A model-driven approach generates families of reinforcement learning environments using a hybrid genetic algorithm. Environment variants are created through model transformations guided by a state-of-the-art model transformation engine, enabling scalable and error-resistant development. The method is validated in wildfire mitigation and curriculum learning scenarios.
A single ReLU recurrent neural network with fixed weights and hidden dimension can uniformly approximate any continuous function on [-1,1] as its runtime increases. This is achieved via a new model, the Turing machine with neural units (TMNU), which balances algorithmic flexibility with bounded simulation by RNNs. The convergence rates match polynomial approximation rates, and minimax lower bounds confirm that runtime is an essential, unavoidable resource.
QCPIKAN is the first quantum-classical physics-informed Kolmogorov-Arnold network designed to solve partial differential equations. It uses Chebyshev-polynomial KAN layers and parameterized quantum circuits to embed physical constraints into training, achieving exponential error convergence and reduced numerical dispersion. Validated on seepage scenarios in porous media, it outperforms existing quantum-classical neural networks in prediction accuracy, error control, and dynamic tracking.
A quantum version of ring all-reduce reduces per-link communication by a factor of two using entanglement and superdense coding, without altering model or gradient computations. It achieves information-theoretically secure aggregation via verified entanglement, with a 2x overhead in GHZ copies, and provides exponential communication advantages in gradient conflict detection for specific auditing tasks.
Temporal difference learning reduces variance by aggregating over multiple trajectories. The study shows TD variance is asymptotically bounded above by Monte Carlo estimators, and shorter horizon updates reduce variance for fixed samples. Direct Advantage Estimation acts as a regression-adjusted control variate, achieving tighter variance bounds than TD in large samples.
A study of sequential Direct Preference Optimization finds that later training does not uniformly degrade earlier learned preferences. The effect varies by objective relationship, signal strength, and training order, ranging from partial degradation to positive transfer. Pair-level analysis reveals heterogeneous changes, with high-confidence preference pairs sometimes improving despite aggregate metric stability.
Manifold Bandits introduces Bayesian Manifold Curriculum (BMC), a framework that models problem sampling as a structured bandit problem in LLMs' latent space. BMC organizes tasks into a hierarchical tree and uses Bayesian learning to guide sampling, revealing tradeoffs between learning signal, task diversity, and evaluation relevance. Prioritizing difficulty alone fails to achieve strong downstream performance, underscoring the need for structure and type-aware sampling.
A new framework enables large language models to learn 'Connect the Dots' by using reinforcement learning with long rollout sequences. The method includes tailored tasks and environments to foster meta-capability development, showing strong cross-domain generalization and performance in out-of-distribution settings. Implementations are available at https://github.com/agentscope-ai/Trinity-RFT/tree/research/cod/examples/research_cod.
This work identifies a dual collapse in latent reasoning: gradient attenuation and representational drift. It proposes Trajectory and Space Supervision, showing that generative reconstruction preserves information capacity better than geometric compression. The Unified Latent Probe measures mutual information between latent trajectories and reasoning steps, revealing an information-performance binding in reasoning accuracy.
HydraHead introduces a head-level hybridization of Full and Linear Attention, leveraging interpretability to select retrieval-critical heads and fuse outputs via a scale-normalized module. Trained on 15B tokens, it achieves over 69% improvement over baseline at 512K context length, outperforming layer-wise hybrids and approaching Qwen3.5's performance on long-context tasks.
GLM-5.2 UD-IQ2_M runs at ~7.3 tokens per second on 4×RTX 3090s with 192GB DDR5 RAM using llama.cpp expert offload. Reducing quantization from IQ2 to IQ1 provided no speed gain, while increasing CPU threads from 6 to 12 improved performance by 22%. Decode is limited by CPU compute, not memory bandwidth, and the offloaded experts must be explicitly distributed across GPUs to avoid out-of-memory errors.
The AI scaling debate overlooks that maximizing model FLOP utilization is more critical than buying more GPUs. Frontiers like xAI operate at sub-10% MFU, while historical models achieved 21% to 70% MFU, indicating systemic inefficiencies in scheduling, networking, and cluster management. Anjney Midha argues that AI infrastructure must evolve into efficient, aligned, and responsible systems, with 'output maxing' emerging as a new discipline for frontier AI.
A user asks if anyone with sufficient computing resources can create a large distillation dataset of 70-1 million examples from GLM5.2. The goal is to enable better training of smaller models like Qwen3.5, benefiting the broader community.
Discriminator-Guided RL (DRL) uses a pretrained representation space to train a discriminator that separates real data from model-generated samples. Its logit is used as a reward in KL-regularized RL, aligning model outputs with visual and semantic realism without human preferences. DRL improves FID and semantic FD across models like SiT and JiT, and enhances the Pareto frontier between preference and fidelity.
Essential Subspace Merging (ESM) reduces inter-task interference by focusing on principal directions of activation shifts. ESM++ extends this with dynamic expert selection via prototype-based routing, enabling efficient, training-free multi-task model merging.
Safety Reflection Pretraining inserts short safety reflections into pretraining data to enable self-monitoring in language models. Experiments with 1.7B models on FineWeb-Edu show improved safety accuracy and reduced attack success rates, with MedSafetyWorld demonstrating that the method better prevents unsafe behaviors from being generalized from safe data than data filtering or rewriting.
Stochastic momentum methods like HB and ASGD show distinct batch-size tradeoffs in compute efficiency and serial runtime. HB maintains SGD-level compute efficiency over a batch-size window up to a factor \sqrt{\kappa} larger than SGD's critical batch size, while ASGD improves small-batch efficiency for rapidly decaying spectra but sacrifices it for larger batches in exchange for reduced serial runtime.
AGDN introduces a graph neural network framework that addresses topological priors and connectivity loss in TSP. It uses a MixScore transition matrix and anisotropic diffusion to enable efficient information exchange, outperforming existing methods across diverse problem sizes and distributions while maintaining competitive computation time. The implementation is available on GitHub.
A new decision-focused RL framework jointly trains a forecaster and charging controller to handle unknown EV departure times. By aligning forecast accuracy with downstream decision quality, the method achieves up to 14% higher total reward and a 55% reduction in unsupplied energy compared to standard RL approaches.
MAST, a mechanism-guided unlearning method, achieves targeted forgetting of RLVR-induced reasoning with minimal collateral damage. On Qwen2.5-Math-1.5B and Qwen3-1.7B-Base, it significantly reduces MATH performance (45/150 to 37/15-0) while preserving GSM8K accuracy by +0.8 points and maintaining MATH retention at -0.5 points. Results hold across different seeds, objectives, and models, showing superior stability over full-parameter unlearning.