Abstract
Large Language Models (LLMs) often fail complex reasoning tasks not due to a lack of logical capability, but because they misinterpret or overlook critical constraints within the prompt—a phenomenon termed "factual drift." This paper introduces SIFT (Stick to the Facts), a training-free, post-training framework that leverages inference-time compute to ground reasoning in structured context. By generating and iteratively refining a "Sticker"—a structured summary of essential conditions—SIFT significantly enhances reasoning performance. Notably, SIFT enables DeepSeek-R1 to achieve a state-of-the-art 85.67% accuracy on the AIME2024 benchmark and provides substantial gains for smaller models like Llama 3.2-3B.
Introduction
While recent advancements in reasoning models, such as OpenAI’s o1 and DeepSeek-R1, have focused on improving the "reasoning path" (the how), they remain vulnerable to errors in problem interpretation (the what). Factual drift occurs when a model systematically misinterprets semantic relationships—such as confusing a unit rate with a total cost—or ignores specific constraints within a query. SIFT addresses this by forcing the model to externalize and verify its understanding of the problem context before and during the reasoning process.
The SIFT Framework
SIFT introduces a systematic loop of four core operations designed to ensure the model is solving the correct problem. This framework utilizes extra tokens at test-time to "think" about the problem definition more clearly.
- Sticker Generation (SG): The model summarizes the input query into a structured "Sticker." This artifact contains atomic conditions, variables, and the core question, serving as a form of external memory.
- Consensus Prediction (CP): The model generates two independent answers: one derived solely from the Sticker and another derived from the original query combined with the Sticker. A discrepancy between these answers signals a fundamental misunderstanding of the context.
- Forward Optimization (FO): If a consensus is not reached, the model refines the Sticker to better align with the original query's semantics, specifically targeting omissions or errors identified during the comparison.
- Inverse Generation (IG): To ensure the Sticker is optimized for the model's internal processing, SIFT asks the model to "reverse-engineer" the facts from its own prediction. This aligns the Sticker with the model's inherent reasoning preferences.
This iterative process continues until the model achieves consensus or reaches a predefined computational limit.
Experimental Evaluation
Datasets and Models
The authors evaluated SIFT across a diverse array of benchmarks and model architectures:
- Benchmarks: GSM8K (grade school math), MATH-500 (high school competition math), AIME 2024 & 2025 (Olympiad-level math), and GPQA-Diamond (expert-level science).
- Models: Testing ranged from small-scale models like Llama 3.2-3B to state-of-the-art reasoning models like DeepSeek-R1.
Key Results
SIFT demonstrated consistent performance improvements across the board:
- DeepSeek-R1: The model’s pass@1 accuracy on AIME2024 rose from 78.33% to 85.67%, a new record for open-source models. On MATH-500, accuracy improved from 97.3% to 98.33%.
- Small Model Gains: Llama 3.2-3B showed remarkable resilience when paired with SIFT, seeing a +8.8% improvement on MATH-500 and a +4.86% jump on GSM8K.
- Test-Time Scaling: A clear trend emerged showing that as the model spends more tokens refining the Sticker through successive stages, accuracy consistently increases, validating the benefit of inference-time compute for context grounding.
Discussion
The Sticker as External Memory
Much like a human mathematician using a scratchpad to list known variables before solving an equation, the Sticker forces the LLM to commit to a specific set of facts. This reduces the stochastic nature of self-verification found in standard reasoning paths.
Efficiency vs. Self-Consistency
Traditional Self-Consistency (SC) methods require sampling numerous full reasoning paths and using majority voting to find the correct answer. SIFT achieves comparable or superior accuracy with significantly fewer total output tokens. By fixing the problem definition early, SIFT prevents the model from wasting compute on reasoning paths derived from a flawed premise.
Conclusion and Future Work
SIFT provides a robust, training-free method for improving LLM reliability in complex reasoning tasks. While the current implementation increases inference latency due to the iterative refinement loop, it offers a clear path toward more dependable AI. Future research may focus on "internalizing" the SIFT behavior through supervised fine-tuning, allowing models to generate these Stickers natively within their hidden thought processes. Additionally, expanding the application of SIFT to non-mathematical domains, such as open-ended summarization or creative writing, remains a promising area for exploration.
Summary updated