Abstract
Chain-of-Thought (CoT) reasoning, while effective for large-scale language models, often introduces performance degradation in smaller models due to logical hallucinations and flawed reasoning chains. This paper introduces the SelF-Reasoner (Selective Filtering Reasoner), a framework designed to mitigate these "CoT traps." By employing a selective filter to evaluate the entailment relationship between a question and its generated reasoning, the system dynamically decides whether to trust the reasoning chain or bypass it in favor of a direct answer. Our experiments demonstrate that SelF-Reasoner consistently outperforms standard CoT methods, enabling smaller models like T5 to surpass the performance of significantly larger models on complex reasoning benchmarks.
Introduction
Chain-of-Thought (CoT) prompting—the technique of asking a model to "think step-by-step"—has significantly improved the performance of Large Language Models (LLMs) on complex tasks. However, this benefit does not always scale down. For smaller models, such as the T5 family, CoT can become a liability. These models frequently generate "misleading" reasoning chains characterized by logical leaps, missing "bridging objects," or pure hallucinations.
When a model relies on broken logic to derive a conclusion, the final answer is often incorrect. Furthermore, many questions are "indecomposable"—simple enough that adding a reasoning step merely introduces unnecessary opportunities for error. The challenge lies in capturing the benefits of CoT while filtering out the noise of flawed logic.
The SelF-Reasoner Framework
The SelF-Reasoner addresses this challenge by acting as a cautious editor. Rather than blindly following a generated reasoning chain, the framework utilizes a three-component architecture to verify the validity of the logic before committing to an answer.
1. The Reasoner
The process begins with a fine-tuned model (specifically UnifiedQA-T5) that generates a candidate reasoning chain for a given input question. This chain represents the model's attempt to decompose the problem into logical steps.
2. The CoT Filter
The Filter serves as the system's decision-maker. It is a verifier trained to assess the entailment relationship between the question and the generated reasoning chain. It assigns a label of "valid" or "invalid" based on whether the reasoning logically supports the path to an answer.
3. The Answerer
The Answerer operates in a dual-mode "Switch" mechanism:
- Valid Path: If the Filter deems the reasoning valid, the Answerer extracts the final answer directly from the reasoning chain.
- Invalid Path: If the Filter identifies the reasoning as flawed, the Answerer ignores the chain entirely and predicts the answer directly from the question text.
This mechanism ensures that the model only utilizes CoT when it is confident that the reasoning is sound, using direct prediction as a robust safety net.
Experimental Evaluation
The researchers evaluated the SelF-Reasoner across three distinct reasoning domains:
- ScienceQA: A multimodal science benchmark requiring scientific knowledge and lecture-based reasoning.
- ECQA (Explainable CommonSense QA): A dataset focused on general commonsense reasoning with human-annotated explanations.
- LastLetter: A symbolic reasoning task requiring the concatenation of the last letters of various words.
Key Results
The SelF-Reasoner consistently outperformed both vanilla fine-tuning and standard CoT pipelines across all benchmarks:
- ScienceQA: The "Large" version (770M parameters) achieved 87.24% accuracy, surpassing the previous state-of-the-art (Chameleon/GPT-4) among text-only models and beating the standard CoT pipeline by over 2%.
- ECQA: The model achieved 58.48% accuracy, a 3.5% improvement over traditional two-level pipelines.
- LastLetter: The framework reached 80.06% accuracy, outperforming the standard pipeline by 3.26%.
Human evaluation further validated these results. In a manual check of 50 instances, reasoning chains identified as valid by the filter were logically correct 96% of the time when they led to a correct answer.
Insights and Analysis
The "Bridging Object" Failure
Analysis of model failures revealed that small models often struggle with "bridging objects"—the intermediate concepts that link facts together. For example, a model might correctly identify a feature of an object but, during the reasoning process, accidentally attribute that feature to a different object mentioned in the prompt. SelF-Reasoner successfully identifies these mismatches and reverts to direct prediction.
The Scaling Law of Filtering
The study found that as the size of the filter increases, the overall system accuracy improves. However, there are diminishing returns; a very large generator may produce sophisticated hallucinations that are increasingly difficult for a filter to detect. This suggests that the balance between generator and filter capacity is crucial for performance.
Limitations and Future Work
While effective, the SelF-Reasoner introduces additional complexity by requiring the training of a separate filter, which increases total parameter count and training time. Additionally, "filter leakage" remains a challenge, where incorrect rationales occasionally pass through the verification stage.
Future research may explore "joint training" methods to optimize the reasoner and filter simultaneously. Furthermore, the development of higher-quality, human-annotated reasoning datasets will be essential for training more robust filters across a wider variety of tasks.
Conclusion
The SelF-Reasoner demonstrates that smaller language models can achieve state-of-the-art performance by being selective about their own reasoning. By implementing a filtering mechanism that treats CoT as an optional tool rather than a mandatory path, the framework mitigates the risks of hallucination and provides a scalable way for smaller models to "punch up" against much larger counterparts.
Summary updated