This paper identifies a critical yet underexplored challenge in distilling from multi-modal large language models (MLLMs): the reasoning trajectories generated by multiple drifting teachers exhibit concept drift, whereby their reasoning distributions evolve unpredictably and transmit biases to the student model, ultimately compromising its performance. To tackle this issue, we pioneer a theoretical connection between concept drift and knowledge distillation, casting the non-stationary reasoning dynamics from multiple MLLM teachers as next-token prediction of multi-stream reasoning trajectories. Guided by concept drift, we introduce the "learn-compare-critique" paradigm, culminating in autonomous preference optimization (APO). Under the active guidance of the teachers, the student model first learns and self-distils preferred thinking by comparing multiple teachers. It then engages in critical reflection over the drifting inference from teachers, performing concept alignment through APO, ultimately yielding a robust, consistent, and generalizable model. Extensive experiments demonstrate our superior performance of consistency, robustness and generalization within knowledge distillation. Besides, we also contributed a large-scale dataset CXR-MAX (Multi-teachers Alignment X-rays), comprising 170,982 distilled reasoning trajectories derived from publicly accessible MLLMs based on MIMIC-CXR. Our code and data are public at: https://anonymous.4open.science/r/

Reinforcement Learning with Verifiable Rewards (RLVR) has become a widely adopted technique for enhancing the reasoning ability of Large Language Models (LLMs). However, the effectiveness of RLVR strongly depends on the capability of base models. This issue arises because it requires the model to have sufficient capability to perform high-quality exploration, which involves both effectiveness and diversity. Unfortunately, existing methods address this issue by imitating expert trajectories, which improve effectiveness but neglect diversity. To address this, we argue that the expert only needs to provide guidance only at critical decision points rather than the entire reasoning path. Based on this insight, we propose MENTOR: Mixed-policy Expert Navigation for Token-level Optimization of Reasoning, a framework that provides expert guidance only at critical decision points to perform effective and diverse exploration in RLVR. Extensive experiments show that MENTOR enables models capture the essence of expert strategies rather than surface imitation, thereby performing high-quality exploration and achieving superior overall performance. Our code is available online.

The rise of Large Language Models has not been inclusive of all cultures. The models are mostly trained on English texts and culture which makes them underperform in other languages and cultural contexts. By developing a generalizable method for preparing culturally relevant datasets and post-training the Gemma 2 model, this project aimed to increase the performance of Gemma 2 for an underrepresented language and showcase how others can do the same to unlock the power of Generative AI in their country and preserve their cultural heritage.

Prior work has shown that a significant driver of performance in reasoning models is their ability to reason and self-correct. A distinctive marker in these reasoning traces is the token wait, which often signals reasoning behavior such as backtracking. Despite being such a complex behavior, little is understood of exactly why models do or do not decide to reason in this particular manner, which limits our understanding of what makes a reasoning model so effective. In this work, we address the question whether model's latents preceding wait tokens contain relevant information for modulating the subsequent reasoning process. We train crosscoders at multiple layers of DeepSeek-R1-Distill-Llama-8B and its base version, and introduce a latent attribution technique in the crosscoder setting. We locate a small set of features relevant for promoting/suppressing wait tokens' probabilities. Finally, through a targeted series of experiments analyzing max activating examples and causal interventions, we show that many of our identified features indeed are relevant for the reasoning process and give rise to different types of reasoning patterns such as restarting from the beginning, recalling prior knowledge, expressing uncertainty, and double-checking.

Metaphor analysis is a complex linguistic phenomenon shaped by context and external factors. While Large Language Models (LLMs) demonstrate advanced capabilities in knowledge integration, contextual reasoning, and creative generation, their mechanisms for metaphor comprehension remain insufficiently explored. This study examines LLMs' metaphor-processing abilities from three perspectives: (1) Concept Mapping: using embedding space projections to evaluate how LLMs map concepts in target domains (e.g., misinterpreting "fall in love" as "drop down from love"); (2) Metaphor-Literal Repository: analyzing metaphorical words and their literal counterparts to identify inherent metaphorical knowledge; and (3) Syntactic Sensitivity: assessing how metaphorical syntactic structures influence LLMs' performance. Our findings reveal that LLMs generate 15\%-25\% conceptually irrelevant interpretations, depend on metaphorical indicators in training data rather than contextual cues, and are more sensitive to syntactic irregularities than to structural comprehension. These insights underline the limitations of LLMs in metaphor analysis and call for more robust computational approaches.

Competitive search is a setting where document publishers modify them to improve their ranking in response to a query. Recently, publishers have increasingly leveraged LLMs to generate and modify competitive content. We introduce Reinforcement Learning from Ranker Feedback (RLRF), a framework that trains LLMs using preference datasets derived from ranking competitions. The goal of a publisher (LLM-based) agent is to optimize content for improved ranking while accounting for the strategies of competing agents. We generate the datasets using approaches that do not rely on human-authored data. We show that our proposed agents consistently and substantially outperform previously suggested approaches for LLM-based competitive document modification. We further show that our agents are effective with ranking functions they were not trained for (i.e., out of distribution) and they adapt to strategic opponents. These findings provide support to the significant potential of using reinforcement learning in competitive search.

Large language models (LLMs) have exhibited significant capabilities in addressing challenging problems throughout various fields, often through the use of agentic workflows that adhere to structured instructions and multi-step procedures. However, designing such workflows demands substantial manual effort, posing challenges to scalability and generalizability. Recent studies have aimed to minimize the human intervention needed for their construction, leading to advances in automated techniques for optimizing agentic workflows. However, current approaches are often constrained by their limited representational capacity, insufficient adaptability, weak scalability, and pairwise comparison paradigm -- issues that stem primarily from a dependence on discrete optimization techniques. To overcome these limitations, we introduce a new score-based preference approach, refereed as SPOGW, which operates directly on cardinal reward signals through group-wise comparison and enables more efficient and stable optimization in a continuous space. SPOGW incorporates Iterative offline GRPO (ioGRPO) with advantage-masked KL divergence (mKL), which regulates training update by placing greater emphasis on the advantageous regions of the policy response. In five benchmark datasets covering mathematical reasoning, coding, and question answering, SPOGW matches or exceeds the performance of current state-of-the-art approaches, presenting a viable and forward-looking methodology for automated generation and optimization of agentic workflows.

Reasoning capability is pivotal for Large Language Models (LLMs) to solve complex tasks, yet achieving reliable and scalable reasoning remains challenging. While Chain-of-Thought (CoT) prompting has become a mainstream approach, existing methods often suffer from uncontrolled generation, insufficient quality, and limited diversity in reasoning paths. Recent efforts leverage code to enhance CoT by grounding reasoning in executable steps, but such methods are typically constrained to predefined mathematical problems, hindering scalability and generalizability. In this work, we propose Caco (Code-Assisted Chain-of-ThOught), a novel framework that automates the synthesis of high-quality, verifiable, and diverse instruction-CoT reasoning data through code-driven augmentation. Unlike prior work, Caco first fine-tunes a code-based CoT generator on existing math and programming solutions in a unified code format, then scales the data generation to a large amount of diverse reasoning traces. Crucially, we introduce automated validation via code execution and rule-based filtering to ensure logical correctness and structural diversity, followed by reverse-engineering filtered outputs into natural language instructions and language CoTs to enrich task adaptability. This closed-loop process enables fully automated, scalable synthesis of reasoning data with guaranteed executability. Experiments on our created Caco-1.3M dataset demonstrate that Caco-trained models achieve strong competitive performance on mathematical reasoning benchmarks, outperforming existing strong baselines. Further analysis reveals that Caco's code-anchored verification and instruction diversity contribute to superior generalization across unseen tasks. Our work establishes a paradigm for building self-sustaining, trustworthy reasoning systems without human intervention.

Conditional Semantic Textual Similarity (C-STS) measures the semantic proximity between text segments under a specific condition, thereby overcoming the ambiguity inherent in traditional STS. However, existing methods are largely confined to discriminative models, failing to fully integrate recent breakthroughs in the NLP community concerning Large Language Models (LLMs) and Reinforcement Learning (RL). RL is a particularly well-suited paradigm for this task, as it can directly optimize the non-differentiable Spearman ranking metric and guide the reasoning process required by C-STS. However, we find that naively applying listwise RL fails to produce meaningful improvements, as the model is overwhelmed by complex, coarse-grained reward signals. To address this challenge, we introduce PoLi-RL, a novel Point-to-List Reinforcement Learning framework. PoLi-RL employs a two-stage curriculum: it first trains the model with simple pointwise rewards to establish fundamental scoring capabilities, then transitions to a hybrid reward that combines pointwise, pairwise, and listwise objectives to refine the model's ability to discern subtle semantic distinctions. Crucially, we propose an innovative Parallel Slice Ranking Reward (PSRR) mechanism that computes ranking rewards in parallel slices, where each slice comprises same-indexed completions from different samples. This provides a precise, differentiated learning signal for each individual completion, enabling granular credit assignment and effective optimization. On the official C-STS benchmark, PoLi-RL achieves a Spearman correlation coefficient of 48.18, establishing a new SOTA for the cross-encoder architecture. As the first work to successfully apply RL to C-STS, our study introduces a powerful and precise paradigm for training LLMs on complex, ranking-based conditional judgment tasks.

The cross-entropy scaling law has long served as a key tool for guiding the development of large language models. It shows that cross-entropy loss decreases in a predictable power-law rate as the model size increases. However, recent evidence indicates that this law breaks down at very large scales: the loss decreases more slowly than expected, which causes significant trouble for developing large language models. In this paper, we hypothesize that the root cause lies in the fact that cross-entropy itself does not truly scale; instead, only one of its hidden components does. To investigate this, we introduce a novel decomposition of cross-entropy into three parts: Error-Entropy, Self-Alignment, and Confidence. We show both theoretically and empirically that this decomposition precisely captures the training dynamics and optimization objectives. Through extensive experiments on multiple datasets and 32 models spanning five orders of magnitude in size, we find that only error-entropy follows a robust power-law scaling, while the other two terms remain largely invariant. Moreover, error-entropy constitutes the dominant share of cross-entropy in small models but diminishes in proportion as models grow larger. This explains why the cross-entropy scaling law appears accurate at small scales but fails at very large ones. Our findings establish the error-entropy scaling law as a more accurate description of model behavior. We believe it will have wide applications in the training, understanding, and future development of large language models.