Skip to main content
Search

Southeast University

4 related articles

Subscribe to updates
  1. arXiv

    OG-OPSD selects KL direction by outcome correctness and truncates prefixes by entropy, lifting Qwen3 1.7B/4B/8B math averages to 44.4/64.9/66.4

    Analyzing on-policy self-distillation (OPSD) through the advantage formulation of RLVR, this work finds that a fixed divergence objective under-penalizes and over-rewards incorrect trajectories and that teacher supervision reliability varies with trajectory outcome and cumulative average teacher entropy; it proposes OG-OPSD, which applies forward KL across correct trajectories and reverse KL only within a prefix of incorrect trajectories whose cutoff is set by the last crossing of the cumulative average teacher entropy curves, consistently outperforming vanilla OPSD and multiple strong baselines on mathematical reasoning, multimodal reasoning, and out-of-distribution tasks across Qwen3-1.7B/4B/8B and Qwen3-VL-2B.

  2. arXiv

    CIM Decomposes MLLM Uncertainty via Causal-Invariant Masking, with EED Proxy Matching Performance at Nearly 50% Speedup

    Addressing hallucinations in Multimodal Large Language Models (MLLMs), this work proposes Causal-Invariant Masking (CIM), which measures the semantic shift between original predictions and those conditioned on a causally-focused view to decompose uncertainty types, introduces Semantic Divergence as the core UQ metric with theoretical evidence that it converges to the variance of the model's sensitivity to non-causal correlations, and further proposes Expected Embedding Drift (EED), a fast geometric proxy metric; experiments report state-of-the-art performance on various benchmarks, with EED accelerating by nearly 50% at comparable performance.

  3. arXiv

    APEX chains skills to carry a false user-approval record across handoffs, inducing attacker-selected actions in 74.2% of 690 SkillsBench attempts

    The work introduces APEX, which builds and refines adversarial skill chains tailored to a user task and an attacker-selected action: an upstream skill induces the agent to write a record of genuine task progress, and a downstream skill uses that record to carry a false claim of user approval, inducing the selected action in 512 of 690 attempts (74.2%) across four targeted-action families and six models on SkillsBench, with 84.3% success on GPT-5.4 versus 17.4% when the workflow is merged into one skill; a prompting defense lowers GPT-5.4 targeted-action success from 84.3% to 59.1% while the verifier test-pass rate across 72 benign native-skill tasks falls from 86.7% to 56.3%.

  4. medRxiv

    Ischemic Stroke Detection, Segmentation, and Volume Estimation from Multi-sequence MRI with Missing Sequences

    This work presents ISDS-MRI, a unified framework that uses graph neural networks and sequence-specific feature modeling to detect ischemic stroke, segment lesions, and estimate lesion volume from multi-sequence MRI while accommodating incomplete sequence combinations; evaluated across multiple public MRI datasets and a newly curated BGD-MRIS dataset of 532 scans from three hospitals in Bangladesh, it achieves a Dice score of 0.725, an AUC of 0.962, and a lesion volume estimation relative error of 8.4%, outperforming comparison methods by 3.2% in Dice, 2.6% in detection, and reducing volume relative error by 1.9%.

You’re all caught up