Skip to main content
Search

Duke University

18 related articles

Subscribe to updates
  1. arXiv

    Neural circuits that grow while working: trained at just three developmental snapshots, a controller keeps visuomotor behavior at untrained ages

    A recurrent neural controller must keep behaving while neurons are born, move, and change their intrinsic time constants during development; controllers optimized only on the initial or the mature substrate generally failed at other ages. The study feeds developmental time into a compact connectivity generator and optimizes the same generator at three, five, or fifteen developmental snapshots, testing it on two geometrically distinct substrates and two visuomotor tasks. Optimization at only three snapshots was often enough to keep performance high and consistent at trained and untrained ages, while the generated connectomes changed substantially with age and did not converge on a single sparsity or weight trajectory.

  2. arXiv

    A verifiable proving ground for egocentric AI assistants: humans act on instructions, and the best model tops out at 20% on long-horizon guidance

    Testing egocentric AI assistants has usually required real people, and the expensive, scarce interaction data left open whether today's multimodal models can actually help someone in the physical world. InteractiveWorld moves that loop into a verifiable environment: the assistant issues instructions, a human agent observes the scene, understands the instruction and acts independently, and the environment executes the action and updates what both see. InteractiveBench built on it holds 120 multi-turn tasks across six categories, each instruction shaping the outcome and scored by verifiable state-based rewards; across 12 frontier multimodal models, Claude Fable 5.1 reaches at most 20% and 25% on long-horizon Guidance and Planning tasks, respectively.

  3. arXiv

    Visual edits lose their grip mid-decoder: equal-strength interventions fade before answers settle, and calibrated routing turns that into attribute accuracy

    Applying the same visual-residual edit at every decoder layer with matched RMS strength reveals a persistent drop in effect, called the Evidence Anchoring Boundary; routing interventions by question type and calibration improves full-set attribute accuracy across three checkpoints, while universal subtraction is neutral or harmful.

  4. arXiv

    Robot memory stops guessing from keyframes: per-tick text lines lift success from 44.5% to 66.0% under a fixed policy

    Gamma replaces pixel keyframe memory with auditable text lines, raising average success on RoboMME's sixteen memory tasks from 44.5% to 66.0% under a fixed execution policy and matching the privileged subgoal ceiling on three tasks.

  5. arXiv

    Attention already sees the harm, but the MLP writes the refusal only at answer onset

    Across 18 Base–Instruct model pairs, harmfulness is more decodable from attention updates while refusal evidence is written mainly by MLPs, and post-training strengthens that writing without improving recognition.

  6. arXiv

    Perovskite models stumble on new alloys: encoding solid-solution continuity cuts unseen-alloy error by 15.7–19.0%

    V-PaSS encodes the physical continuity of perovskite solid solutions, cutting error by 15.7–19.0% on unseen absorber alloys under a protocol where every test device comes from an unseen laboratory.

  7. arXiv

    Perovskite model scoring gets a new split: a lookup table matches the published best on random splits, and neural models overtake trees on unseen institutions

    Evaluating perovskite solar-cell machine learning with provenance-grouped splits shows a lookup table that learns nothing matches the published state of the art on a random split, while neural models overtake tree ensembles on unseen institutions and later years.

  8. arXiv

    Expectations under a product of densities now need no density: one regression merges component samples, error 0.0001 on HIGGS

    Expectations under a product of densities become a regression on a consensus and a discrepancy: using only Monte Carlo draws from each component and no density evaluations, the method estimates arbitrary expectations. On 11 million simulated HIGGS collision events, PMC-MLP reaches an estimation error of 0.0001 with 0.1092 minutes of merge time, about 10% of full-data MCMC runtime; in a privacy-restricted respiratory clinical trial, every PMC variant merges more cheaply than GBF.

  9. arXiv

    Mid-training a language model on raw video lifts video benchmarks by 2.9 points and image benchmarks by 5.1 points while preserving text ability

    This work asks whether raw web video, with no captions and no text loss, can serve as mid-training data for a pretrained language model: frames are encoded into continuous visual tokens and the model learns to predict the next visual token, Qwen3-1.7B is mid-trained on raw clips from YT-Temporal-1B and then given the same image-text instruction tuning as a model without mid-training, yielding 2.9 points higher on video benchmarks and 5.1 points higher on image benchmarks, with text average rising from 48.0 to 48.9, and caption prediction offering no advantage over visual-token prediction.

  10. arXiv

    MASA recasts sparse-attention selection as matrix approximation, lifting long-context accuracy across three sparse frameworks at no extra budget

    The authors argue that existing sparse-attention selectors keep large or high-mass entries of the attention matrix even though Transformers use that matrix through multiplication with value vectors, and propose MASA, which replaces attention-mass ranking with a closed-form score measuring each sparse unit's reduction in matrix-product approximation error; plugged into MInference, SeerAttention, and FlexPrefill without changing their sparse kernels, candidate units, or budgets, it yields consistent accuracy gains on RULER, InfiniteBench, and LongBench with negligible latency overhead.