Searcharxiv⌕ Search

arXiv subjects

Eunho Yang

Publications and source records attributed to Eunho Yang.

At least 37 records · Page 2Linked to original sources

Efficient Reinforcement for Visual-Textual Thinking with Discrete Diffusion Model

RL-based post-training has been widely adopted to enable interleaved visual and textual reasoning in unified multimodal models capable of both text and image generation. However, most existing approaches are built upon autoregressive (AR) unified models, which require full image regeneration during visual reasoning. In this work, we demonstrate that multimodal discrete diffusion models are effective alternatives to AR models for reinforcement learning in interleaved reasoning, owing to their ability to perform efficient visual rollouts via localized visual editing rather than full image-token regeneration. This reduces rollout computation during GRPO by 26.9\% compared to AR baselines, with minimal performance drop. Despite the improved efficiency, we find that joint reward assignment, which employs a shared reward signal across modalities, introduces cross-modal interference between unrelated image and text token sequences during RL updates. To address this issue, we propose factorized reward assignment, a strategy that assigns rewards independently to text and vision segments. With factorized reward assignment, our RL approach achieves an 11.2% improvement over joint reward assignment and a 38.04% improvement over the base model.

cs.CV↗

On the Collapse of Generative Paths: A Criterion and Correction for Diffusion Steering

Inference-time steering adapts pretrained diffusion and flow models to new tasks without retraining, often utilizing ratio-of-densities constructions that reweight time-indexed marginals with fixed exponents. We identify Marginal Path Collapse, a failure mode in which the intermediate density defined by such compositions becomes non-normalizable despite valid endpoints. This collapse can arise when composing heterogeneous experts trained with mismatched noise schedules (and/or negative exponents / partial supports). To address this, we provide (i) a sharp sufficient Path Existence Criterion that characterizes when the composed intermediate densities are mathematically well-defined, and (ii) Adaptive Path Correction with Exponents (ACE), which generalizes Feynman-Kac steering to support time-varying exponents. Our analysis reveals that ACE controls the quantile radius of the intermediate distributions, providing a theoretical mechanism for path stabilization observed in experiments. On flexible-pose scaffold decoration, a drug design task composed of de-novo, conformer, and protein-conditioned experts, ACE prevents collapse and significantly outperforms constant-exponent baselines. Furthermore, ACE improves attribute success rates in compositional image generation, establishing it as a general framework for compositional sampling. Project Page: https://ziseoklee.github.io/projects/ACE/

cs.AI↗

Verifying Meta-Awareness via Predictive Rewards in Reasoning Models

Recent research on reasoning models explores the meta-awareness of language models, including their ability to determine optimal thinking duration, recognize knowledge boundaries, and structure concept-level thinking. While current large reasoning models depend solely on answer-based verification, we show that adding meta-awareness objectives leads to significant performance gains over models without such meta-knowledge. MAPR (Meta-Awareness via Predictive Reward) utilizes a self-generated task of predicting rollout statistics - specifically length, pass-rate, and concepts used - allowing for verification against the actual statistics. Furthermore, by leveraging this self-predictive capability, the model can regulate its reasoning behavior by i) filtering out trivial or unsolvable prompts, ii) reducing lengthy generations that tend to be incorrect, and iii) generating hints relevant to the problem. The results are inspiring: MAPR yields significant improvements in both accuracy and training efficiency on various reasoning benchmarks. More specifically, our method can speed up GRPO training by over 1.28x to reach the same performance, and achieve 83.18% gain in accuracy on AIME25, and a 13.04% average gain over six mathematics benchmarks. The code is publicly available at https://github.com/akatigre/MAPR-RL.

cs.LG↗

LFQ: Logit-aware Final-block Quantization for Boosting the Generation Quality of Low-Bit Quantized LLMs

As large language models continue to scale, low-bit weight-only post-training quantization (PTQ) offers a practical solution to their memory-efficient deployment. Although block-wise PTQ is capable of matching the full-precision (FP) baseline on basic language modeling and understanding, its quality is degraded for generative tasks -- especially at longer responses and extended chains of thought, which is critical in boosting task accuracy. We attribute this shortfall to two factors: (i) the omission of the unembedding layer (the LM head) in block-wise optimization and (ii) the reliance on the mean squared error (MSE) objective. Both factors cause the token probability distribution of the quantized model to misalign with that of the FP model, yielding notable accuracy drops on text generation benchmarks. To rectify the discrepancy, we introduce Logit-aware Final-block Quantization (LFQ), a simple yet effective enhancement to block-wise PTQ that quantizes the final Transformer block by minimizing the cross-entropy between the logits of the FP model and those of its quantized counterpart. By aligning token probabilities at the logit level in the final block, LFQ consistently improves the accuracy of complex generation tasks over state-of-the-art block-wise PTQ across diverse model families, while maintaining parity with FP baselines on language modeling and understanding.

cs.AI↗

Discounted Beta-Bernoulli Reward Estimation for Sample-Efficient Reinforcement Learning with Verifiable Rewards

Reinforcement learning with verifiable rewards (RLVR) has emerged as an effective post-training paradigm for improving the reasoning capabilities of large language models. However, existing group-based RLVR methods often suffer from severe sample inefficiency. This inefficiency stems from reliance on point estimation of rewards from a small number of rollouts, leading to high estimation variance, variance collapse, and ineffective utilization of generated responses. In this work, we reformulate RLVR from a statistical estimation perspective by modeling rewards as samples drawn from a policy-induced distribution and casting advantage computation as the problem of estimating the reward distribution from finite data. Building on this view, we propose Discounted Beta-Bernoulli (DBB) reward estimation, which leverages historical reward statistics for the non-stationary distribution. Although biased, the resulting estimator exhibits reduced and stable variance, theoretically avoids estimated variance collapse, and achieves lower mean squared error than standard point estimation. Extensive experiments across six in-distribution and three out-of-distribution reasoning benchmarks demonstrate that GRPO with DBB consistently outperforms naive GRPO, achieving average Acc@8 improvements of 3.22/2.42 points in-distribution and 12.49/6.92 points out-of-distribution on the 1.7B and 8B models, respectively, without additional computational cost or memory usage.

cs.LG↗

Argument Reconstruction as Supervision for Critical Thinking in LLMs

To think critically about arguments, human learners are trained to identify, reconstruct, and evaluate arguments. Argument reconstruction is especially important because it makes an argument's underlying inferences explicit. However, it remains unclear whether LLMs can similarly enhance their critical thinking ability by learning to reconstruct arguments. To address this question, we introduce a holistic framework with three contributions. We (1) propose an engine that automatically reconstructs arbitrary arguments (GAAR), (2) synthesize a new high-quality argument reconstruction dataset (Arguinas) using the GAAR engine, and (3) investigate whether learning argument reconstruction benefits downstream critical thinking tasks. Our experimental results show that, across seven critical thinking tasks, models trained to learn argument reconstruction outperform models that do not, with the largest performance gains observed when training on the proposed Arguinas dataset.

cs.CL↗

Evaluating the Pre-Consultation Ability of LLMs using Diagnostic Guidelines

We introduce EPAG, a benchmark dataset and framework designed for Evaluating the Pre-consultation Ability of LLMs using diagnostic Guidelines. LLMs are evaluated directly through HPI-diagnostic guideline comparison and indirectly through disease diagnosis. In our experiments, we observe that small open-source models fine-tuned with a well-curated, task-specific dataset can outperform frontier LLMs in pre-consultation. Additionally, we find that increased amount of HPI (History of Present Illness) does not necessarily lead to improved diagnostic performance. Further experiments reveal that the language of pre-consultation influences the characteristics of the dialogue. By open-sourcing our dataset and evaluation pipeline on https://github.com/seemdog/EPAG, we aim to contribute to the evaluation and further development of LLM applications in real-world clinical settings.

cs.CL↗

CollabVR: Collaborative Video Reasoning with Vision-Language and Video Generation Models

Recent "Thinking with Video" approaches use Video Generation Models (VGMs) for visual reasoning by producing temporally coherent Chain-of-Frames as reasoning artifacts. Even strong VGMs, however, exhibit two recurring failure modes on goal-directed tasks: long-horizon drift on multi-step tasks and mid-clip simulation errors that compound. Both stem from the absence of explicit reasoning built upon the VGM's short-horizon visual prior, a role naturally filled by Vision-Language Models (VLMs), but where to place the VLM is non-trivial: upfront plans commit before any frame is generated and post-hoc critiques over whole videos intervene too late. We propose VLM-VGM Collaborative Video Reasoning (CollabVR), a closed-loop framework that couples the VLM with the VGM at step-level granularity: the VLM plans the immediate next action, inspects the clip the VGM generates, and folds the verifier's diagnosis directly into the next action prompt to repair detected failures. On Gen-ViRe and VBVR-Bench, CollabVR improves both open-source and closed-source VGMs over single-inference, Pass@$k$, and prior test-time scaling baselines at matched compute, with the largest gains on the hardest tasks. It also yields further improvements on top of a reasoning-fine-tuned VGM, indicating that step-level VLM supervision is orthogonal to and stackable with reasoning-oriented fine-tuning. We provide video samples and additional qualitative results at our project page: https://joow0n-kim.github.io/collabvr-project-page.

cs.CV↗

Delta-XAI: A Unified Framework for Explaining Prediction Changes in Online Time Series Monitoring

Explaining online time series monitoring models is crucial across sensitive domains such as healthcare and finance, where temporal and contextual prediction dynamics underpin critical decisions. While recent XAI methods have improved the explainability of time series models, they mostly analyze each time step independently, overlooking temporal dependencies. This results in further challenges: explaining prediction changes is non-trivial, methods fail to leverage online dynamics, and evaluation remains difficult. To address these challenges, we propose Delta-XAI, which adapts 14 existing XAI methods through a wrapper function and introduces a principled evaluation suite for the online setting, assessing diverse aspects, such as faithfulness, sufficiency, and coherence. Experiments reveal that classical gradient-based methods, such as Integrated Gradients (IG), can outperform recent approaches when adapted for temporal analysis. Building on this, we propose Shifted Window Integrated Gradients (SWING), which incorporates past observations in the integration path to systematically capture temporal dependencies and mitigate out-of-distribution effects. Extensive experiments consistently demonstrate the effectiveness of SWING across diverse settings with respect to diverse metrics. Our code is publicly available at https://github.com/AITRICS/Delta-XAI.

cs.LG↗

Accelerating Vision Transformers with Adaptive Patch Sizes

Vision Transformers (ViTs) partition input images into uniformly sized patches regardless of their content, resulting in long input sequence lengths for high-resolution images. We present Adaptive Patch Transformers (APT), which addresses this by using multiple different patch sizes within the same image. APT reduces the total number of input tokens by allocating larger patch sizes in more homogeneous areas and smaller patches in more complex ones. APT achieves a drastic speedup in ViT inference and training, increasing throughput by 40% on ViT-L and 50% on ViT-H while maintaining downstream performance, and can be applied to a previously fine-tuned ViT, converging in as little as 1 epoch. It also significantly reduces training and inference time without loss of performance in high-resolution dense visual tasks, achieving up to 30\% faster training and inference in visual QA, object detection, and semantic segmentation.

cs.CV↗

CURaTE: Continual Unlearning in Real Time with Ensured Preservation of LLM Knowledge

The inability to filter out in advance all potentially problematic data from the pre-training of large language models has given rise to the need for methods for unlearning specific pieces of knowledge after training. Existing techniques overlook the need for continuous and immediate action, causing them to suffer from degraded utility as updates accumulate and protracted exposure of sensitive information. To address these issues, we propose Continual Unlearning in Real Time with Ensured Preservation of LLM Knowledge (CURaTE). Our method begins by training a sentence embedding model on a dataset designed to enable the formation of sharp decision boundaries for determining whether a given input prompt corresponds to any stored forget requests. The similarity of a given input to the forget requests is then used to determine whether to answer or return a refusal response. We show that even with such a simple approach, not only does CURaTE achieve more effective forgetting than existing methods, but by avoiding modification of the language model parameters, it also maintains near perfect knowledge preservation over any number of updates and is the only method capable of continual unlearning in real-time.

cs.CL↗

Integrating Multimodal Large Language Model Knowledge into Amodal Completion

With the widespread adoption of autonomous vehicles and robotics, amodal completion, which reconstructs the occluded parts of people and objects in an image, has become increasingly crucial. Just as humans infer hidden regions based on prior experience and common sense, this task inherently requires physical knowledge about real-world entities. However, existing approaches either depend solely on the image generation ability of visual generative models, which lack such knowledge, or leverage it only during the segmentation stage, preventing it from explicitly guiding the completion process. To address this, we propose AmodalCG, a novel framework that harnesses the real-world knowledge of Multimodal Large Language Models (MLLMs) to guide amodal completion. Our framework first assesses the extent of occlusion to selectively invoke MLLM guidance only when the target object is heavily occluded. If guidance is required, the framework further incorporates MLLMs to reason about both the (1) extent and (2) content of the missing regions. Finally, a visual generative model integrates these guidance and iteratively refines imperfect completions that may arise from inaccurate MLLM guidance. Experimental results on various real-world images show impressive improvements compared to all existing works, suggesting MLLMs as a promising direction for addressing challenging amodal completion.

cs.CV↗

Instruction-Guided Lesion Segmentation for Chest X-rays with Automatically Generated Large-Scale Dataset

The applicability of current lesion segmentation models for chest X-rays (CXRs) has been limited both by a small number of target labels and the reliance on complex, expert-level text inputs, creating a barrier to practical use. To address these limitations, we introduce instruction-guided lesion segmentation (ILS), a medical-domain adaptation of referring image segmentation (RIS) designed to segment diverse lesion types based on simple, user-friendly instructions. Under this task, we construct MIMIC-ILS, the first large-scale instruction-answer dataset for CXR lesion segmentation, using our fully automated multimodal pipeline that generates annotations from CXR images and their corresponding reports. MIMIC-ILS contains 1.1M instruction-answer pairs derived from 192K images and 91K unique segmentation masks, covering seven major lesion types. To empirically demonstrate its utility, we present ROSALIA, a LISA model fine-tuned on the MIMIC-ILS dataset. ROSALIA can segment diverse lesions and provide textual explanations in response to user instructions. The model achieves high accuracy in our newly proposed task, highlighting the effectiveness of our pipeline and the value of MIMIC-ILS as a foundational resource for pixel-level CXR lesion grounding. The dataset and model are available at https://github.com/checkoneee/ROSALIA.

cs.CV↗

ReviewScore: Misinformed Peer Review Detection with Large Language Models

Peer review serves as a backbone of academic research, but in most AI conferences, the review quality is degrading as the number of submissions explodes. To reliably detect low-quality reviews, we define misinformed review points as either "weaknesses" in a review that contain incorrect premises, or "questions" in a review that can be already answered by the paper. We verify that 15.2% of weaknesses and 26.4% of questions are misinformed and introduce ReviewScore indicating if a review point is misinformed. To evaluate the factuality of each premise of weaknesses, we propose an automated engine that reconstructs every explicit and implicit premise from a weakness. We build a human expert-annotated ReviewScore dataset to check the ability of LLMs to automate ReviewScore evaluation. Then, we measure human-model agreements on ReviewScore using eight current state-of-the-art LLMs. The models show F1 scores of 0.4--0.5 and kappa scores of 0.3--0.4, indicating moderate agreement but also suggesting that fully automating the evaluation remains challenging. A thorough disagreement analysis reveals that most errors are due to models' incorrect reasoning. We also prove that evaluating premise-level factuality shows significantly higher agreements than evaluating weakness-level factuality.

cs.CL↗

When to Ensemble: Identifying Token-Level Points for Stable and Fast LLM Ensembling

Ensembling Large Language Models (LLMs) has gained attention as a promising approach to surpass the performance of individual models by leveraging their complementary strengths. In particular, aggregating models' next-token probability distributions to select the next token has been shown to be effective in various tasks. However, while successful for short-form answers, its application to long-form generation remains underexplored. In this paper, we show that using existing ensemble methods in long-form generation requires a careful choice of ensembling positions, since the standard practice of ensembling at every token often degrades performance. We identify two key factors for determining the ensembling positions: tokenization mismatch across models and consensus in their next-token probability distributions. Based on this, we propose SAFE, (Stable And Fast LLM Ensembling), a framework that selectively ensembles by jointly considering these factors. To further improve stability, we apply a probability sharpening strategy when the ensemble distribution becomes overly smooth, enabling the selection of more confident tokens during ensembling. Our experiments on diverse benchmarks, including MATH500 and BBH, demonstrate that SAFE outperforms existing methods in both accuracy and efficiency, with gains achieved even when ensembling fewer than 1% of tokens.

cs.CL↗

Structure-Aware Set Transformers: Temporal and Variable-Type Attention Biases for Asynchronous Clinical Time Series

Electronic health records (EHR) are irregular, asynchronous multivariate time series. As time-series foundation models increasingly tokenize events rather than discretizing time, the input layout becomes a key design choice. Grids expose time$\times$variable structure but require imputation or missingness masks, risking error or sampling-policy shortcuts. Point-set tokenization avoids discretization but loses within-variable trajectories and time-local cross-variable context (Fig.1). We restore these priors in STructure-AwaRe (STAR) Set Transformer by adding parameter-efficient soft attention biases: a temporal locality penalty $-|Δt|/τ$ with learnable timescales and a variable-type affinity $B_{s_i,s_j}$ from a learned feature-compatibility matrix. We benchmark 10 depth-wise fusion schedules (Fig.2). On three ICU prediction tasks, STAR-Set achieves AUC/APR of 0.7158/0.0026 (CPR), 0.9164/0.2033 (mortality), and 0.8373/0.1258 (vasopressor use), outperforming regular-grid, event-time grid, and prior set baselines. Learned $τ$ and $B$ provide interpretable summaries of temporal context and variable interactions, offering a practical plug-in for context-informed time-series models.

cs.LG↗

No Prompt Left Behind: Exploiting Zero-Variance Prompts in LLM Reinforcement Learning via Entropy-Guided Advantage Shaping

Reinforcement Learning with Verifiable Rewards (RLVR) is a powerful framework for improving the reasoning abilities of Large Language Models (LLMs). However, current methods such as GRPO rely only on problems where the model responses to the same input differ in correctness, while ignoring those where all responses receive the same reward -- so-called zero-variance prompts. In this work, we argue that such prompts are not useless but can, in fact, provide meaningful feedback for policy optimization. To this end, we introduce Reinforcement Learning with Zero-Variance Prompts (RL-ZVP), a novel algorithm that extract learning signals from zero-variance prompts. RL-ZVP directly rewards correctness and penalizes errors even without contrasting responses, modulating feedback with token-level characteristics to preserve informative, nuanced signals. Across six math reasoning benchmarks, RL-ZVP achieves significant improvements of up to 8.61 points in accuracy and 7.77 points in pass rate over GRPO, while consistently outperforming other baselines that filter out zero-variance prompts. These results highlight the untapped potential of learning from zero-variance prompts in RLVR. The project page is available at https://bltnynk.github.io/publications/rl-zvp/.

cs.CL↗

Towards Reliable Test-Time Adaptation: Style Invariance as a Correctness Likelihood

Test-time adaptation (TTA) enables efficient adaptation of deployed models, yet it often leads to poorly calibrated predictive uncertainty - a critical issue in high-stakes domains such as autonomous driving, finance, and healthcare. Existing calibration methods typically assume fixed models or static distributions, resulting in degraded performance under real-world, dynamic test conditions. To address these challenges, we introduce Style Invariance as a Correctness Likelihood (SICL), a framework that leverages style-invariance for robust uncertainty estimation. SICL estimates instance-wise correctness likelihood by measuring prediction consistency across style-altered variants, requiring only the model's forward pass. This makes it a plug-and-play, backpropagation-free calibration module compatible with any TTA method. Comprehensive evaluations across four baselines, five TTA methods, and two realistic scenarios with three model architecture demonstrate that SICL reduces calibration error by an average of 13 percentage points compared to conventional calibration approaches.

cs.LG↗