Chain of Thought (CoT) reasoning has demonstrated remarkable deep reasoning capabilities in both large language models (LLMs) and multimodal large language models (MLLMs). However, its reliability is often undermined by the accumulation of errors in intermediate steps. This paper introduces an novel approach to calibrate the CoT reasoning accuracy by leveraging the model's intrinsic veracity encoding. We discover that specific attention head activations reliably reflect the truthfulness of reasoning steps in CoT. Based on this insight, we train a confidence predictor to evaluate the correctness of each reasoning step using these truthfulness-sensitive activations, dynamically selecting the most plausible reasoning path via beam search. Experimental results demonstrate that our method significantly outperforms the state-of-the-art baselines (e.g., Few-Shot CoT, Self-Consistency, and Self-Evaluation Guided Beam Search) across the mathematical, symbolic, and commonsense reasoning tasks, exhibiting superior accuracy and reliability in both unimodal and multimodal settings. We further validate the approach on large reasoning models, confirming its applicability to specialized reasoning models. Additionally, we explore the role of the model's self-correction ability in CoT reasoning. This work provides a novel reliability improvement path for CoT reasoning with broad application potential.

Multilingual large language models (LLMs) often exhibit language confusion, a tendency to generate responses in a dominant language irrespective of the prompt's language. To address this, we propose Smoothie-Qwen, a lightweight, post-hoc method that mitigates language bias without retraining. This technique selectively adjusts token-level output probabilities to effectively suppress undesired language generation. Applied to the Qwen model, our method reduces unintended Chinese output by over 95% while preserving task accuracy on multilingual benchmarks. This work provides a practical and efficient solution for enhancing the language controllability of LLMs, making them more reliable for global applications.

Reward models are critical for improving large language models (LLMs), particularly in reinforcement learning from human feedback (RLHF) or inference-time verification. Current reward modeling typically relies on scores of overall responses to learn the outcome rewards for the responses. However, since the response-level scores are coarse-grained supervision signals, the reward model struggles to identify the specific components within a response trajectory that truly correlate with the scores, leading to poor generalization on unseen responses. In this paper, we propose to leverage generation probabilities to establish reward consistency between processes in the response trajectory, which allows the response-level supervisory signal to propagate across processes, thereby providing additional fine-grained signals for reward learning. Building on analysis under the Bayesian framework, we develop an intra-trajectory consistency regularization to enforce that adjacent processes with higher next-token generation probability maintain more consistent rewards. We apply the proposed regularization to the advanced outcome reward model, improving its performance on RewardBench. Besides, we show that the reward model trained with the proposed regularization induces better DPO-aligned policies and achieves better best-of-N (BON) inference-time verification results. Our code is provided in https://github.com/chaoyang101/ICRM.

The notion of language modeling has gradually shifted in recent years from a distribution over finite-length strings to general-purpose prediction models for textual inputs and outputs, following appropriate alignment phases. This paper analyzes the distinction between distribution estimation and response prediction in the context of LLMs, and their often conflicting goals. We examine the training phases of LLMs, which include pretraining, in-context learning, and preference tuning, and also the common use cases for their output probabilities, which include completion probabilities and explicit probabilities as output. We argue that the different settings lead to three distinct intended output distributions. We demonstrate that NLP works often assume that these distributions should be similar, which leads to misinterpretations of their experimental findings. Our work sets firmer formal foundations for the interpretation of LLMs, which will inform ongoing work on the interpretation and use of LLMs' induced distributions.

Driven by privacy protection laws and regulations, unlearning in Large Language Models (LLMs) is gaining increasing attention. However, current research often neglects the interpretability of the unlearning process, particularly concerning sample-level unlearning difficulty. Existing studies typically assume a uniform unlearning difficulty across samples. This simplification risks attributing the performance of unlearning algorithms to sample selection rather than the algorithm's design, potentially steering the development of LLM unlearning in the wrong direction. Thus, we investigate the relationship between LLM unlearning and sample characteristics, with a focus on unlearning difficulty. Drawing inspiration from neuroscience, we propose a Memory Removal Difficulty ($\mathrm{MRD}$) metric to quantify sample-level unlearning difficulty. Using $\mathrm{MRD}$, we analyze the characteristics of hard-to-unlearn versus easy-to-unlearn samples. Furthermore, we propose an $\mathrm{MRD}$-based weighted sampling method to optimize existing unlearning algorithms, which prioritizes easily forgettable samples, thereby improving unlearning efficiency and effectiveness. We validate the proposed metric and method using public benchmarks and datasets, with results confirming its effectiveness.

Large Language Models (LLMs) have demonstrated remarkable capabilities in various educational tasks, yet their alignment with human learning patterns, particularly in predicting which incorrect options students are most likely to select in multiple-choice questions (MCQs), remains underexplored. Our work investigates the relationship between LLM generation likelihood and student response distributions in MCQs with a specific focus on distractor selections. We collect a comprehensive dataset of MCQs with real-world student response distributions to explore two fundamental research questions: (1). RQ1 - Do the distractors that students more frequently select correspond to those that LLMs assign higher generation likelihood to? (2). RQ2 - When an LLM selects a incorrect choice, does it choose the same distractor that most students pick? Our experiments reveals moderate correlations between LLM-assigned probabilities and student selection patterns for distractors in MCQs. Additionally, when LLMs make mistakes, they are more likley to select the same incorrect answers that commonly mislead students, which is a pattern consistent across both small and large language models. Our work provides empirical evidence that despite LLMs' strong performance on generating educational content, there remains a gap between LLM's underlying reasoning process and human cognitive processes in identifying confusing distractors. Our findings also have significant implications for educational assessment development. The smaller language models could be efficiently utilized for automated distractor generation as they demonstrate similar patterns in identifying confusing answer choices as larger language models. This observed alignment between LLMs and student misconception patterns opens new opportunities for generating high-quality distractors that complement traditional human-designed distractors.

Natural Language Generation (NLG) popularity has increased owing to the progress in Large Language Models (LLMs), with zero-shot inference capabilities. However, most neural systems utilize decoder-only causal (unidirectional) transformer models, which are effective for English but may reduce the richness of languages with less strict word order, subject omission, or different relative clause attachment preferences. This is the first work that analytically addresses optimal text generation order for non-causal language models. We present a novel Viterbi algorithm-based methodology for maximum likelihood word order estimation. We analyze the non-causal most-likelihood order probability for NLG in Spanish and, then, the probability of generating the same phrases with Spanish causal NLG. This comparative analysis reveals that causal NLG prefers English-like SVO structures. We also analyze the relationship between optimal generation order and causal left-to-right generation order using Spearman's rank correlation. Our results demonstrate that the ideal order predicted by the maximum likelihood estimator is not closely related to the causal order and may be influenced by the syntactic structure of the target sentence.

In this work, we show the pre-trained language models return distinguishable generation probability and uncertainty distribution to unfaithfully hallucinated texts, regardless of their size and structure. By examining 24 models on 6 data sets, we find out that 88-98% of cases return statistically significantly distinguishable generation probability and uncertainty distributions. Using this general phenomenon, we showcase a hallucination-reducing training algorithm. Our algorithm outperforms other baselines by achieving higher faithfulness metrics while maintaining sound general text quality measures.