Self-correction has emerged as a promising paradigm for enhancing LLM reasoning, yet recent studies reveal a troubling phenomenon: without external oracle feedback, iterative reflection often degrades rather than improves performance. Our analysis uncovers that toxic reflection, where correct answers deteriorate into incorrect ones, occurs at rates comparable to effective corrections. We identify the root cause as the shortsightedness of existing methods: models focus only on the previous output while neglecting the global reasoning state, rendering them unable to quantify uncertainty. To address this, we propose two complementary reflection signals — Reflecting Entropy for snapshot disagreement among reasoning trajectories and Belief Shift for inter-round belief revision — paired with a two-stage hierarchical controller that schedules reflection adaptively. We instantiate these ideas as REGAR (Reflecting Entropy-Guided Adaptive Reflection), a framework that maintains a pool of candidate reasoning trajectories and applies a single, uniform reflection primitive when correction is needed. Experiments on mathematical and commonsense reasoning demonstrate consistent improvements over strong baselines while reducing computational overhead, highlighting both effectiveness and efficiency.

Self-reflection for Large LanguageModels (LLMs) has gained significant attention. Existing approaches involve models iterating and improving their previous responses based on LLMs’ internal reflection ability or external feedback. However, recent research has raised doubts about whether intrinsic self-correction without external feedback may even degrade performance. Based on our empirical evidence, we find that current static reflection methods may lead to redundant, drift, and stubborn issues. To mitigate this, we introduce Instruct-of-Reflection (IoRT), a novel and general reflection framework that leverages dynamic-meta instruction to enhance the iterative reflection capability of LLMs. Specifically, we propose the instructor driven by the meta-thoughts and self-consistency classifier, generates various instructions, including refresh, stop, and select, to guide the next reflection iteration. Our experiments demonstrate that IoRT achieves an average improvement of 10.1% over established baselines in mathematical and commonsense reasoning tasks, highlighting its efficacy and applicability. Our code is available at https://github.com/llp635/IoRT.