In briefState-Isolated Vision Mamba for Underwater Image Enhancement
SummaryThe paper identifies State Contention, where tokens with different update behaviors interfere when sharing one recurrent hidden state in Vision Mamba. It proposes State Isolation, grouping update-compatible tokens into independent state streams, and implements it as IsoMamba, which uses label-free clustering and routes tokens to isolated Mamba experts. The authors report state-of-the-art results on several underwater image enhancement benchmarks.
In briefMachine unlearning for fair federated k-means clustering, explained
SummaryThis paper studies exact machine unlearning for federated k-means clustering that aims to keep group-level fairness. The method keeps separate local summaries for each demographic group, so summaries from unaffected institutions stay valid when fairness weights change. Experiments on three datasets show replay is much faster than retraining when a whole client withdraws, but gains are modest for individual record deletions.
In briefC²LUE: Patient-level ultrasound diagnosis using evidence-aware multi-instance learning, aligned with clinical risk
SummaryThe paper proposes C²LUE, a framework for patient-level ultrasound diagnosis that combines multi-instance learning with risk-aligned regularization. It uses a multi-query cross-attention mechanism on a LoRA-adapted DINOv3 backbone to keep diverse morphological cues from multi-view exams. A dynamic trade-off regularization term penalizes sensitivity-specificity imbalance during training. The authors report consistent gains over prior methods on biopsy-proven thyroid and breast ultrasound datasets.
In briefUnderstanding GFlowNet Training: A Theoretical View
SummaryThis paper studies the theory behind training Generative Flow Networks (GFlowNets), which sample discrete objects in proportion to a positive reward. The authors build a unified theory for the Flow Matching, Detailed Balance, and Trajectory Balance objectives, linking nonzero balance loss and approximate stationarity to sampling accuracy. They establish finite-time convergence results and explain how parameterization, coverage, and moving components affect training.
WhyUnifies sampling-error bounds and finite-time convergence rates for three GFlowNet balance objectives; theoretically solid, though lacking empirical validation.
In briefOpenSocialJax: Testing How AI Agents Adapt to Hidden Rules and Others in Social Dilemmas
SummaryThis paper introduces OpenSocialJax, a benchmark extending Clean Up and Commons Harvest with varied layouts and hidden rules for crafting, payoffs, and resources. It evaluates eight large language models on discovering these rules and adapting to other agents in homogeneous and mixed teams. Rule discovery often improved performance but did not reliably sustain cooperation, with cleaning work concentrated on fewer agents in OpenCleanup and resource depletion in OpenHarvest.
In briefAdaptive DHiSparse: Speculative Decoding on a Load-Adaptive KV Memory Hierarchy
SummaryThis paper addresses KV memory limits in long-context LLM serving by proposing Adaptive DHiSparse, which supports speculative decoding over hierarchical sparse KV and adjusts GPU residency to load. It keeps the required working set, retains more history when GPU memory is free, and fetches the rest via exact fused retrieval. The authors report 793 output tokens/s per decode GPU for GLM-5.3 with DFlash2 at 100K context and concurrency 128, 1.45 times HiSparse.
In briefGLIMPSE: Global-local memory for consistent camera-controlled video generation
SummaryThis paper addresses keeping scenes consistent in camera-controlled video world models when the camera returns to a previously seen view. It presents GLIMPSE, which combines local patch memory for view-specific appearance with recurrent global memory for scene context, using band-wise multi-precision coordinates and geometry-conditioned spectral attenuation to reduce sensitivity to inaccurate projected addresses. Under matched retrieval and read budgets, the authors report better geometric correspondence and photometric consistency than deterministic positional transport.
