Skip to main content
Search

The University of Melbourne

9 related articles

Subscribe to updates
  1. arXiv

    Multimodal reasoning models can forget one face without forgetting the person: VKCR confines unlearning to the image channel

    Multimodal large reasoning models can pull sensitive information back through cross-modal associations inside intermediate reasoning, so suppressing the final answer is not enough. VKCR reframes the forget target as one designated image-to-knowledge association: it localizes the edit in a compact association-conditioned low-rank parameter subspace, then counterfactually rebinds the affected reasoning branch, weakening target-bearing or unsupported continuations while favoring evidence-grounded non-sensitive alternatives. On CLEAR-R and MLLMU-R it reduces both image-conditioned target leakage and hallucinated reassociation while preserving non-sensitive visual reasoning and permitted text-only access to the same knowledge.

  2. arXiv

    EvoGuard screens every online skill rewrite: 56.5% average success on ALFWorld versus 45.5% for accepting all updates

    Language agents can store experience as natural-language skill documents, but rewriting a skill from a single trajectory after every deployed task lets its quality drift. EvoGuard turns each update into a two-way choice: offline, skills are scored on a private test set, rejection sampling keeps pairs with a clear utility gap, and a selector trained on the logits of two label tokens decides which skill to keep; at deployment, one forward pass settles which of the old and new skills stays. On ALFWorld, with Qwen3-4B and Qwen3-8B executors, it averages 56.5% success, against 45.5% for accepting every rewrite and 36.7% with no skill.

  3. arXiv

    Last-layer weights alone can rank model quality: a statistical account of heavy-tailed spectra, with an overfitting marker during training

    A Bayes free energy computed from trained final-layer weights alone tracks model quality across convolutional networks and LLMs, and under a heavy-tailed spectral model it yields an optimal heaviness whose crossing marks the onset of overfitting.

  4. arXiv

    PTH checks show four harness details can reverse the ranking of stale-data RL methods, and after correction TIS matches SAN on verl

    The work introduces PTH (Probe The Harness), a set of checks that makes the experimental harness visible, and uses a comparison between SAN, a behaviour-free method, and truncated importance sampling (TIS) on verl and in a single-GPU trainer to show that four harness details—the PPO ratio taken against the learner's own recomputed probabilities, the data seed not reaching the TIS arm, the replay queue reusing its first batch for 33 updates, and two loss normalisers differing from their description—can reverse the observed ranking; with the harness checked, TIS matches SAN on verl, and in the trainer TIS learns steadily while SAN keeps a margin.

  5. arXiv

    BPO replaces single-response rewards with counterfactual-view behavior packs, lifting Qwen2.5-VL-7B temporal-hard accuracy by 7.8 points

    The work proposes Behavior Pack Optimization (BPO), which replaces the single-response reward in video MLLM post-training with a jointly scored behavior pack of outputs across counterfactual views chosen by question type, requiring stability under irrelevant interventions, sensitivity when key evidence is removed, and abstention when no evidence remains, using the original-view response as an anchor-relative advantage; on TempCompass, MVBench, and NExT-QA, BPO improves Qwen2.5-VL-7B-Instruct macro accuracy by 4.7 pp, the temporal-hard subset by 7.8 pp, and abstention F1 by 20.0 pp over a budget-matched vanilla GRPO baseline from the same SFT checkpoint, transferring to Video-MME, LongVideoBench, and LLaVA-Video-7B.

  6. After reviewing 922 citations, 41 international experts extended the acute exacerbation definition from IPF to all fibrotic interstitial lung diseases and proposed a new acute respiratory worsening framework

    An international multidisciplinary working group of 41 experts systematically reviewed 922 citations up to March 17, 2025, revised the 2016 acute exacerbation (AE) definition—previously limited to idiopathic pulmonary fibrosis (IPF)—into an AE-fILD definition applicable across all fibrotic interstitial lung diseases (anchored in diffuse alveolar damage with or without organizing pneumonia), proposed the broader concept of acute respiratory worsening (ARW) to capture acute deteriorations not attributable to DAD, and summarized evidence on epidemiology, risk factors, prognosis, and management while discussing AE as a clinical trial endpoint.

  7. BJPsych Open

    Australian team publishes the 5W-PL protocol: linking 14 Victorian health and human services datasets to trace mental health service pathways for young people aged 12–25

    The protocol describes the design of the 5W-PL study: using the Centre for Victorian Data Linkage's Victorian Linkage Map to link 14 datasets covering mortality, mental health, hospital, emergency, ambulance and human services (child protection, disability, sexual assault, homelessness, alcohol and drug) through deterministic and probabilistic methods, building a cross-sector longitudinal cohort of people born between 1970 and 2010 with service-use records from 2015 to 2025, in order to identify subgroups by care pathway, map geographical patterns of service need, and develop predictive models of service type and intensity for young people aged 12–25.

  8. medRxiv

    Time-resolved predictability of end-of-therapy outcome and relapse after cure in Phase 3 tuberculosis trials

    Using harmonised clinical data from two Phase 3 trials (2,918 participants), this study trained monthly tabular models from baseline to therapy end for time-resolved prediction of end-of-therapy (EOT) outcomes and post-treatment relapse, finding that EOT outcome prediction improved after month 3 (ROC-AUC up to 0.84, driven by sputum-smear and solid culture), whereas relapse prediction among those with favourable EOT outcomes and completed follow-up improved only modestly through month 3 (ROC-AUC 0.58-0.63) before declining, with age, sex, clinical symptoms and bacterial burden contributing most; large language model-derived embedding models matched tabular relapse models throughout therapy and outperformed tabular EOT models at months 3 and 4 (ΔROC-AUC 0.12 and 0.

  9. Climatic Change

    Generative Debunking of Climate Misinformation: Automatically Producing 'Truth Sandwich' Rebuttals with Large Language Models

    This study introduces a 'generative debunking' framework that integrates climate contrarian claim classification (CARDS) and fallacy detection (FLICC) into an LLM prompting pipeline, so that a model takes a climate myth as input and produces a debunking that follows the fact-myth-fallacy-fact ('truth sandwich') structure; the authors pair three prompting strategies of increasing complexity with GPT-4, Palm2, and Mixtral, and four authors (including one climate misinformation expert) rate 60 debunkings for 20 myths on fact, fallacy, and structure, finding that GPT-4 with a simple prompt and Mixtral with a structured prompt perform relatively well, that fallacy explanations score higher than facts, and that non-expert annotators show poor agreement with the expert on fact quality.

You’re all caught up