Skip to main content
Search

University of Birmingham

9 related articles

Subscribe to updates
  1. JAMA Cardiology

    Coronary inflammation read from noncontrast CT: a fat attenuation index score sharpens risk in people with zero calcium

    Coronary inflammation could previously be measured only on contrast-enhanced coronary CT angiography, while routine calcium scoring uses low-dose noncontrast CT. A deep-learning pipeline trained and externally validated in a multicenter cohort computes a pericoronary fat attenuation index (FAI) score from noncontrast CT; the score performed comparably to the contrast-based FAI and further separated cardiovascular death and major adverse cardiovascular events among people with a calcium score of zero.

  2. medRxiv

    IBDex freezes 3,168 gut biopsies into one reference atlas: deep remission quiets inflammation but leaves a disease state behind

    Bulk mucosal RNA sequencing in inflammatory bowel disease has been analyzed one cohort at a time, so historical data accumulate without becoming a reusable reference. IBDex trained a tissue-aware conditional variational autoencoder on 3,168 biopsies from 1,680 patients across four cohorts and suppressed cohort identity adversarially, then froze the gene interface, encoder, decoder, pathway basis and phenotype directions. After freezing, dataset identity fell to balanced accuracy 0.307 against chance 0.250 while inflammation stayed recoverable (AUC 0.82); in 93 patients with concordant endoscopic and histological remission, inflammation approached controls (AUC 0.58) but disease state stayed displaced (AUC 0.725).

  3. arXiv

    Differentiable decision trees stop spreading across dozens of features: a Euclidean penalty cuts Wine's effective node features from 8.94 to 1.54

    Differentiable decision trees keep explicit paths and end-to-end training, but an oblique soft split at each internal node can depend on many input features at once, weakening node-level readability. EU-DDT adds a normalised squared Euclidean regulariser to the training objective that constrains how far each node's relative feature weights depart from uniform, without changing the sigmoid soft routing. Over five matched random seeds, mean effective features fall from 8.94 to 1.54 on Wine and from 45.28 to 9.06 on Digits, while mean test accuracy moves from 100.0% to 93.9% and from 92.3% to 94.6%, a dataset-dependent trade-off between concentration and performance.

  4. Nature Medicine

    A medical vision-language model that scores 0.895 on diagnosis and points back to the lesion

    Medical imaging foundation models have traded generality against pixel-level precision, and most return a label without evidence. EVLF-FM feeds both disease-level and pixel-level visual features into a language model, first with supervised fine-tuning and then with reinforcement learning on verifiable rewards, spanning six specialties and 10 imaging modalities. On internal diagnostic validation it reached the highest mean accuracy among evaluated models, 0.895, and averaged 0.842 Acc@0.5 for medical visual grounding across nine modalities.

  5. arXiv

    3D medical self-supervised pretraining drops random cropping: an offline probe bank lifts AMOS low-label Dice from 55.81 to 58.64

    3D medical self-supervised pretraining has long defaulted to random fixed-size sub-volumes from each scan, implicitly treating all spatial regions as equally worth observing. VolumeProbe replaces that step with an offline probe bank: candidate crop centers are scored for anatomical plausibility, informativeness and diversity, then sampled uniformly from the cache during pretraining. Across the VoCo, S2DC and MedGMAE backbones and segmentation, classification and registration tasks, swapping random cropping for this bank improves transfer consistently, most clearly in low-label segmentation, where VoCo on AMOS rises from 55.81 to 58.64 Dice at 1% labels.

  6. arXiv

    SymCE generates counterexamples with per-theorem symbolic verifiers: counterexample-only SFT collapses true-theorem recognition from 0.27 to 0.00, while sparse-reward RL repairs it to 0.66

    The work frames counterexample generation as constrained witness emission against deterministic per-theorem Python verifiers, releases SymCE with 4,707 false conjectures each paired with an executable verifier, and trains Qwen3-4B with the verifier as reward: counterexample-only SFT collapses true-theorem recognition from 0.27 to 0.00, while RLVR with a sparse outcome-only reward repairs and exceeds the base to 0.66, with the collapse replicating across four seeds and on Gemma-3-4B.

  7. arXiv

    SAKIKO audits tool-call interventions across seven LLMs: behavioral movement is not repair, and only Qwen3-8B earns a formal ADMIT

    The work presents SAKIKO, an auditing framework that decomposes internal activation interventions on pre-execution tool decisions into directional error discovery, router-conditioned intervention, destination-resolved verification, and prospectively frozen statistical licensing; across seven LLMs on When2Call and MetaTool, channel-keyed interventions induce direction-specific net gains in five models, yet destination auditing shows behavioral movement is not repair—an intervention with +55 net gain corrupts over half of the baseline-correct decisions it touches, Qwen3-4B and Gemma-2-9B are formally declined due to finite-sample uncertainty, and only Qwen3-8B passes all ten conditions for an ADMIT.

  8. arXiv

    Tracking fine-tuning trajectories shows safety-aligned LLMs are not steered toward harm but lose their safe subspace, and projecting out the harmful gradient subspace cuts free-generation emergent misalignment by up to 80.0% on Qwen2.5-14B-IT

    The work presents a dynamic second-order geometric study of emergent misalignment (EM): tracking training trajectories shows directional Hessian curvature concentrates on semantic pivot tokens, Grassmannian projections show the harmful-safe gap widens mainly because safe-gradient overlap declines rather than through rotation toward a new malicious circuit, and on this basis it introduces a parameter-level Geometric Mitigation Framework that orthogonally projects the empirical harmful gradient subspace out of LoRA updates, suppressing free-generation EM by up to 80.0% on Qwen2.5-14B-IT and, across four open-weight instruction-based model families (3B–20B), teacher-forced evaluation shows the same harmful subspace controls the conditional support of frozen EM responses.

  9. bioRxiv

    Agentic-AI-ready genome-wide poxvirus-host interaction screen refined by a protein language model

    This work proposes ICARus, a positive-unlabelled read-out refinement framework that integrates protein-protein interaction information derived from a protein language model into a genome-wide RNA interference screen to boost the discovery of vaccinia virus-host interactions and human genes with potential antiviral function, and releases the raw and refined read-outs of the screen as an agentic-AI-enabled community resource.

You’re all caught up