Recent large language models have demonstrated relevant capabilities in solving problems that require logical reasoning; however, the corresponding internal mechanisms remain largely unexplored. In this paper, we show that a small language model can solve a deductive reasoning task by learning the underlying rules (rather than operating as a statistical learner). A low-level explanation of its internal representations and computational circuits is then provided. Our findings reveal that induction heads play a central role in the implementation of the rule completion and rule chaining steps involved in the logical inference required by the task. Introduction Recent Large Language Models (LLMs) have demonstrated remarkable capabilities in reasoning and problem-solving (Huang and Chang, 2023). Many approaches have focused on enhancing logical reasoning in LLMs, with a growing body of work introducing formal and symbolic logic-based benchmarks (Liu et al., 2025). While much of the literature emphasizes solving reasoning benchmarks, comparatively less attention has been devoted to understanding and explaining the underlying low-level computational mechanisms. Y et, interpretability is crucial for designing more robust and targeted models, that are less prone to errors.

Although large language models (LLM) have achieved remarkable performance, their enormous parameter counts hinder deployment on resource-constrained hardware. Low-rank compression can reduce both memory usage and computational demand, but applying a uniform compression ratio across all layers often leads to significant performance degradation, and previous methods perform poorly during decoding. To address these issues, we propose the Fine-grained Low-Rank Compressor (FLRC), which efficiently determines an optimal rank allocation for each layer, and incorporates progressive low-rank decoding to maintain text generation quality. Comprehensive experiments on diverse benchmarks demonstrate the superiority of FLRC, achieving up to a 17% improvement in ROUGE-L on summarization tasks compared to state-of-the-art low-rank compression methods, establishing a more robust and efficient framework to improve LLM inference.

A prompt is a natural language instruction that defines a specific task for a large language model (LLM) and serves as the primary interface for human-LLM interaction. With the growing deployment of LLMs, diverse prompt datasets are emerging from platforms such as GitHub and social media. These datasets span a wide array of applications and content types, facilitating both broader LLM utilization and improved prompt engineering. In this work, we--for the first time--have compiled an extensive list of prompt datasets sourced from various channels, representing a spectrum of downstream tasks, languages, engineering techniques, attributes, and modalities. We select key representative datasets for systematic analysis, revealing commonalities and differences in prompt construction across categories, distinguishing them from other text corpora like literature and web. We further propose a prompt optimization approach that leverages syntactic embeddings of part-of-speech and dependency structures. By identifying a centroid representation of prompts and guiding LLMs to rewrite prompts toward this centroid, our method improves the meaningfulness of model outputs. We have made our datasets and code available.

Current Chain-of-Thought (CoT) verification methods predict reasoning correctness based on outputs (black-box) or activations (gray-box), but offer limited insight into why a computation fails. We introduce a white-box method: Circuit-based Reasoning Verification (CRV). We hypothesize that attribution graphs of correct CoT steps, viewed as execution traces of the model's latent reasoning circuits, possess distinct structural fingerprints from those of incorrect steps. By training a classifier on structural features of these graphs, we show that these traces contain a powerful signal of reasoning errors. Our white-box approach yields novel scientific insights unattainable by other methods. (1) We demonstrate that structural signatures of error are highly predictive, establishing the viability of verifying reasoning directly via its computational graph. (2) We find these signatures to be highly domain-specific, revealing that failures in different reasoning tasks manifest as distinct computational patterns. (3) We provide evidence that these signatures are not merely correlational; by using our analysis to guide targeted interventions on individual transcoder features, we successfully correct the model's faulty reasoning. Our work shows that, by scrutinizing a model's computational process, we can move from simple error detection to a deeper, causal understanding of LLM reasoning.

Diffusion large language models (dLLMs) present a promising alternative to dominant autoregressive models (ARMs) by the ability of parallel decoding at the expense of substantial computation and memory costs. Specifically, the cache mechanism for bidirectional attention in dLLMs demands large memory footprint, restricting their ability to handle long contexts under resource-limited settings. Existing cache eviction strategies are designed for ARMs and ignore the unique characteristics of dLLMs, thus leading to unsatisfactory performance. To address these challenges, we introduce MaskKV, a training-free cache eviction framework tailored to dLLMs, focusing on the effect of mask tokens in dLLMs. MaskKV is built on two key innovations: (1) a mask-query guided scoring mechanism that leverages attention weights to identify and evict less critical prompt tokens for each head; (2) an adaptive cache budgeting strategy that improves efficiency by reducing allocation in intermediate layers and concentrating resources on prompt-preferring heads. On LLaDA with MaskKV, compressing the KV cache to only 256 pairs (less than 5% of tokens) retains 94% of the full-cache performance on LongBench and achieves up to 31x acceleration at 32k prompt length. The code is publicly available at: https://github.com/jianuo-huang/MaskKV

Geometric reasoning remains a core challenge for Multimodal Large Language Models (MLLMs). Even the most advanced closed-source systems, such as GPT-O3 and Gemini-2.5-Pro, still struggle to solve geometry problems reliably, despite exhibiting strong textual reasoning abilities on tasks like the International Mathematical Olympiad (IMO). This gap suggests that the bottleneck lies in understanding geometric diagrams rather than reasoning itself. Since geometric figures can often be faithfully described in concise textual form, converting visual content into captions offers a promising direction. Motivated by this insight, we introduce CapGeo, a caption-assisted reasoning framework that bridges visual and textual modalities. Experiments show substantial improvements when models are equipped with captions: Qwen2.5-VL-72B improves from 8.6% (vision-only) to 59.0%, while Claude-Opus-4 rises from 44.8% to 73.0%. To systematically evaluate and identify high-quality geometric captioning models, we further propose CapGeo-Bench, a dataset of 4,641 curated figure-caption pairs. Crucially, CapGeo-Bench incorporates a keypoint-based evaluation metric that correlates strongly with downstream CapGeo performance, enabling reliable assessment of geometric captioning ability. Together, our framework and benchmark highlight a new pathway toward advancing geometric reasoning in MLLMs.

Reliable evaluation is fundamental to the progress of Large Language Models (LLMs), yet the evaluation process during pre-training is plagued by significant instability that obscures true learning dynamics. In this work, we systematically diagnose this instability, attributing it to two distinct sources: \textit{Parameter Instability} from training stochasticity and \textit{Evaluation Instability} from noisy measurement protocols. To counteract both sources of noise, we introduce \textbf{MaP}, a dual-pronged framework that synergistically integrates checkpoint \underline{M}erging \underline{a}nd the \underline{P}ass@k metric. Checkpoint merging smooths the parameter space by averaging recent model weights, while Pass@k provides a robust, low-variance statistical estimate of model capability. Extensive experiments show that MaP yields significantly smoother performance curves, reduces inter-run variance, and ensures more consistent model rankings. Ultimately, MaP provides a more reliable and faithful lens for observing LLM training dynamics, laying a crucial empirical foundation for LLM research.

Training expert LLMs in domains with scarce data is difficult, often relying on multiple-choice questions (MCQs). However, standard outcome-based reinforcement learning (RL) on MCQs is risky. While it may improve accuracy, we observe it often degrades reasoning quality such as logical consistency. Existing solutions to supervise reasoning, such as large-scale Process Reward Models (PRMs), are prohibitively expensive. To address this, we propose CLARity, a cost-effective RL framework that enhances reasoning quality using only a small, general-purpose LLM. CLARity integrates a consistency-aware reward mechanism with a 2-stage refine-then-monitor training pipeline to enhance reasoning consistency, and a dynamic data reformulation strategy to to better exploit limited data. Experiments demonstrate that CLARity improves response consistency by 16.5% and accuracy by 7.5% over baselines. Human evaluations further confirm holistic improvements in coherence and professionalism. Thus, CLARity offers a generalizable solution that enables smaller models to effectively guide expert models by reasoning consistency.Our code is open sourced at: https://github.com/Infinite-set/CLARity

Medical diagnostics is a high-stakes and complex domain that is critical to patient care. However, current evaluations of large language models (LLMs) are fundamentally misaligned with real-world clinical practice. Most of them rely on static benchmarks derived from public medical exam items, which tend to overestimate model performance and ignore the difference between textbook cases and the ambiguous, varying conditions in the real world. Recent efforts toward dynamic evaluation offer a promising alternative, but their improvements are limited to superficial perturbations and a narrow focus on accuracy. To address these gaps, we propose DyReMe, a dynamic benchmark for medical diagnostics that better reflects real clinical practice. Unlike static exam-style questions, DyReMe generates fresh, consultation-like cases that introduce distractors such as differential diagnoses and common misdiagnosis factors. It also varies expression styles to mimic diverse real-world query habits. Beyond accuracy, DyReMe evaluates LLMs on three additional clinically relevant dimensions: veracity, helpfulness, and consistency. Our experiments demonstrate that this dynamic approach yields more challenging and realistic assessments, revealing significant misalignments between the performance of state-of-the-art LLMs and real clinical practice. These findings highlight the urgent need for evaluation frameworks that better reflect the demands of trustworthy medical diagnostics.

Long-horizon tasks that require sustained reasoning and multiple tool interactions remain challenging for LLM agents: small errors compound across steps, and even state-of-the-art models often hallucinate or lose coherence. We identify context management as the central bottleneck -- extended histories cause agents to overlook critical evidence or become distracted by irrelevant information, thus failing to replan or reflect from previous mistakes. To address this, we propose COMPASS (Context-Organized Multi-Agent Planning and Strategy System), a lightweight hierarchical framework that separates tactical execution, strategic oversight, and context organization into three specialized components: (1) a Main Agent that performs reasoning and tool use, (2) a Meta-Thinker that monitors progress and issues strategic interventions, and (3) a Context Manager that maintains concise, relevant progress briefs for different reasoning stages. Across three challenging benchmarks -- GAIA, BrowseComp, and Humanity's Last Exam -- COMPASS improves accuracy by up to 20% relative to both single- and multi-agent baselines. We further introduce a test-time scaling extension that elevates performance to match established DeepResearch agents, and a post-training pipeline that delegates context management to smaller models for enhanced efficiency.