Large Vision-Language Models (LVLMs) have demonstrated strong capabilities in multimodal tasks. Nevertheless, their outputs remain prone to object hallucinations that are inconsistent with visual content. Contrastive decoding, as a training-free inference-time mitigation approach, typically intervenes in the output distribution at every decoding step. However, such indiscriminate full-stage intervention may disrupt originally correct generations and even induce new hallucinations. To address this issue, we first analyze the relationship between internal attention features and hallucinated token segments in LVLMs, revealing that hallucination-related abnormalities already emerge progressively in preceding token segments. Based on this observation, we propose HALMES, a lightweight plug-and-play module that identifies hallucination precursor segments from attention features during generation and selectively triggers contrastive decoding. HALMES can be integrated with various full-stage contrastive decoding methods, transforming indiscriminate intervention into on-demand targeted correction. Extensive experiments across different LVLM architectures and multiple benchmarks show that HALMES further improves the hallucination mitigation performance of existing contrastive decoding methods while reducing the negative impact of full-stage intervention on normal outputs, validating the importance of modeling and selectively intervening on hallucination-prone segments.
Dataset distillation (DD) synthesizes a small training set from a large one, but applying it to 3D medical classification is constrained by the memory and optimization costs of synthesizing full volumetric samples. Under a storage-equivalent budget, distilling in the latent space of a frozen autoencoder reduces this cost and fits more synthetic samples (prototypes) than in voxel space. However, class-level matching, such as distribution matching, causes prototypes to concentrate in dense latent regions while underrepresenting sparser intra-class regions, a consistent imbalance that we term prototype concentration. To address this, we propose MAPA (Mode-Aware Prototype Allocation), a latent distillation method that spreads prototypes across intra-class modes while preserving class-level matching. MAPA marks these modes with fixed anchors estimated by clustering the real latents and distributes prototypes over them through a soft assignment, together with a lightweight anti-collapse regularizer and curriculum weighting that stabilize the optimization. Across five 3D medical classification tasks, MAPA improves the aggregate balanced accuracy by +4.2 points over the strongest latent baseline while using the same number of prototypes. An exploratory severity analysis on MosMedData suggests improved worst-stratum recall without loss of normal-case specificity.
Chest radiograph interpretation errors often arise from failures in visual search: although trainees may correctly identify an abnormality once pointed out, they frequently fail to fixate on the relevant region, while expert search strategies (where radiologists look, in what sequence, and for how long) remain largely implicit. Eye movement modeling examples (EMME), which replay a recorded expert's gaze over the image, have been shown to improve novice detection of findings on chest radiographs, but producing one requires an eye-tracking session on that specific case, so teaching sets are limited to recorded sessions by experts. In this work, RadAI is presented, the first end-to-end framework for transforming predicted radiologist gaze into narrated instructional videos, enabling scalable gaze-guided training without collecting new eye-tracking data. Given only a chest X-ray and its report, RadAI predicts an expert fixation sequence, associates each interval of that sequence with the diagnostic finding it pursues, extracts the corresponding image region, and renders the result as an EMME-style video in which each highlighted finding is narrated with synthesized speech. RadAI is evaluated on 97 cases from the EGD-CXR dataset, assessing whether the highlighted regions correspond to annotated abnormality locations. We position this work as a framework for scalable generation of educational content for residents and medical students in the field of radiology.
Federated Learning (FL), Split Learning (SL), and Split Federated Learning (SplitFed) enable collaborative medical image analysis without sharing raw patient data. However, intermediate features and gradients exchanged during SplitFed training may reveal sensitive information through reconstruction attacks. To address this vulnerability, we propose \textbf{DiffusedSplitFed}, a SplitFed framework that integrates Latent Denoising Diffusion Models (LDDMs) at client--server split points to obfuscate transmitted representations while preserving segmentation performance. Three architectural variants (V1--V3) are developed using different combinations of dual conditioning and global feature fusion to investigate trade-offs among reconstruction resistance, segmentation accuracy, and computational complexity. Experiments on three medical imaging datasets (HAM10K, Blastocysts, and FHPsAOPMSB), supported by multiple independent runs and paired statistical analyses, demonstrate improved resistance to the evaluated image reconstruction attacks compared with Baseline SplitFed, while maintaining competitive segmentation performance. Comparisons with existing diffusion-based segmentation methods further demonstrate competitive segmentation accuracy under the evaluated settings. A conditional convergence analysis is also provided to characterize the effects of LDDM-induced gradient perturbations and local optimization on training. These findings highlight the potential of latent diffusion-based representation obfuscation for reconstruction-resilient collaborative medical image segmentation. The implementation is available at https://github.com/ChamaniS/DiffusedSplitFed.
Handwritten character recognition systems can achieve high accuracy while retaining concentrated errors among visually similar classes. Motivated by residual letter confusions in earlier geometry-based recognition work, we investigate whether these errors can be addressed through the training objective. We introduce Confusion-Adaptive Margin Learning (CAML), which constructs a sparse weighted confusion graph from a discovery partition held out from pilot training and applies pair-specific cosine-separation constraints during final training. CAML adds no classifier, routing rule, or inference-time parameters. We evaluate CAML on EMNIST Letters against a strictly matched cross-entropy CNN baseline, using the same architecture, final training data, deterministic augmentation, optimizer, and checkpoint selection. Eight confusion pairs are targeted: I/L, G/Q, U/V, R/V, D/O, C/E, I/J, and H/N. On the official clean test set, accuracy changes from 95.6346% to 95.6731%, and macro-F1 from 0.956326 to 0.956724. Mean accuracy across ten predefined corrupted conditions increases from approximately 95.3812% to 95.4303%; symmetric errors within targeted pairs decrease from 614 to 605 (1.47%). The global differences are small and not statistically reliable. CAML fixes 124 baseline errors but introduces 116 new errors on clean data (exact McNemar p=0.6515), and no corruption-specific difference remains significant after Holm correction. Thus, CAML does not establish global superiority, but shows that a leakage-controlled, pair-specific margin objective can alter targeted error structure with limited overall benefit for an already strong classifier.
Humans answer complex visual questions by inspecting short moments, updating what they know, and redirecting attention toward missing evidence. Existing video benchmarks, however, often provide the relevant video, evaluate only final answers or isolated localization, or permit textual shortcuts, leaving multi-round visual evidence seeking underexplored. We introduce Video-DCI, a diagnostic benchmark and functionally matched human-agent protocol for Direct Video-Corpus Interaction. It comprises DCI-Loc for temporal evidence localization and DCI-QA for end-to-end discovery, retention, and composition of cross-video evidence. In DCI-QA, videos within each pool are topically related, yet every answer requires hidden temporal evidence from multiple videos. Given only the question and video pool, humans and agents iteratively inspect bounded local windows, retain supporting evidence, and answer with cited temporal spans. Human trajectories provide reference curves for evidence acquisition and use across rounds. Experiments reveal a consistent retrieval-to-grounding gap. On DCI-Loc, frontier VLMs often retrieve and expose the reference moment, yet convert far less of this exposure into precise final citations than the human reference. On DCI-QA, the loss begins earlier and compounds across videos. Ablation studies further confirm that iterative search outperforms one-shot retrieval. Multi-round search is therefore necessary but insufficient: current agents accumulate candidates and observations more readily than precise, jointly sufficient evidence.
Large language models can recognize when they are being evaluated and change their behavior accordingly, undermining the reliability of safety benchmarks. Evaluation awareness (EA) is usually measured by having an LLM judge scan reasoning traces for explicit statements of recognition, but this signal is increasingly unreliable. Recent frontier-model system cards report a drop in verbalized rates between releases, which is indistinguishable from awareness that has gone unverbalized. Interpretability methods could detect such unverbalized evaluation awareness (UEA) in individual interactions, but lack instance-level ground truth. Asking the model alters the input, and dataset labels do not reveal whether it recognized a particular prompt as an evaluation. We address this gap with a model organism. Using synthetic document fine-tuning, SFT, and DPO, we train Qwen3.5-27B to emit a marker (a designated token) in its final response when it recognizes an evaluation, while suppressing the marker and verbalized awareness in its reasoning. The marker largely tracks the evaluation concept as it can be strongly modulated by evaluation framing in the system prompt. When not told whether it is being evaluated, the organism recognizes evaluations from test-question cues, mirroring what its base model often verbalizes. Using the marker as a label, we benchmark linear probes, the Jacobian Lens, and Activation Oracles, built on the base model and blind to the marker. None recovers the label reliably. These results suggest that current interpretability methods are not yet reliable UEA detectors, even when recognition relies on simple cues. Our model organism offers a testbed for improving them.
Clinical agents must retrieve patient-specific evidence and correctly update records despite unreliable tool responses. We introduce MedToolFault, a fault-injection benchmark linking six tool-failure classes to answer correctness, required state changes and forbidden mutations in public clinical-agent tasks. We also introduce MedGuard, a tool-layer method that verifies read evidence and attributes ambiguous writes through public interfaces. Across 127,861 formal episode evaluations, we assess fault recovery, mechanism controls, task expansion, constrained tool budgets and external question answering. Across nine model configurations, unguarded agents fail in 32.2-67.8% of fault-injected evaluations on tasks each configuration solves without faults. MedGuard improves paired safe success by 13.7-46.6 percentage points on these tasks.
Mental health conditions affect millions of people worldwide, yet many cases go undetected until they reach a critical stage. Social media platforms like Reddit have become spaces where people openly share their experiences, but the support is often provided by individuals without clinical training. Existing multi-task approaches jointly model emotion and sentiment for mental health classification but frequently rely on noisy labels generated by pre-trained models and overlook topic information as a useful signal. This paper introduces two key contributions: First, an existing Core Fusion Network framework is extended by adding Topic alongside existing Mental Health Condition, Emotion, and Sentiment Classification tasks, with topic labels derived from the therapist-client interactions in the CounselChat dataset. Second, an Active Learning-based Stepwise Relabeling approach is proposed to refine a small subset of noisy emotion and topic labels using pre-trained models in two stages: targeting medium-uncertainty samples first, followed by high-uncertainty samples. The predictions above a confidence threshold (0.7) are accepted to prevent error propagation. The experiments are conducted on the SuicideWatch and Mental Health dataset of 54,412 Reddit posts collected from five mental health subreddits, evaluating domain-specific transformer encoders including MentalBERT and MentalRoBERTa. The proposed Active Learning-based Stepwise Relabeling, when applied to Core Fusion Network with MentalRoBERTa, improves Mental Health Condition by 0.09%, Emotion by 0.51%, Sentiment by 0.35%, and Topic Classification by 1.57% over the baseline Core Fusion Network. The results suggest that the proposed framework significantly reduces label noise in multi-task mental health classification, and that topic information provides additional context.
Multimodal Large Language Models (MLLMs) can memorize and reproduce sensitive information from their training data, posing significant privacy risks. Their ability to encode knowledge across both textual and visual modalities further complicates the selective removal of such private information. Many existing methods guide model unlearning by training the model toward constructed labels or output distributions, which we refer to as supervision targets. However, these methods often suffer from a trade-off between forgetting effectiveness and model utility. We identify supervision target limitations as a key factor in this trade-off and empirically show that targets containing more private information lead to insufficient forgetting, while larger distribution shifts from the original output distribution are associated with greater utility degradation. Motivated by these findings, we propose FFN-Perturbed Self-Distillation (FPD), which starts from the teacher's original output distribution as a zero-shift reference and selectively perturbs FFN activations associated with private information across textual and visual modalities. By precisely reducing private information from the teacher's original output distribution, FPD constructs supervision targets with less private information and reduced distribution shift. Extensive experiments on S-MLLMUn Bench demonstrate that FPD achieves a better forgetting-utility trade-off than existing methods.
LLMs can now draft complete multi-episode dramas, and LLM judges are increasingly used to evaluate them. Assessing the reliability of these judges is difficult because drama quality encompasses specification-backed errors, reproducible preferences, and legitimate disagreement among human raters. Most existing evaluations reduce these cases to a single score from one judge call and leave tie calibration, call-to-call variation, and the vote margin underlying the human reference unreported. We introduce DramaGrow, a regime-aware benchmark and audit protocol for AI-generated Chinese five-episode drama. The benchmark assigns verifiable constraint violations, four-of-five preferences, and within-panel disagreements to separate denominators and evidence rules. It contains 800 episodes, 120 byte-preserving interventions, and 1,155 blinded five-rater records, including a frozen extension with 114 supermajority pairs and 21 disagreement pairs. On the frozen sets of 114 preference pairs and 35 constraint interventions, the repeated absolute-score rubrics yield overlapping performance, with mean accuracies of 53.8-55.8% and three-call agreement of 29.8-32.5%, and their accuracy falls from 71.6-71.9% to 34.8-37.1% as the tie band widens. DirectPair-GLM reaches 85.1% preference accuracy, 89.5% constraint F1, and 81.9% Detect+Locate on the same items, and exact subpanel enumeration identifies four-of-five as the smallest stable human vote margin. Evaluation protocol, tie calibration, judge-call variation, and panel margin therefore affect performance as much as rubric design. We will release the specification packages, scripts, interventions, anonymized human votes and rationales, frozen judge runs, and analysis code.
Differentiable model predictive control (MPC) enables structured, constraint-aware policies to be optimized end-to-end. However, learning MPC policies faces two bottlenecks: (1) nonsmooth sensitivities at active-set transitions and (2) expensive policy evaluation, since each requires an optimization solve. We first introduce SmoothedTurboMPC, a GPU-parallel differentiable MPC solver combining sequential quadratic programming (SQP) with a log-barrier-smoothed alternating direction method of multipliers (ADMM) inner solver. Its design enables smooth gradients, exact sensitivities at convergence, and efficient execution on GPUs. Second, to efficiently train MPC policies, we pair our solver with first-order reinforcement learning (FORL) using Short-Horizon Actor-Critic (SHAC), which reuses rollout MPC solutions instead of repeatedly re-evaluating them. Comparisons with state of the art solvers show that the approach achieves up to $4.4\times$ higher end-to-end training throughput. On a vehicle racing task, our method achieves higher returns than MPC PPO with a $46\times$ speedup, training on 3.3 million transitions in only 7 minutes. Finally, we use the framework to train a more reliable MPC racing policy via RL with domain randomization, and validate it on a full-scale race car. We integrate our advances into the open-source TurboMPC solver.
Natural language inference (NLI) is a fundamental task of natural language processing and crucial for evaluating the strengths and weaknesses of AI methods. Research has mostly focused on encoder or encoder-decoder architectures. However, the field of decoder-only Large Language Models (LLMs) remains underexplored. While prior work has mainly investigated complex sampling and prompting techniques, the effectiveness of a minimalistic prompting approach is yet to be examined in decoder-only LLMs. In this work, we examine and provide a comprehensive benchmark of LLMs (Llama, Mistral, Gemma, and Phi 4) in this domain, across the MNLI, QNLI and SciTail datasets. The performance of these models was evaluated using zero-shot and few-shot prompting strategies with simple and direct prompts. The results of the article indicate that LLMs generally perform exceptionally well when they are provided with a simple, straightforward prompt, challenging the notion that complex prompting techniques are always required for NLI tasks. Additionally, we find that the choice of demonstration label in one-shot prompting significantly affects performance for most models, with entailment-primed examples performing best. Our findings also show that the use of a few representative examples, in the case of few-shot, produced performance comparable or superior to prompts with higher example counts, although the efficacy of the prompting techniques is model and dataset-dependent. Moreover, our simple prompting approach consistently matched or outperformed more elaborate methods such as ICS and MCQ-based prompting.
Resource-constrained IoT devices cache learned models for real-time decisions, but frequent updates incur substantial bandwidth, energy, and service overhead. Because decision-relevant model differences need not grow with elapsed time, we define Age of Intelligence (AoInt) as a model-version-relative, client-specific, and action utility distortion, and formulate synchronization as a tradeoff between stale-decision regret and refresh cost. Furthermore, we develop Kullback-Leibler-aware AoInt Synchronization (KL-AoIntSync), which accumulates output-distribution changes using client-specific unlabeled probes, whereas Parameter-aware AoInt Synchronization (Param-AoIntSync) uses a server-broadcast scalar parameter change. Then, we establish pathwise guarantees with fixed client distributions, action sets, and utilities. The numerical results show that calibrated Param-AoIntSync reduces joint synchronization cost by 6.15% relative to an Instant-Parameters policy while performing comparably to KL-AoIntSync.
As large language model (LLM) agents get actively adopted for autonomous research, they must not only reason over existing evidence, but also design informative experiments, allocate resources, analyze sequential observations, and decide when to stop and make conclusions. Yet these capabilities remain poorly measured. We introduce Causal-Discovery-Bench (https://github.com/mshen1019/causal-discovery-bench), a verifiable sandbox with configurable structural equation models and resource constraints to evaluate sequential experimental design and causal discovery without domain-semantic cues. Across 3,802 episodes with 13 models from 5 vendors, we analyze agent behavior along 40 metrics in seven families. Agents' performance exhibits a four-level difficulty hierarchy spanning single-experiment design, sequential exploration of causal relationships, resource-aware experimentation balancing cost and statistical power, and autonomous resolution of conflicting evidence. We identify a substantial performance gap between causal inference from correctly curated evidence and navigating causal discovery through autonomous experimentation: Analyzing agents' action trajectories attributes 94.5% of their mistakes to experimental design rather than to inference, which is further consolidated with ablation studies. Relaxing resource constraints alone does not reliably resolve failures in sequential discovery. In contrast, basic statistical tools substantially improve performance across multiple model families, while advanced tools provide little additional benefit when models lack prerequisite lower-level capabilities.
Large Language Models (LLMs) are emerging as natural-language interfaces for Building Information Modeling (BIM), successfully achieving information retrieval, but no benchmark exists to evaluate their ability to edit Industry Foundation Classes (IFC). We introduce BIBIMBAP, a Text-to-BIM benchmark with 100 curated tasks covering Create-Read-Update-Delete (CRUD) operations across spatial, geometric, topological, numeric, and conceptual categories. Each task includes a natural-language prompt, an IFC model, expected structured outputs, and test scripts for automated evaluation. Baseline results show limited performance (best: 50.2%) and frequent failures in constraint preservation and spatial reasoning, with some correct outputs achieved despite flawed reasoning.
Ensuring safe and reliable navigation under diverse sources of uncertainty remains a central challenge for autonomous systems. Within distributional reinforcement learning (DistRL), Implicit Quantile Networks (IQN) have attracted interest for navigation under uncertainty because they enable risk adaptation through flexible quantile sampling. Existing approaches, however, typically use a fixed or randomly sampled Conditional Value-at-Risk (CVaR) level during training and adjust the risk level during deployment using hand-tuned hyperparameters or external sensor signals. This scalar-CVaR formulation confines learning to a narrow family of tail-risk preferences, creating a mismatch between training-time preference exposure and deployment-time risk adjustment. We instead represent variations in risk preference as shifts in quantile-weight distributions and expose the agent to diverse preferences through preference-guided exploration. At deployment, we optimize against the worst-case preference distribution to derive a conservative policy that accounts for multiple risk profiles. This reduces hyperparameter tuning and reliance on limited sensor signals under partial observability. Experiments across simulated environments and vehicle platforms show that our method reduces collisions while improving task success.
Electroencephalography (EEG) data plays an important role in understanding human brain activity, with various major applications. While current approaches, which largely build on EEG foundation models, have demonstrated strong performance across datasets and tasks, they typically rely on a fixed temporal scale. This design choice limits their ability to capture characteristics of neural dynamic behavior, often spanning across multiple timescales, which are naturally captured in EEG recordings. To address this limitation, we propose MAGE (\textbf{Multi-timescale Adaptive Generalization for EEG}), a foundation model that dynamically adapts to capture and integrate information across different timescales. Specifically, MAGE combines two key mechanisms: (i) a novel multi-timescale attention mechanism, capable of learning representations across multiple temporal scales and (ii) a trainable aggregation module, that dynamically identifies and adapts to the most informative timescale of the input signal. Experimental results across seven datasets demonstrate that MAGE consistently outperforms existing EEG foundation models while using substantially fewer parameters. Moreover, our results support the hypothesis that different downstream tasks benefit from distinct temporal scales. All source code will be released.
Both humans and AI exhibit systemic biases, yet little is known about how interface design can foster fairness in hybrid human-AI decision-making. This paper introduces a method for extracting fair data samples from biased datasets and adapts a fairness-explanation technique to highlight potential AI bias during decisions. We conducted a within-subjects study (n=48) where non-experts made decisions across three conditions: independent human judgment, unexplained AI assistance, and AI assistance augmented with visual ethnic-bias explanations (Fair-XAI). To isolate effects, the AI conditions provided systematically balanced recommendations. Results show that fairness explanations significantly reduced both sex and ethnic bias compared to human-only decisions. Fair-XAI also marginally improved fairness over balanced AI recommendations while enhancing task performance. These findings demonstrate that fairness-centered explanation interfaces can meaningfully shape user behavior, providing concrete design insights for developing equitable and accountable human-AI systems.
Does prolonged interaction with Large Language Models reduce users' observable autonomy during conversation? We test this degradation hypothesis by scoring user autonomy across 2,877 multi-turn conversations (46,952 turns) from WildChat and validating on 500 conversations from LMSYS-Chat-1M. Our core finding is a null result: population-level autonomy remains stable throughout conversations, only 4.7% of conversations show meaningful decline, and an apparent decrease at later turn positions is explained by survivorship bias. Instead, the dominant pattern is not erosion but absence: 80.8% of conversations exhibit zero or low autonomy throughout. When autonomy engagement occurs, it is self-reinforcing (persistence doubles from 40% to 85% over consecutive turns). These findings reframe the problem: the risk is not that LLMs progressively degrade users' observable autonomy, but that most users never engage in independent reasoning during the interaction.