Skip to main content
Search

The University of Sydney

24 related articles

Subscribe to updates
  1. arXiv

    SinkState stores recurrent states as adaptive low-rank factors: hybrid-model decoding peak memory falls to as little as 1/2.6 of baseline

    Hybrid language models compress history into fixed-size recurrent states, yet serving systems still treat those states as dense matrices, reading and rewriting them whole at every decoding step, so memory traffic grows with batch size. SinkState exploits the concentration of queries, keys and values in linear attention into a stable low-dimensional subspace, keeping each state in an adaptive low-rank factorization throughout decoding, merging new tokens with a local SVD and consolidating the frozen part asynchronously on the CPU. Across four Mamba-2 and GDN hybrid models from 4B to 120B parameters, peak recurrent-state memory falls to 0.38–0.63 of the baseline and decoding speeds up by up to 1.58x, with accuracy maintained.

  2. OpenAI News

    Australian universities bring ChatGPT Edu into teaching, with 81% of surveyed students already using AI weekly

    Several Australian universities are collaborating with OpenAI to bring frontier AI tools such as ChatGPT Edu into teaching, clinical simulation and teacher training. A survey commissioned by OpenAI of 1,018 university students found 81% use AI several times a week and 28% have already built an AI agent or assistant. UNSW plans to bring ChatGPT Edu to more than 80,000 students and staff, while the University of Sydney and La Trobe University roll out in stages, with full access scheduled for Semester 1, 2027 and the first quarter of 2027 respectively.

  3. bioRxiv

    A bounded agent for paleomicrobiome data processing passes 23 of 24 frozen runs, and every run can be replayed

    Ancient microbial data are scarce and contamination-prone, and a language model left to plan an analysis may cite files that do not exist, apply operations a format forbids, or answer questions the data cannot settle. PaleoRigor confines the model's planner to predefined local modules, validates file references, formats and step dependencies before anything runs, refuses out-of-scope requests with a named reason code, and saves the plan, intermediate files, checksums and decisions. In a preregistered frozen held-out evaluation it passed 23 of 24 runs (95.8%), including all 12 boundary decisions; the same model without the planning contract passed 19.

  4. Why deep-learning runs resist reproduction: interwoven basins leave outcomes unpredictable no matter how precise the initialization

    Deep-learning training outcomes are hard to reproduce, usually blamed on random seeds, batch order and other external noise. This work treats training as a dynamical system in parameter space and proposes that the basins leading to different solutions interleave arbitrarily closely, forming a riddled-basin geometry. In a minimal two-layer model this yields a near-zero uncertainty exponent: initialization precision must improve by almost 24 orders of magnitude to halve the fraction of disagreeing outcomes, and across roughly 500,000 deep-network trainings the analogous parameter-space and function-space exponents are also near zero.

  5. Science Advances

    AlphaFold 3 scores T cell receptor specificity from sequence alone, reaching a median AUC of 0.92 on held-out class I antigens

    Whether a T cell receptor recognizes a given antigen has usually been settled by binding assays or tetramer staining, one receptor at a time. Researchers instead used AlphaFold 3 to predict the peptide-MHC-receptor triad and scored the predicted alignment error at the receptor-peptide-MHC interface. On held-out sets excluded from training, the antigen-centric area under the ROC curve had a median of 0.92 for class I and 0.81 for class II, 0.24 to 0.33 AUC units above the earlier AF2-multimer model.

  6. Nature

    Chromatin-level QTLs in immune cells recover 4.5–22.6% more disease signals than expression QTLs alone

    Across 3.47 million peripheral blood mononuclear cell nuclei from 922 donors, chromatin accessibility was treated as a heritable molecular trait, yielding 243,225 chromatin accessibility quantitative trait loci (caQTLs) across 28 immune cell types, 60% of them cell-type specific. Compared with single-cell expression QTLs (eQTLs), these loci colocalized with 4.5–22.6% more genome-wide association signals for 16 diseases and 44 blood traits; adding cell-type context, wider windows and multi-causal-variant models raised the share of inflammatory bowel disease loci with colocalization evidence from about 16% to about 65%.

  7. Flood alerts gain a 90% band: quantile LSTM ensemble flags risk across Australian catchments

    Flood warning for Australian catchments now comes as a 90% prediction band rather than a single flow curve: three independently trained LSTMs give the mean flow and the 5th and 95th percentiles, and the band's position against a 15% annual exceedance threshold becomes a high, moderate or low risk label. On CAMELS-AUS, the q=0.05 Quantile-LSTM reached the lowest RMSE of 0.0801 in South Australia against 0.0963 for a conventional LSTM, while peak timing error across states ran about 1.94 to 3.12 days with peaks underestimated by roughly 30%.

  8. arXiv

    Brain age from white matter fiber orientation: 3.54-year mean error across 692 people, and predictions rotate with the input

    Brain age prediction no longer relies only on diffusion tensor scalar metrics: white matter fiber orientation distributions (WM-FODs) enter directly as spherical harmonic coefficients, processed by an SE(3) equivariant CNN that preserves their transformation properties under rotation. On 692 participants aged 36-90 from the Human Connectome Project Aging dataset, this combined framework outperformed the evaluated DTI-based models and conventional CNNs; averaging five cross-validation models gave a mean absolute error of 3.54 years and a Pearson correlation of 0.958. The advantage depended on combining WM-FODs with equivariant processing, and accuracy was highest at maximum spherical harmonic order Lmax = 2.

  9. arXiv

    Beyond whether an image is AI: EviScene checks the depicted situation, reaching 52.26% evidence sufficiency

    Image factuality assessment no longer stops at whether a picture is AI-generated: the EviScene multimodal agent refines its own questions during investigation and reaches 79.05% balanced accuracy and 52.26% evidence sufficiency on 1,527 test images.

  10. arXiv

    Reasoning post-training without the full corpus: two-stage contraction cuts the optimization budget to 0.22%

    Reasoning post-training no longer feeds every teacher chain-of-thought to the student at equal weight. Scaling the corpus had raised training cost without proportional gains. Here Monte Carlo reasoning trees estimate process uncertainty at intermediate states to contract the corpus into a teacher-informed pool, and online routing during supervised fine-tuning concentrates updates on what the student has not yet learned, cutting the optimization budget to 0.22–1.91% of full-data SFT across reasoning benchmarks and model families while matching or improving performance.