A fully differentiable emulator of the Atari 2600 VCS is presented, reproducing all 64 ALE games with bit-for-bit accuracy in RAM and screen output. The system enables gradient-based explainable AI by providing a complex, fully known ground truth, with both Julia and JAX implementations validated against a reference emulator and supporting high-throughput training on GPUs.
AdaR enables flexible test-time computing on graphs without parameter changes by using adaptive recurrence. It derives step dependence as a necessary and sufficient condition for convergence and incorporates normalized step information and representation-target relations into recurrent updates, guided by gradient-based supervision signals. Empirical results show AdaR outperforms strong baselines in both inductive and transductive graph learning settings.
Interleaved speech-language models undergo an implicit transcription phase where spoken words become decodable as text tokens in intermediate layers, despite no speech recognition training. Up to 77% of the data shows the spoken word appearing as a top candidate text prediction, followed by a transition to text-based next-word prediction before returning to speech. This behavior is influenced by interleaved training and text LM initialization, and correlates with spoken knowledge performance.
Fed-CausalDiff introduces a federated causal diffusion framework that enables do-simulation and policy evaluation in decentralized settings. It decomposes latent state evolution into global and local components, allowing decoupled synchronisation to reduce communication cost while maintaining accurate causal inference.
A structured prompt framework improves local LLM accuracy in environmental monitoring by transforming raw sensor data into enriched textual representations. Evaluations on indoor and outdoor datasets show local model accuracy increases from 50.9% to 81.7% indoors and from 63.7% to 89.3% outdoors with enriched prompts, while maintaining low latency near 0.22 seconds in no-chain-of-thought mode.
Input diversity, a common practice in transfer attacks, improves success on standard surrogates but reduces it on robust ones. This regime-dependent effect, called the Scissors Effect, is driven by gradient geometry, with resize operations degrading alignment in robust models. A training-free rule (CG-DI) adjusts diversity based on local gradient consistency to preserve attack success across surrogate types.
Generative Robust Optimisation (GRO) introduces a deep generative model to define uncertainty sets, capturing nonlinear correlations, asymmetry, and multimodality. A five-point evaluation framework assesses neural network-based uncertainty sets across reconstruction fidelity, distribution matching, latent regularity, robust relevance, and computational tractability, with experiments validating GRO's effectiveness in production planning and facility location problems.
A learnable scaling parameter called Quantum Measurement Temperature (QMT) is introduced to rescale quantum measurement outputs in hybrid quantum neural networks. This approach mitigates measurement-induced logit contraction, enhancing gradient magnitude and stability during training without altering the quantum circuit or measurement operators. Experiments show improved logit separation, gradient strength, and classification accuracy in protein and image classification tasks.
A piezoelectric deep material network (PDMN) is proposed to efficiently homogenize two-phase piezoelectric composites. The framework embeds electromechanical homogenization relations into its architecture, enabling physics-informed, semi-analytical predictions with over three orders of magnitude lower computational cost than direct numerical simulation, validated on PVDF-LiNbO3 and viscoelastic-piezoelectric composites under nonlinear loading.
CCPL introduces a lightweight framework that anchors class prompts to frozen concept prototypes, improving few-shot CLIP adaptation. It achieves better base-to-new performance on DTD and EuroSAT compared to CoOp, with consistent gains from text-space concept regularization, though results vary by dataset and protocol.
This paper introduces a stationary mean-field game framework that directly incorporates distributional model uncertainty into population-coupled dynamics. It establishes a robust dynamic programming principle, proves existence of a stationary robust equilibrium, and presents the first algorithm with convergence guarantees. The mean-field solution approximates finite-population equilibria and provides explicit non-asymptotic error bounds under model uncertainty.
SiM enables dynamic routing in multi-task model merging without additional training or task ID access. It uses SVD-based manifold approximations and projects test inputs onto precomputed task manifolds to route inputs to relevant experts, improving performance and reducing the gap to individual expert levels.
On-Policy Distillation (OPD) suffers from position bias where later tokens provide poor supervision. We introduce Importance-Weighted On-Policy Distillation (IW-OPD), which assigns weights based on distribution discrepancy, prioritizing early tokens. IW-OPD converges faster and achieves up to 6.9 point performance gains on AIME-2025.
A generative surrogate framework using a Variational Autoencoder approximates Gaussian Process priors, bypassing costly covariance operations. The VAE+Hidden Markov Model architecture enables fast, scalable stellar flare detection in large astronomical time series, matching exact models in structural fidelity while reducing computational time significantly.
A new scheduling algorithm, Smallest Volume First (SVF), reduces LLM inference latency by optimizing key-value cache management. Theoretical analysis shows a worst-case competitive ratio reduced from 48 to 5, with 1-bit SVF achieving strong performance using minimal information. Evaluations on Llama-3.1 models confirm improvements in both average and tail latency, with the approach integrated into vLLM.
HDSO enables safe, auditable skill updates for LLM agents without training, using falsifiable hypotheses and validation. On ALFWorld, it improves Qwen3-8B by +6.9 Avg. SR points and maintains a +7.1-point gain under noisy feedback, with validated skills transferable across runs and models when diagnostic alignment is achieved.
FAPS is the first function-space posterior sampling framework that unifies stochastic-process regression and PDE inverse problems. It uses pretrained flow-matching priors and Langevin correction with low-rank covariance preconditioning to enable efficient, accurate posterior inference from sparse, noisy data with coherent uncertainty quantification.
HRLLI introduces a hierarchical reinforcement learning framework that adapts natural-language instructions dynamically during decision-making. It decomposes instructions into stage-specific guidance elements and uses a select-to-act paradigm to enable real-time selection of relevant instruction pieces, improving sample efficiency and performance in complex environments.
SAFER is a training-free framework that enhances robustness of test-time adaptation by using reliability-guided augmentation. It generates stochastic augmentations, pools predictions via correlation-weighted aggregation with outlier detection, and includes adaptive mixing to preserve clean performance under adversarial attacks. Evaluations on PACS, VLCS, and OfficeHome show improved resilience without sacrificing clean accuracy.
Parsimoniously activated dictionary learning (PADL) establishes a structured generative model with auxiliary latent variables, enabling maximum a posteriori estimation. This framework provides generalization guarantees and an analytical characterization of the tradeoff between sparsity, storage cost, and reconstruction accuracy, allowing data-driven hyperparameter estimation. The resulting algorithm achieves better reconstruction performance and accelerates inference in vision-language models.