Open-ended evaluation is essential for deploying large language models in real-world settings. In studying HealthBench, we observe that using the model itself as a grader and generating rubric-based reward signals substantially improves reasoning performance. Remarkably, the trained model also becomes a stronger grader. Motivated by this, we introduce Self-Rewarding Rubric-Based Reinforcement Learning for Open-Ended Reasoning, a lightweight framework that enables faster and more resource-efficient training while surpassing baselines. Remarkably, on Qwen3-32B, training with just the 4000-sample HealthBench Easy subset is sufficient to obtain a model that exceeds GPT-5 on HealthBench Hard. Incorporating a small amount of teacher-graded data further enhances performance for less capable models.

As Vision Language Models (VLMs) are deployed across safety-critical applications, understanding and controlling their behavioral patterns has become increasingly important. Existing behavioral control methods face significant limitations: system prompting approaches could easily be overridden by user instructions, while applying activation-based steering vectors requires invasive runtime access to model internals, precluding deployment with API-based services and closed-source models. Finding steering methods that transfer across multiple VLMs is still an open area of research. To this end, we introduce universal visual input based steering for output redirection (VISOR++), to achieve behavioral control through optimized visual inputs alone. We demonstrate that a single VISOR++ image can be generated for an ensemble of VLMs to emulate each of their steering vectors. By crafting universal visual inputs that induce target activation patterns, VISOR++ eliminates the need for runtime model access while remaining deployment-agnostic. This means that when an underlying model supports multimodal capability, model behaviors can be steered by inserting an image input replacing runtime steering vector based interventions. We first demonstrate the effectiveness of the VISOR++ images on open-access models such as LLaVA-1.5-7B and IDEFICS2-8B along three alignment directions: refusal, sycophancy and survival instinct. Both the model-specific steering images and the jointly optimized images achieve performance parity closely following that of steering vectors for both positive and negative steering tasks. We also show the promise of VISOR++ images in achieving directional behavioral shifts for unseen models including both open-access and closed-access ones. Furthermore, VISOR++ images are able to preserve 99.9% performance on 14,000 unrelated MMLU evaluation tasks.

Calibrated confidence estimates are necessary for large language model (LLM) outputs to be trusted by human users. While LLMs can express their confidence in human-interpretable ways, verbalized LLM-generated confidence scores have empirically been found to be miscalibrated, reporting high confidence on instances with low accuracy and thereby harming trust and safety. We hypothesize that this overconfidence often stems from a given LLM's heightened suggestibility when faced with claims that it encodes little information about; we empirically validate this hypothesis, finding more suggestibility on lower-accuracy claims. To further improve calibration, we leverage generator-validator disagreement, augmenting normalized validator confidence with a consistency-based estimate of generator confidence. Users often rely on information obtained from these models to make important decisions, but the information is not always accurate. Thus, we seek to qualify LLM responses with confidence estimates that are calibrated, i.e. match the probability of correctness. Users and agentic frameworks often use LLMs in a zero-shot manner without task-specific tuning (Manakul et al., 2023; Geng et al., 2024; Feng et al., 2024; Shorinwa et al., 2025), motivating the development of confidence estimation methods that work in off-the-shelf settings - both gray-box settings with logit access, and black-box settings with only textual input and output. In these settings, verbalized confidence is a simple and commonly-used approach that prompts the model to report its confidence in an answer (Lin et al., 2022; Xiong et al., 2024; Wei et al., 2024). For brevity, we use verbalized confidence as a blanket term for (1) asking the model to decode a numerical confidence like "80%" (Tian et al., 2023) and (2) asking the model whether an answer is correct and taking P(True) (Kadavath et al., 2022). V erbalized confidence is appealing for several reasons, including that it resembles one way humans express confidence, making it easy to interpret and integrate into decision-theoretic frameworks (Sun et al., 2025; Steyvers et al., 2025). However, verbalized confidence has several drawbacks. First, it empirically tends to exhibit overconfidence (Tian et al., 2023; Xiong et al., 2024; Wei et al., 2024; Xu et al., 2025); Figure 1 (left) shows that verbalized confidence scores generally outstrip average accuracy within a confidence bin. For each bar, we label the number of instances whose confidence falls in the interval and we darken larger bins. In other words, no rejection threshold can be chosen to reject a high proportion of false claims.

Retrieval-augmented generation (RAG) is a powerful paradigm for improving large language models (LLMs) on knowledge-intensive question answering. Graph-based RAG (GraphRAG) leverages entity-relation graphs to support multi-hop reasoning, but most systems still rely on static retrieval. When crucial evidence, especially bridge documents that connect disjoint entities, is absent, reasoning collapses and hallucinations persist. Iterative retrieval, which performs multiple rounds of evidence selection, has emerged as a promising alternative, yet its role within GraphRAG remains poorly understood. We present the first systematic study of iterative retrieval in GraphRAG, analyzing how different strategies interact with graph-based backbones and under what conditions they succeed or fail. Our findings reveal clear opportunities: iteration improves complex multi-hop questions, helps promote bridge documents into leading ranks, and different strategies offer complementary strengths. At the same time, pitfalls remain: naive expansion often introduces noise that reduces precision, gains are limited on single-hop or simple comparison questions, and several bridge evidences still be buried too deep to be effectively used. Together, these results highlight a central bottleneck, namely that GraphRAG's effectiveness depends not only on recall but also on whether bridge evidence is consistently promoted into leading positions where it can support reasoning chains. To address this challenge, we propose Bridge-Guided Dual-Thought-based Retrieval (BDTR), a simple yet effective framework that generates complementary thoughts and leverages reasoning chains to recalibrate rankings and bring bridge evidence into leading positions. BDTR achieves consistent improvements across diverse GraphRAG settings and provides guidance for the design of future GraphRAG systems.

As personalized learning gains increasing attention in mathematics education, there is a growing demand for intelligent systems that can assess complex student responses and provide individualized feedback in real time. In this study, we present a personalized auto-grading and feedback system for constructive geometry tasks, developed using large language models (LLMs) and deployed on the Algeomath platform, a Korean online tool designed for interactive geometric constructions. The proposed system evaluates student-submitted geometric constructions by analyzing their procedural accuracy and conceptual understanding. It employs a prompt-based grading mechanism using GPT-4, where student answers and model solutions are compared through a few-shot learning approach. Feedback is generated based on teacher-authored examples built from anticipated student responses, and it dynamically adapts to the student's problem-solving history, allowing up to four iterative attempts per question. The system was piloted with 79 middle-school students, where LLM-generated grades and feedback were benchmarked against teacher judgments. Grading closely aligned with teachers, and feedback helped many students revise errors and complete multi-step geometry tasks. While short-term corrections were frequent, longer-term transfer effects were less clear. Overall, the study highlights the potential of LLMs to support scalable, teacher-aligned formative assessment in mathematics, while pointing to improvements needed in terminology handling and feedback design.

Multimodal large language models (MLLMs) have demonstrated remarkable capabilities in processing and reasoning over diverse modalities, but their advanced abilities also raise significant privacy concerns, particularly regarding Personally Identifiable Information (PII) leakage. While relevant research has been conducted on single-modal language models to some extent, the vulnerabilities in the multimodal setting have yet to be fully investigated. In this work, we investigate these emerging risks with a focus on vision language models (VLMs), a representative subclass of MLLMs that covers the two modalities most relevant for PII leakage, vision and text. We introduce a concept-guided mitigation approach that identifies and modifies the model's internal states associated with PII-related content. Our method guides VLMs to refuse PII-sensitive tasks effectively and efficiently, without requiring re-training or fine-tuning. We also address the current lack of multimodal PII datasets by constructing various ones that simulate real-world scenarios. Experimental results demonstrate that the method can achieve an average refusal rate of 93.3% for various PII-related tasks with minimal impact on unrelated model performances. We further examine the mitigation's performance under various conditions to show the adaptability of our proposed method.

Large language models (LLMs) are increasingly used in newsroom workflows, but their tendency to hallucinate poses risks to core journalistic practices of sourcing, attribution, and accuracy. We evaluate three widely used tools - ChatGPT, Gemini, and NotebookLM - on a reporting-style task grounded in a 300-document corpus related to TikTok litigation and policy in the U.S. We vary prompt specificity and context size and annotate sentence-level outputs using a taxonomy to measure hallucination type and severity. Across our sample, 30% of model outputs contained at least one hallucination, with rates approximately three times higher for Gemini and ChatGPT (40%) than for NotebookLM (13%). Qualitatively, most errors did not involve invented entities or numbers; instead, we observed interpretive overconfidence - models added unsupported characterizations of sources and transformed attributed opinions into general statements. These patterns reveal a fundamental epistemological mismatch: While journalism requires explicit sourcing for every claim, LLMs generate authoritative-sounding text regardless of evidentiary support. We propose journalism-specific extensions to existing hallucination taxonomies and argue that effective newsroom tools need architectures that enforce accurate attribution rather than optimize for fluency.

Large audio-language models (LALMs) exhibit strong zero-shot performance across speech tasks but struggle with speech emotion recognition (SER) due to weak paralinguistic modeling and limited cross-modal reasoning. We propose Compositional Chain-of-Thought Prompting for Emotion Reasoning (CCoT-Emo), a framework that introduces structured Emotion Graphs (EGs) to guide LALMs in emotion inference without fine-tuning. Each EG encodes seven acoustic features (e.g., pitch, speech rate, jitter, shimmer), textual sentiment, keywords, and cross-modal associations. Embedded into prompts, EGs provide interpretable and compositional representations that enhance LALM reasoning. Experiments across SER benchmarks show that CCoT-Emo outperforms prior SOTA and improves accuracy over zero-shot baselines.

Large Language Models (LLMs) have achieved significant advances in reasoning tasks. A key approach is tree-based search with verifiers, which expand candidate reasoning paths and use reward models to guide pruning and selection. Although effective in improving accuracy, these methods are not optimal in terms of efficiency: they perform simple decomposition on the reasoning process, but ignore the planning-execution nature of tasks such as math reasoning or code generation. This results in inefficient exploration of reasoning process. To address this, we propose a dual-phase test-time scaling framework that explicitly separates reasoning into planning and execution, and performs search over the two phases individually. Specifically, we decompose reasoning trajectories and develop reward models for each phase, enabling the search to explore and prune plans and executions separately. We further introduce a dynamic budget allocation mechanism that adaptively redistributes sampling effort based on reward feedback, allowing early stopping on confident steps and reallocation of computation to more challenging parts of the reasoning process. Experiments on both mathematical reasoning and code generation benchmarks demonstrate that our approach consistently improves accuracy while reducing redundant computation.

Large language models are often adapted using parameter-efficient techniques such as Low-Rank Adaptation (LoRA), formulated as $y = W_0x + BAx$, where $W_0$ is the pre-trained parameters and $x$ is the input to the adapted layer. While multi-adapter extensions often employ multiple LoRAs, prior studies suggest that the inner $A$ matrices are highly similar during training and thus suitable for sharing. We revisit this phenomenon and find that this similarity is largely attributable to the identical initialization rather than shared knowledge, with $B$ playing a more critical role in knowledge encoding and transfer. Motivated by these insights, we propose \textbf{ALoRA}, an asymmetric multi-LoRA design with multiple $A$ matrices and a single shared $B$ in multi-task fine-tuning, and \textbf{Fed-ALoRA}, which shares $B$ across clients in federated fine-tuning under both homogeneous and heterogeneous settings, through a novel matrix decomposition strategy to accommodate heterogeneous ranks across clients. Experiments on commonsense reasoning, math reasoning, multi-task NLP dataset, and federated NLP dataset demonstrate that our methods achieve more balanced performance across tasks with comparable or superior average accuracy relative to existing multi-LoRA approaches. Codes are available at https://github.com/OptMN-Lab/ALoRA.