GEAR: Granularity-Adaptive Advantage Reweighting for LLM Agents via Self-Distillation

Reinforcement learning has become a widely used post-training approach for LLM agents, where training commonly relies on outcome-level rewards that provide only coarse supervision. While finer-grained credit assignment is promising for effective policy updates, obtaining reliable local credit and assigning it to the right parts of the long-horizon trajectory remains an open challenge. In this paper, we propose Granularity-adaptivE Advantage Reweighting (GEAR), an adaptive-granularity credit assignment framework that reshapes the trajectory-level GRPO advantage using token- and segment-level signals derived from self-distillation. GEAR compares an on-policy student with a ground-truth-conditioned teacher to obtain a reference-guided divergence signal for identifying adaptive segment boundaries and modulating local advantage weights. This divergence often spikes at the onset of a semantic deviation, while later tokens in the same autoregressive continuation may return to low divergence. GEAR therefore treats such spikes as anchors for adaptive credit regions: where the student remains aligned with the teacher, token-level resolution is preserved; where it departs, GEAR groups the corresponding continuation into an adaptive segment and uses the divergence at the departure point to modulate the segment' s advantage. Experiments across eight mathematical reasoning and agentic tool-use benchmarks with Qwen3 4B and 8B models show that GEAR consistently outperforms standard GRPO, self-distillation-only baselines, and token- or turn-level credit-assignment methods. The gains are especially strong on benchmarks with lower GRPO baseline accuracy, reaching up to around 20\% over GRPO, suggesting that the proposed adaptive reweighting scheme is especially useful in more challenging long-horizon settings.

What to Ignore, What to React: Visually Robust RL Fine-Tuning of VLA Models

Reinforcement learning (RL) fine-tuning has shown promise for Vision-Language-Action (VLA) models in robotic manipulation, but deployment-time visual shifts pose practical challenges. A key difficulty is that standard task rewards supervise task success, but offer limited guidance on whether a visual change is task-irrelevant or changes the behavior required for manipulation. We propose PAIR-VLA (Paired Action Invariance & Sensitivity for Visually Robust VLA), an RL fine-tuning framework to address this difficulty by adding two auxiliary objectives over paired visual variants during PPO optimization: an invariance term that reduces the discrepancy between action distributions for a task-preserving pair (e.g., different distractors), and a sensitivity objective that encourages separable action distributions for a task-altering pair (e.g., target object in a different pose). Together, these objectives turn visual variants from mere observation diversity into behavior-level guidance on policy responses during RL fine-tuning. We evaluate on ManiSkill3 across two representative VLA architectures, OpenVLA and $π_{0.5}$, under diverse out-of-distribution visual shifts including unseen distractors, texture changes, target object pose variation, viewpoint shifts, and lighting changes. Our method consistently improves over standard PPO, achieving average improvements of 16.62% on $π_{0.5}$ and 9.10% on OpenVLA. Notably, ablations further show generalization across visual shifts: invariance guidance learned from distractor and texture variants transfers to target-pose and lighting shifts, while adding sensitivity guidance on target-pose variants further improves robustness to nuisance shifts, highlighting the broader transferability of behavior-level RL guidance.

Orchestrating Spatial Semantics via a Zone-Graph Paradigm for Intricate Indoor Scene Generation

Autonomous 3D indoor scene synthesis breaks down in non-convex rooms with tightly coupled spatial constraints. Data-driven generators lack topological priors for long-horizon planning, while iterative agents fragment semantics and become geometrically brittle. We present ZoneMaestro, a unified framework that shifts the paradigm from object-centric synthesis to Zone-Graph Orchestration. By internalizing a novel zone-based logic, ZoneMaestro translates high-level semantic intent into functional zones and topological constraints, enabling robust adaptation to diverse architectural forms. To support this, we construct Zone-Scene-10K, a large-scale dataset enriched with explicit Zone-Graph annotations. We further introduce an Alternating Alignment Strategy that cycles between reasoning internalization and Zone-Aware Group Relative Policy Optimization (Z-GRPO), effectively reconciling the tension between semantic richness and geometric validity without relying on external physics engines. To rigorously evaluate spatial intelligence beyond convex primitives, we formally define the task of Intricate Spatial Orchestration and release SCALE, a stress-test benchmark for irregular indoor scenarios with complex, dense spatial relations. Extensive experiments demonstrate that ZoneMaestro resolves the density-safety dichotomy, significantly outperforming state-of-the-art baselines in both structural coherence and intent adherence.

OphMAE: Bridging Volumetric and Planar Imaging with a Foundation Model for Adaptive Ophthalmological Diagnosis

The advent of foundation models has heralded a new era in medical artificial intelligence (AI), enabling the extraction of generalizable representations from large-scale unlabeled datasets. However, current ophthalmic AI paradigms are predominantly constrained to single-modality inference, thereby creating a dissonance with clinical practice where diagnosis relies on the synthesis of complementary imaging modalities. Furthermore, the deployment of high-performance AI in resource-limited settings is frequently impeded by the unavailability of advanced three-dimensional imaging hardware. Here, we present the Ophthalmic multimodal Masked Autoencoder (OphMAE), a multi-imaging foundation model engineered to synergize the volumetric depth of 3D Optical Coherence Tomography (OCT) with the planar context of 2D en face OCT. By implementing a novel cross-modal fusion architecture and a unique adaptive inference mechanism, OphMAE was pre-trained on a massive dataset with of 183,875 paired OCT images derived from 32,765 patients. In a rigorous benchmark encompassing 17 diverse diagnostic tasks with 48,340 paired OCT images from 8,191 patients, the model demonstrated state-of-the-art performance, achieving an Area Under the Curve (AUC) of 96.9% for Age-related Macular Degeneration (AMD) and 97.2% for Diabetic Macular Edema (DME), consistently surpassing existing single-modal and multimodal foundation models. Crucially, OphMAE exhibits robust engineering adaptability: it maintains high diagnostic accuracy, such as 93.7\% AUC for AMD, even when restricted to single-modality 2D inputs, and demonstrates exceptional data efficiency by retaining 95.7% AUC with as few as 500 labeled samples. This work establishes a scalable and adaptable framework for ophthalmic AI, ensuring robust performance across different tasks.

SAIL: Structure-Aware Interpretable Learning for Anatomy-Aligned Post-hoc Explanations in OCT

Optical coherence tomography (OCT), a commonly used retinal imaging modality, plays a central role in retinal disease diagnosis by providing high-resolution visualization of retinal layers. While deep learning (DL) has achieved expert-level accuracy in OCT-based retinal disease detection, its "black box" nature poses challenges for clinical adoption, where explainability is essential for clinical trust and regulatory approval. Existing post-hoc explainable AI (XAI) methods often struggle to delineate fine-grained lesion structures, respect anatomical boundaries, or suppress noise, limiting the trustworthiness of their explanations. To bridge these gaps, we propose a Structure-Aware Interpretable Learning (SAIL) framework that integrates retinal anatomical priors at the representation level and couples them with semantic features via a fusion design. Without modifying standard post-hoc explainability methods, this representation yields sharper and more anatomically aligned attribution maps. Comprehensive experiments on diverse OCT datasets demonstrate that our structure-aware method consistently enhances interpretability, producing clinically meaningful and anatomy-aware explanations. Ablation studies further show that strong interpretability requires both structural priors and semantic features, and that properly fusing the two is critical to achieve the best explanation quality. Together, these results highlight structure-aware representations as a key step toward reliable explainability in OCT.

Driving risk emerges from the required two-dimensional joint evasive acceleration

Most autonomous driving safety benchmarks use time-to-collision (TTC) to assess risk and guide safe behaviour. However, TTC-based methods treat risk as a one-dimensional closing problem, despite the inherently two-dimensional nature of collision avoidance, and therefore cannot faithfully capture risk or its evolution over time. Here, we report evasive acceleration (EA), a hyperparameter-free and physically interpretable two-dimensional paradigm for risk quantification. By evaluating all possible directions of collision avoidance, EA defines risk as the minimum magnitude of a constant relative acceleration vector required to alter the relative motion and make the interaction collision-free. Using interaction data from five open datasets and more than 600 real crashes, we derive percentile-based warning thresholds and show that EA provides the earliest statistically significant warning across all thresholds. Moreover, EA provides the best discrimination of eventual collision outcomes and improves information retention by 54.2-241.4% over all compared baselines. Adding EA to existing methods yields 17.5-95.5 times more information gain than adding existing methods to EA, indicating that EA captures much of the outcome-relevant information in existing methods while contributing substantial additional nonredundant information. Overall, EA better captures the structure of collision risk and provides a foundation for next-generation autonomous driving systems.

Agent^2 RL-Bench: Can LLM Agents Engineer Agentic RL Post-Training?

We introduce Agent^2 RL-Bench, a benchmark for evaluating agentic RL post-training -- whether LLM agents can autonomously design, implement, and run complete RL pipelines that improve foundation models. This capability is important because RL post-training increasingly drives model alignment and specialization, yet existing benchmarks remain largely static: supervised fine-tuning alone yields strong results, leaving interactive RL engineering untested. Agent^2 RL-Bench addresses this with six tasks across three levels -- from static rule-based training to closed-loop online RL with trajectory collection -- each adding a structural requirement that prior levels do not impose. The benchmark provides isolated workspaces with a grading API, runtime instrumentation that records every submission and code revision, and automated post-hoc analysis that generates structured run reports, enabling the first automated diagnostic of agent-driven post-training behavior. Across multiple agent stacks spanning five agent systems and six driver LLMs, we find that agents achieve striking interactive gains -- on ALFWorld, an RL-only agent improves from 5.97 to 93.28 via SFT warm-up and GRPO with online rollouts -- yet make only marginal progress on others (DeepSearchQA: +2.75 within evaluation noise), and that driver choice has a large effect on interactive tasks -- within the same scaffold, switching drivers changes interactive improvement from near-zero to +78pp. More broadly, the benchmark reveals that supervised pipelines dominate agent-driven post-training under fixed budgets, with online RL succeeding as the final best route only on ALFWorld. Code is available at https://github.com/microsoft/RD-Agent/tree/main/rdagent/scenarios/rl/autorl_bench.

Learning Additively Compositional Latent Actions for Embodied AI

Latent action learning infers pseudo-action labels from visual transitions, providing an approach to leverage internet-scale video for embodied AI. However, most methods learn latent actions without structural priors that encode the additive, compositional structure of physical motion. As a result, latents often entangle irrelevant scene details or information about future observations with true state changes and miscalibrate motion magnitude. We introduce Additively Compositional Latent Action Model (AC-LAM), which enforces scene-wise additive composition structure over short horizons on the latent action space. These AC constraints encourage simple algebraic structure in the latent action space~(identity, inverse, cycle consistency) and suppress information that does not compose additively. Empirically, AC-LAM learns more structured, motion-specific, and displacement-calibrated latent actions and provides stronger supervision for downstream policy learning, outperforming state-of-the-art LAMs across simulated and real-world tabletop tasks.

LiveMathematicianBench: A Live Benchmark for Mathematician-Level Reasoning with Proof Sketches

Mathematical reasoning is a hallmark of human intelligence, and whether large language models (LLMs) can meaningfully perform it remains a central question in artificial intelligence and cognitive science. As LLMs are increasingly integrated into scientific workflows, rigorous evaluation of their mathematical capabilities becomes a practical necessity. Existing benchmarks are limited by synthetic settings and data contamination. We present LiveMathematicianBench, a dynamic multiple-choice benchmark for research-level mathematical reasoning built from recent arXiv papers published after model training cutoffs. By grounding evaluation in newly published theorems, it provides a realistic testbed beyond memorized patterns. The benchmark introduces a thirteen-category logical taxonomy of theorem types (e.g., implication, equivalence, existence, uniqueness), enabling fine-grained evaluation across reasoning forms. It employs a proof-sketch-guided distractor pipeline that uses high-level proof strategies to construct plausible but invalid answer choices reflecting misleading proof directions, increasing sensitivity to genuine understanding over surface-level matching. We also introduce a substitution-resistant mechanism to distinguish answer recognition from substantive reasoning. Evaluation shows the benchmark is far from saturated: Gemini-3.1-pro-preview, the best model, achieves only 43.5%. Under substitution-resistant evaluation, accuracy drops sharply: GPT-5.4 scores highest at 30.6%, while Gemini-3.1-pro-preview falls to 17.6%, below the 20% random baseline. A dual-mode protocol reveals that proof-sketch access yields consistent accuracy gains, suggesting models can leverage high-level proof strategies for reasoning. Overall, LiveMathematicianBench offers a scalable, contamination-resistant testbed for studying research-level mathematical reasoning in LLMs.

Cerebra: A Multidisciplinary AI Board for Multimodal Dementia Characterization and Risk Assessment

Modern clinical practice increasingly depends on reasoning over heterogeneous, evolving, and incomplete patient data. Although recent advances in multimodal foundation models have improved performance on various clinical tasks, most existing models remain static, opaque, and poorly aligned with real-world clinical workflows. We present Cerebra, an interactive multi-agent AI team that coordinates specialized agents for EHR, clinical notes, and medical imaging analysis. These outputs are synthesized into a clinician-facing dashboard that combines visual analytics with a conversational interface, enabling clinicians to interrogate predictions and contextualize risk at the point of care. Cerebra supports privacy-preserving deployment by operating on structured representations and remains robust when modalities are incomplete. We evaluated Cerebra using a massive multi-institutional dataset spanning 3 million patients from four independent healthcare systems. Cerebra consistently outperformed both state-of-the-art single-modality models and large multimodal language model baselines. In dementia risk prediction, it achieved AUROCs up to 0.80, compared with 0.74 for the strongest single-modality model and 0.68 for language model baselines. For dementia diagnosis, it achieved an AUROC of 0.86, and for survival prediction, a C-index of 0.81. In a reader study with experienced physicians, Cerebra significantly improved expert performance, increasing accuracy by 17.5 percentage points in prospective dementia risk estimation. These results demonstrate Cerebra's potential for interpretable, robust decision support in clinical care.