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.

Retrieval-augmented language model (RALM) relies on retrieved external knowledge to generate responses, resulting in vulnerability in the face of retrieval results with noisy documents. Previous works integrate additional filters or finetune Large Language Models (LLMs) to learn adaptive retrieval to reduce the performance damage of noisy documents. However, prior noise filtering may lead to the loss of crucial information, and these methods do not focus on distracting documents with high semantic relevance, which is the most challenging problem. In this study, we propose a training method for fact-centric preference alignment (FPA) to improve the ability of LLMs to directly extract useful information from noisy retrieval results without prior filtering. Our method performs positive document mining based on factual consistency and uses LLMs self-generated synthetic data as training data without manual annotation. We evaluate our FPA on four question answering benchmarks, and the experimental results demonstrate that our method achieves significant improvement with a small scale of training data.