Chain-of-Thought

“errors can propagate through the reasoning chain, and there is no principled mechanism to revisit decisions or explore alternative strategies.”

“However, despite these advances, LLMs frequently exhibit limitations in their logical consistency, accuracy, and self-correction abilities when confronted with highly intricate or chaotic reasoning problems”

“However, CoT is fundamentally linear: once a reasoning step is generated, the model commits to it, often propagating early errors into final failures.”

“Studies have shown that simple approaches like CoT are inadequate for tasks that demand decomposition into sub-tasks”

“sprague2025to reported that on the Massive Multitask Language Understanding (MMLU) benchmark hendrycks2021measuring, 95\% of the performance gain from CoT is attributed to questions involving symbolic reasoning.”

“Nevertheless, CoT alone does not guarantee the factual correctness of the underlying statements within the reasoning chain.”

“Chain-of-Thought (CoT) prompting~wei2022chain, zhang2023multimodal, lyu2023faithful, while enhancing reasoning, may not ensure its steps visually align with the image and can be sensitive to setup or resource-intensive.”

“Although CoT has led to remarkable achievements, it does not always provide positive outcomes and sometimes hinders reasoning performance”

“it still relies on a relatively simple, linear flow of thought, which can become limiting for tasks involving more complex reasoning”

“In this paper, we reveal a strikingly counterintuitive finding: Chain-of-Thought prompting unexpectedly degrades LLM performances in certain problem-solving contexts.”

“this capability typically relies on explicit supervision (annotated CoT data), is confined to a discrete token space, and is ultimately capped by the base model's pre-trained capabilities”

“CoT's linear structure fails to capture the branching nature of mathematical reasoning, where multiple strategies are considered, partial arguments explored, and errors necessitate backtracking”