Skip to main content
Search

University of Southern California

32 related articles

Subscribe to updates
  1. arXiv

    Influence scores for traffic forecasts must pass through conservation constraints: classic rankings shift, and removing the top ten stations raises constrained-model MAE by 44.2%

    When a traffic forecasting model enforces junction flow conservation at its output, a training sample's influence on predictions no longer travels only through the loss but also through the constraint projection. Classical influence functions assume unconstrained learning and cannot separate perturbations the physics keeps from those the projection removes. This work computes influence inside the feasible set, ranks station sensitivity on hourly PeMS District 3 data covering 318 freeway links, and shows that withholding the ten highest-ranked stations raises the constrained model's MAE by 44.2%.

  2. arXiv

    Discharge decisions can retrain the model: capacity-aware discharge cuts readmission-risk learning error by 24%–25%

    Hospitals use readmission-risk predictions to decide who leaves first, but outcomes are recorded only for discharged patients, so retraining a predictor on that data can lock in censoring bias. Coupling a queueing-based Markov decision process with a periodically retrained risk predictor, the work finds that a discharge-up-to policy, which sets a target census and discharges by priority only when occupancy exceeds it, spreads observed lengths of stay through natural arrival variability. In synthetic experiments and calibrated simulations built on patient-level data from a partner hospital, the policy reduced learning error by 24%–25% relative to thresholding on risk and shrank baseline-adjusted readmission cost disparities by 38%–63%.

  3. arXiv

    Unit-test rewards teach models to edit more: PreciseCoder cuts average edits from 7–15 lines to 3 without losing pass rate

    Reinforcement learning from unit tests raises pass rates but also teaches code models to edit more and to invent bugs in programs that are already correct. PreciseCoder adds a signal for semantically verified edit precision and recall against a reference repair; on 4B, 9B and 27B models it cuts the average edit from 7–15 lines to 3 at a pass rate on par with unit-test-only training, and raises the no-edit rate on correct code to 47–52%.

  4. arXiv

    Internal distillation drops the learned projector: RAPID swaps in random projections and cuts up to 21.1% of trainable parameters

    Internal distillation aligns a teacher model's intermediate activations with a student's smaller representation space, and existing methods do this by training extra projection matrices that are discarded after training. RAPID replaces them with a fixed random projection built from the Johnson-Lindenstrauss lemma and balances its losses with exponential moving average and homoscedastic uncertainty weighting. In reasoning post-training with Gemma 3 and Qwen 2.5, its four-benchmark averages stay within 0.15 percentage points of learned projection while trainable parameters fall 8.1% in the Qwen experiment and 21.1% when distilling all activations.

  5. LLMs keep learning after deployment: sleep-style offline consolidation writes new facts into weights without erasing old skills

    Deployed large language models stay frozen at a fixed knowledge cutoff, continual parameter updates cause severe forgetting of earlier tasks, and in-context learning loses information once the session ends. Google Research's sleep framework alternates active interaction with offline consolidation, in which the model opens pre-allocated Mixture-of-Experts memory blocks and updates weights using self-generated practice examples. In multi-stage continual translation into unseen languages such as Manchu and Kalamang, performance dropped near zero, while in-context learning, Cartridges and supervised fine-tuning suffered severe forgetting.

  6. Cross-region neural dynamics stop being drowned by local activity: shared dynamics come first, and nonlinearity gets a location

    Models of cross-region neural dynamics have largely stayed linear, while each region's recorded activity mixes local dynamics with dynamics shared across regions. CroP-DYNO sets learning directly on predicting target-population activity from source activity, so shared dynamics are extracted first, and it lets each model component be specified as linear or nonlinear. On macaque S1–M1 and rat CA1–CA3 data it predicts target activity better than a non-prioritized nonlinear model and a prioritized linear model, and it localizes nonlinear structure and yields interaction strengths consistent with known circuit anatomy.

  7. arXiv

    Frozen time-series models no longer ship as-is: a coding agent rewrites the deployment pipeline and tops six metrics across 198 tasks

    Time-series foundation models forecast well zero-shot, yet deploying them still means hand-building input and output pipelines. TimeRSI has a coding agent repeatedly rewrite preprocessing, features, context selection and post-processing around a frozen forecaster, using errors on historical forecast windows to guide each revision, then picks among the resulting pipelines at deployment by similar past conditions. Across 198 tasks on TIME and FEV-bench it beats all published submissions on six point and probabilistic metrics; against direct inference with the same frozen TimesFM 3.0, TIME MASE falls 1.52% and CRPS 1.61%.

  8. arXiv

    Agentic data synthesis stops just piling up samples: feedback rewrites the synthesis space, lifting task success from 64.5% to 77.8%

    Agentic training-data synthesis has a new shape: EvoSyn turns learner feedback into a revisable synthesis space, so what gets generated follows the task agent's weak spots. Earlier methods used environment feedback, data validation, or learner failures, but observed weaknesses did not directly say what the next round should generate, nor whether existing synthesis dimensions could represent them. On τ-Bench and CAR-Bench with Opus 4.5 as the synthesis model, average task success rises from 64.5% for the strongest synthesis baseline to 77.8%, reaching 70.9% after the first round, with multi-round evolution adding further gains.

  9. arXiv

    Bimanual robots get stereo data without stereo cameras: warp the left view, inpaint the holes, lift transparent-bottle grasping from 11/20 to 18/20

    Large-scale robot datasets are overwhelmingly monocular and single-arm, and capturing bimanual stereo demonstrations would need stereo rigs plus re-teleoperation for every task. StereoEngine offers two complementary routes: when a real monocular video exists, it warps the left view into the right view using monocular depth and inpaints the regions the left camera never observed with a video diffusion model; when no real video exists, it renders a stereo pair in simulation and generates both photorealistic views jointly by tiling them into one frame. On an AgileX PiPER bimanual platform, mid-training or fine-tuning policies on this synthetic data succeeds on 50/60 trials across three tasks versus 36/60 without mid-training, reaching 18/20 on transparent-bottle grasping.

  10. arXiv

    A humanoid catches thrown boxes with onboard vision alone: 65% at 2.5 m, 76% of human catchers

    The humanoid Reflex catches human-thrown boxes using only onboard RGB-D vision, reaching 85.7% catch success in simulation versus 87.9% with privileged box observations, and 65% and 40% on the real robot at 2.5 and 5 m.