Shenzhen University of Advanced Technology

Patent examination asks an agent to find sparse prior art, refine queries, inspect evidence and write a novelty judgment, yet existing benchmarks mostly stop at static candidate sets. PatEx-Gym turns that process into a runnable environment: PatEx-Bench supplies 1K examination tasks, retrieval corpora of over 1M patents and 52K scientific papers, and a runtime supporting iterative search and Office Action generation. PASS synthesizes search trajectories by treating document visits as implicit relevance signals, and the 4B PatEx-Agent fine-tuned on them performs competitively against substantially larger general-purpose agents.

Patent examination asks an agent to find sparse prior art, refine queries, inspect evidence and write a novelty judgment, yet existing benchmarks mostly stop at static candidate sets. PatEx-Gym turns that process into a runnable environment: PatEx-Bench supplies 1K examination tasks, retrieval corpora of over 1M patents and 52K scientific papers, and a runtime supporting iterative search and Office Action generation. PASS synthesizes search trajectories by treating document visits as implicit relevance signals, and the 4B PatEx-Agent fine-tuned on them performs competitively against substantially larger general-purpose agents.

Web-GUI agents locate buttons and links from screenshots, and high-resolution screenshots plus multi-turn histories create excessive visual tokens that raise inference cost. STP-GUI replaces independent token scoring with query-guided, spatially coherent window-level selection, trained through a lightweight selector while the vision-language backbone stays frozen. On Online-Mind2Web, retaining 21% of tokens preserves 77.7% of full-token performance while cutting inference latency by 31% and FLOPs by 57%; results on Android in the Wild suggest the method transfers to mobile GUI environments.

Web-GUI agents locate buttons and links from screenshots, and high-resolution screenshots plus multi-turn histories create excessive visual tokens that raise inference cost. STP-GUI replaces independent token scoring with query-guided, spatially coherent window-level selection, trained through a lightweight selector while the vision-language backbone stays frozen. On Online-Mind2Web, retaining 21% of tokens preserves 77.7% of full-token performance while cutting inference latency by 31% and FLOPs by 57%; results on Android in the Wild suggest the method transfers to mobile GUI environments.

Looped Transformers deepen reasoning by re-running a shared block at test time, but the usual carry-last rule lets later iterations overwrite correct intermediate results. InfiLoop replaces that rule with learned attention over recurrent outputs: a shared pseudo-query scores each candidate once, a learned exponential decay discounts older ones, and the aggregation has an exact streaming recurrence needing only a weighted-state vector and a normalization scalar per token. On Sudoku-Extreme a 7M-parameter model reaches 97.9% exact accuracy versus 94.2% for FPRM, and hits 90.9% at 1,872 layers where FPRM needs 3,744 layers for 90.1%.

Looped Transformers deepen reasoning by re-running a shared block at test time, but the usual carry-last rule lets later iterations overwrite correct intermediate results. InfiLoop replaces that rule with learned attention over recurrent outputs: a shared pseudo-query scores each candidate once, a learned exponential decay discounts older ones, and the aggregation has an exact streaming recurrence needing only a weighted-state vector and a normalization scalar per token. On Sudoku-Extreme a 7M-parameter model reaches 97.9% exact accuracy versus 94.2% for FPRM, and hits 90.9% at 1,872 layers where FPRM needs 3,744 layers for 90.1%.

Nereus is a cost-aware runtime for reinforcement-learning post-training of large language models that represents each model-stage replica as an Elastic Model Unit and orchestrates Split, Merge, Extend, and Destroy primitives through a global transition graph, adapting the DP/TP/PP execution plan online under a payback rule; on a trace built from real data, online TP/PP adaptation cuts average step latency by 27.7% versus the initial fixed TP/PP layout, six transitions in a 1,000-step run scaling to 1,024 GPUs consume 0.079% of total run time, and end-to-end 8B PPO throughput improves by 2.14-7.27x over OpenRLHF and 1.10-1.47x over Verl.

Nereus is a cost-aware runtime for reinforcement-learning post-training of large language models that represents each model-stage replica as an Elastic Model Unit and orchestrates Split, Merge, Extend, and Destroy primitives through a global transition graph, adapting the DP/TP/PP execution plan online under a payback rule; on a trace built from real data, online TP/PP adaptation cuts average step latency by 27.7% versus the initial fixed TP/PP layout, six transitions in a 1,000-step run scaling to 1,024 GPUs consume 0.079% of total run time, and end-to-end 8B PPO throughput improves by 2.14-7.27x over OpenRLHF and 1.10-1.47x over Verl.

This work turns the sparsity coefficient of convolutional sparse coding from a manually fixed hyperparameter into a differentiable variable jointly learned with network parameters through unfolded FISTA iterations, interprets that coefficient within an information bottleneck framework as the controller of the compression-retention trade-off, achieves competitive clean-data recognition on CIFAR-10/100 and ImageNet-1K with greatly improved robustness under various input perturbations, and introduces a label-free post-training adaptation strategy that adjusts compression strength using a small set of unlabeled corrupted samples while keeping the main network parameters frozen.

This work turns the sparsity coefficient of convolutional sparse coding from a manually fixed hyperparameter into a differentiable variable jointly learned with network parameters through unfolded FISTA iterations, interprets that coefficient within an information bottleneck framework as the controller of the compression-retention trade-off, achieves competitive clean-data recognition on CIFAR-10/100 and ImageNet-1K with greatly improved robustness under various input perturbations, and introduces a label-free post-training adaptation strategy that adjusts compression strength using a small set of unlabeled corrupted samples while keeping the main network parameters frozen.

The work proposes HEAL, which applies causal noise intervention to multi-head outputs and counterfactual Difference-in-Differences to categorize attention heads into redundant, visual, language, and synergy types, finds that hallucinations occur when the visual-language information distribution inside synergy heads drifts away from a healthy equilibrium rather than correlating strongly with the quantity or strength of modality-specific heads, and accordingly injects dynamic calibration factors into the value vectors of synergy heads at inference time, reducing hallucinations across multiple MLLMs and benchmarks.

The work proposes HEAL, which applies causal noise intervention to multi-head outputs and counterfactual Difference-in-Differences to categorize attention heads into redundant, visual, language, and synergy types, finds that hallucinations occur when the visual-language information distribution inside synergy heads drifts away from a healthy equilibrium rather than correlating strongly with the quantity or strength of modality-specific heads, and accordingly injects dynamic calibration factors into the value vectors of synergy heads at inference time, reducing hallucinations across multiple MLLMs and benchmarks.