How do vision transformers (ViTs) represent and process the world? This paper addresses this long-standing question through the first systematic analysis of 6.6K features across all layers, extracted via sparse autoencoders, and by introducing the residual replacement model, which replaces ViT computations with interpretable features in the residual stream. Our analysis reveals not only a feature evolution from low-level patterns to high-level semantics, but also how ViTs encode curves and spatial positions through specialized feature types. The residual replacement model scalably produces a faithful yet parsimonious circuit for human-scale interpretability by significantly simplifying the original computations. As a result, this framework enables intuitive understanding of ViT mechanisms. Finally, we demonstrate the utility of our framework in debiasing spurious correlations.

We introduce FinDebate, a multi-agent framework for financial analysis, integrating collaborative debate with domain-specific Retrieval-Augmented Generation (RAG). Five specialized agents, covering earnings, market, sentiment, valuation, and risk, run in parallel to synthesize evidence into multi-dimensional insights. To mitigate overconfidence and improve reliability, we introduce a safe debate protocol that enables agents to challenge and refine initial conclusions while preserving coherent recommendations. Experimental results, based on both LLM-based and human evaluations, demonstrate the framework's efficacy in producing high-quality analysis with calibrated confidence levels and actionable investment strategies across multiple time horizons.

LLMs suffer from critical reasoning issues such as unfaithfulness, bias, and inconsistency, since they lack robust causal underpinnings and may rely on superficial correlations rather than genuine understanding. Successive LRMs have emerged as a promising alternative, leveraging advanced training techniques such as reinforcement learning (RL) and distillation to improve task accuracy. However, the impact of these training methods on causality remains largely unexplored. In this study, we conduct a systematic causal analysis on LLMs and LRMs, examining structural causal models (SCMs) of four key variables: problem instruction (Z), thinking process (T), reasoning steps (X), and answer (Y). Our findings reveal that RLVR-trained LRMs exhibit enhanced causal reasoning capabilities, aligning more closely with ideal causal structures, while LLMs and distilled LRMs fail to address causality-related deficiencies. Our further investigation indicates that RLVR reduces spurious correlations and strengthens genuine causal patterns, thereby mitigating unfaithfulness and bias. In addition, our inspection on the dynamics of the RLVR training process observes a high correlation between reduced spurious features and improved causal structures, where the causal relationships consistently improve in the training process. This study contributes to the understanding of causality in reasoning models, highlights the critical role of RLVR in enhancing causal reasoning, and provides insights for designing future AI systems with stronger causal foundations. We release our code and data at https://github.com/Harryking1999/CoT_Causal_Analysis.

We present a comparative analysis of text complexity across domains using scale-free metrics. We quantify linguistic complexity via Heaps' exponent $β$ (vocabulary growth), Taylor's exponent $α$ (word-frequency fluctuation scaling), compression rate $r$ (redundancy), and entropy. Our corpora span three domains: legal documents (statutes, cases, deeds) as a specialized domain, general natural language texts (literature, Wikipedia), and AI-generated (GPT) text. We find that legal texts exhibit slower vocabulary growth (lower $β$) and higher term consistency (higher $α$) than general texts. Within legal domain, statutory codes have the lowest $β$ and highest $α$, reflecting strict drafting conventions, while cases and deeds show higher $β$ and lower $α$. In contrast, GPT-generated text shows the statistics more aligning with general language patterns. These results demonstrate that legal texts exhibit domain-specific structures and complexities, which current generative models do not fully replicate.

Automating radiology report generation poses a dual challenge: building clinically reliable systems and designing rigorous evaluation protocols. We introduce a multi-agent reinforcement learning framework that serves as both a benchmark and evaluation environment for multimodal clinical reasoning in the radiology ecosystem. The proposed framework integrates large language models (LLMs) and large vision models (LVMs) within a modular architecture composed of ten specialized agents responsible for image analysis, feature extraction, report generation, review, and evaluation. This design enables fine-grained assessment at both the agent level (e.g., detection and segmentation accuracy) and the consensus level (e.g., report quality and clinical relevance). We demonstrate an implementation using chatGPT-4o on public radiology datasets, where LLMs act as evaluators alongside medical radiologist feedback. By aligning evaluation protocols with the LLM development lifecycle, including pretraining, finetuning, alignment, and deployment, the proposed benchmark establishes a path toward trustworthy deviance-based radiology report generation.

Continual learning (CL) is essential for deploying large language models (LLMs) in dynamic real-world environments without the need for costly retraining. Recent model merging-based methods have attracted significant attention, but they still struggle to effectively manage the trade-off between learning new knowledge and preventing forgetting, a challenge largely stemming from suboptimal number of merges and merging frequency. In this paper, we introduce Adaptive Iterative Model Merging (AimMerging), a novel CL framework that utilizes learning and forgetting signals from the training trajectory to dynamically monitor the model's training status. Guided by dynamic monitoring, the training trajectory-guided merge controller adaptively determines the timing and frequency of iterative fusion, while the rehearsal-based knowledge fusion module computes the merging weights and executes the fusion. Comprehensive experiments on three CL benchmarks with various model sizes (from 770M to 13B) demonstrate that AimMerging achieves significant performance improvements over existing state-of-the-art methods, with an average relative improvement of 80% and 59% on FWT and BWT, respectively. The source code is provided for reproducibility.

Increasing complexity in software systems places a growing demand on reasoning tools that unlock vulnerabilities manifest in source code. Many current approaches focus on vulnerability analysis as a classifying task, oversimplifying the nuanced and context-dependent real-world scenarios. Even though current code large language models (LLMs) excel in code understanding, they often pay little attention to security-specific reasoning. We propose LLaVul, a multimodal LLM tailored to provide fine-grained reasoning about code through question-answering (QA). Our model is trained to integrate paired code and natural queries into a unified space, enhancing reasoning and context-dependent insights about code vulnerability. To evaluate our model performance, we construct a curated dataset of real-world vulnerabilities paired with security-focused questions and answers. Our model outperforms state-of-the-art general-purpose and code LLMs in the QA and detection tasks. We further explain decision-making by conducting qualitative analysis to highlight capabilities and limitations. By integrating code and QA, LLaVul enables more interpretable and security-focused code understanding.

Tool-augmented large language models (LLMs), hereafter LLM agents, leverage external tools to solve diverse tasks and interface with the real world. However, current training practices largely rely on supervised fine-tuning (SFT) over static trajectories or reinforcement learning (RL) on narrow tasks, and generalize poorly beyond development settings, leading to brittleness with new tools and unseen workflows. Because code execution reflects many structures of real-world workflows, coding problems provide a natural basis for building agent training environments. Motivated by this, we introduce CodeGym, a scalable framework that synthesizes diverse, verifiable, and controllable multi-turn tool-use environments for agent RL, enabling LLM agents to explore and master various workflows actively. CodeGym rewrites static coding problems into interactive environments by extracting atomic functions or logic into callable tools, yielding verifiable tasks that span various tool-execution workflows. Models of varying sizes and chain-of-thought configurations, trained in CodeGym, exhibit consistent out-of-distribution generalizability; for example, Qwen2.5-32B-Instruct achieves an absolute accuracy gain of 8.7 points on the OOD benchmark $τ$-Bench. These results highlight CodeGym as a step toward scalable general-purpose RL environments that align with real-world agent workflows.

Most languages lack sufficient data for large-scale monolingual pretraining, creating a "data wall." Multilingual pretraining helps but is limited by language imbalance and the "curse of multilinguality." An alternative is to translate high-resource text with machine translation (MT), which raises three questions: (1) How does MT-derived data scale with model capacity? (2) Can source-side transformations (e.g., simplifying English with an LLM) improve generalization to native text? (3) How well do models pretrained on MT-derived data adapt when continually trained on limited native text? We investigate these questions by translating English into Indonesian and Tamil--two typologically distant, lower-resource languages--and pretraining GPT-2 models (124M-774M) on native or MT-derived corpora from raw and LLM-simplified English. We evaluate cross-entropy loss on native text, along with accuracy on syntactic probes and downstream tasks. Our results show that (1) MT-pretrained models benefit from scaling; (2) source-side simplification harms generalization to native text; and (3) adapting MT-pretrained models on native text often yields better performance than native-only models, even with less native data. However, tasks requiring cultural nuance (e.g., toxicity detection) demand more exposure to native data.

Cognitive distortions have been closely linked to mental health disorders, yet their automatic detection remained challenging due to contextual ambiguity, co-occurrence, and semantic overlap. We proposed a novel framework that combines Large Language Models (LLMs) with Multiple-Instance Learning (MIL) architecture to enhance interpretability and expression-level reasoning. Each utterance was decomposed into Emotion, Logic, and Behavior (ELB) components, which were processed by LLMs to infer multiple distortion instances, each with a predicted type, expression, and model-assigned salience score. These instances were integrated via a Multi-View Gated Attention mechanism for final classification. Experiments on Korean (KoACD) and English (Therapist QA) datasets demonstrate that incorporating ELB and LLM-inferred salience scores improves classification performance, especially for distortions with high interpretive ambiguity. Our results suggested a psychologically grounded and generalizable approach for fine-grained reasoning in mental health NLP.