.

DRG operates through a three-stage mechanism:

1. Detection: Generate baseline output via greedy decoding, and parallelly detect three types of pathological signals: repetition (>0.7), length (>P85), and stagnation (no new terms for 4 consecutive lines);
2. Retry: When triggered, retry with sampling strategy (temperature=0.7, top_p=0.95), with prompts including the original problem and the last 1200 characters of the baseline;
3. Gate: Decide based on the count of pathological signals. If the retry result is consistent with the baseline, accept the baseline; if highly pathological and the retry result can be extracted, accept the retry; otherwise, fall back to SC-2 (select answer from two samples).

DRG provides a reproducible experimental framework:

• Multi-GPU Support: Data sharding (parallel processing for large datasets) and model sharding (memory optimization for large models);
• Datasets and Models: Supports mathematical reasoning datasets like MATH-500 and AIME2024, compatible with models such as Qwen3 and distilled versions of DeepSeek-R1;
• Answer Extraction and Scoring: Strip thinking content, extract \boxed{} expressions, and score via string normalization and sympy symbolic verification.

DRG's innovations include:

1. Zero Training Cost: No parameter updates needed, can be applied to off-the-shelf models immediately;
2. Fine-Grained Detection: Capture output quality issues from multiple dimensions;
3. Cost-Quality Tradeoff: Hierarchical strategy (greedy → sampling → SC-2) controls additional overhead;
4. Interpretable Path: Record decision paths for easy diagnosis of failure modes.

DRG has limitations:

• Threshold Sensitivity: Trigger thresholds need to be calibrated for different models/datasets;
• Domain Specificity: Currently adapted for mathematical reasoning, migration requires adjustments;
• Sampling Randomness: May lead to decreased retry quality. Future directions: Explore learning-based triggers, integrate acceleration technologies, and verify effectiveness on large-scale models.

DRG provides a practical solution for reasoning model deployment in resource-constrained scenarios, proving that output quality can be improved without modifying model parameters. Its detailed code and documentation lay the foundation for reproduction and expansion, and future training-free optimization methods will play an important role in real-world deployments.