Skip to main content
Search

# Curriculum Learning

12 related articles

Subscribe to updates
  1. arXiv

    Why curriculum learning pays off in reinforcement learning: sparse rewards can stall learning without one, while dense rewards gain little

    Curriculum learning often helps more in reinforcement learning than in supervised learning, and a tractable target-interception model now explains why. With dense rewards, curricula bring only marginal gains; with sparse rewards, a curriculum can be required for learning to occur at all. When the agent must both minimize effort and accumulate reward, competing optima appear and can trap it in a quitting solution. Useful curricula do two things: they make rewards less sparse to speed learning, and they shape the reward landscape to steer agents away from bad local optima.

  2. arXiv

    Ultrasound labels don't need every frame: EvoDiff generates diffusion probes, then sends only retrieved real images to experts

    EvoDiff recasts active-learning sample selection as generate-then-retrieve: diffusion sampling in feature space produces synthetic probes, and only the real unlabeled ultrasound images those probes retrieve are sent to experts for annotation. Across 15 datasets and 16,696 images spanning six organ tasks, it beat random sampling and several active-learning baselines in most settings on both ResNet-50 and the USFM foundation model, and a model-based annotation-effort metric estimated a 24% to 84% workload reduction relative to random sampling.

  3. Splitting a humanoid's upper and lower body into two controllers cuts upper-body joint error to about 0.1 radians on H1 and H1-2

    Humanoid whole-body control usually drives the upper and lower body with one policy, so large arm motions can disturb balance and the two tasks interfere during training. JAEGER splits the body into two independent controllers that share observations and rewards, initializes the upper body with supervised learning, then trains the whole body with reinforcement learning while supporting both root-velocity and joint-angle command modes. In simulation comparisons on the H1 and H1-2 platforms, JAEGER's upper-body joint tracking error is lower than HumanPlus, ExBody and OmniH2O, with pose-mode joint errors about 30% to 60% of those methods.

  4. arXiv

    GRPO curricula stop following a fixed schedule: probe accuracy steers difficulty sampling, with the largest gains at intermediate levels

    In GRPO reinforcement fine-tuning, a fixed easy-to-hard curriculum can drift from the model's actual progress. ClosedLoopGaussian keeps the Gaussian sampler's form but uses periodic probe accuracy to adjust how fast training advances and how widely it rehearses, leaving the reward, architecture and optimizer unchanged. On Countdown, mean exact-match accuracy over four seeds exceeds both the fixed curriculum and vanilla GRPO, with gains concentrated at intermediate difficulty, though seed dispersion is large and the comparison is read descriptively.

  5. arXiv

    Agents iterate their own data recipes: math and code selection strategies gain up to 13.34% under fixed training

    Agents use experimental feedback to revise data-engineering strategies, and across four tasks—math and code data selection, math reward design, and math curriculum learning—the selected strategies gain up to 13.34% absolute over generic baselines on AIME25 Pass@32 under fixed training conditions.

  6. Nature Methods

    A billion-scale mass spectral corpus arrives: 1.54 billion spectra across 35 species give de novo sequencing models a shared starting point

    MassNet is an AI-ready mass spectrometry proteomics corpus of about 1.54 billion tandem mass spectra from 648 projects across 35 species, paired with a columnar tensor storage format called MSDT. Previously, mass spectra were scattered across heterogeneous formats such as MGF and mzML, with peak lists stored separately from peptide annotations and coverage skewed toward human samples. XuanjiNovo, a non-autoregressive model trained on this corpus, improves over state-of-the-art methods by 38.35% to 143.57% on human Astral data and also supports retention time prediction and PSM rescoring.

  7. arXiv

    X-ray security scans skip human labels: mining objects from unlabeled scans lifts cluttered-bag localization by up to 23 points

    Universal object localization in X-ray security screening has been limited by scarce expert annotation, forcing reliance on RGB-pretrained segmentation models that falter on overlapping, weakly bounded baggage scans. LAO-X mines object instances from unlabeled X-ray scans, recomposes them in the absorbance domain under the Beer-Lambert law to build cluttered scenes with holistic and part pseudo-masks, then fine-tunes a SAM2 localizer through an occlusion-controlled curriculum. Across six X-ray benchmarks, it raises category-agnostic mAP over SAM2 and X-ray-specific baselines by 2% to 23% with no human-annotated labels, reaching 53.33% mAP on DET-COMPASS.

  8. arXiv

    Long-MDR pushes online RL for multimodal deep-research agents to 128k context and 75 tool turns, with a 9B model leading five of six benchmarks among 7B–9B agents at a 50-turn evaluation budget

    This work scales online reinforcement learning for multimodal deep-research agents to 128k context and 75+ tool-interaction turns and introduces Long-MDR, a three-component recipe of On-Policy Distillation Warmup, Progressive Horizon Expansion, and Entropy-Triggered Rescue, to address slow early reward growth and later entropy collapse in long-horizon training; at a 50-turn tool-interaction evaluation budget, the trained Long-MDR-9B ranks first on five of six benchmarks among the compared 7B–9B agents.

  9. arXiv

    ActiveSaddler co-evolves training scenarios with the harness, raising Pass@1 by 4.4 and 7.5 points on GAIA2 and Terminal-Bench 2.0

    The work formulates training-scenario selection in offline harness optimization as an automated curriculum-learning problem and introduces ActiveSaddler, which models the curriculum as a non-stationary bandit with failure-pattern arms, estimates each arm's potential learning progress, and adaptively balances revisiting known weaknesses against exploring unseen scenarios, improving test Pass@1 by 4.4 and 7.5 percentage points over the same harness optimizer with a scenario order fixed before optimization on GAIA2 and Terminal-Bench 2.0 respectively.

  10. arXiv

    LT-OPD uses on-policy self-distillation on its own generated trajectories to lift average retained performance at 5% visual tokens from 68.6% to 82.3%

    The work proposes LT-OPD, which supervises a student keeping only 5% of visual tokens on its own generated trajectories using a frozen full-token copy of the same model as teacher, together with a curriculum that progressively lowers the token budget, raising average retained performance on nine benchmarks with Qwen3.5-4B from 68.6% to 82.3% while cutting KV cache by 85.2% and prefill FLOPs by 85.4% with no added inference overhead.