TurnSight: Turn-Level Hindsight Self-Distillation for Tool-Integrated Reasoning

Tool-Integrated Reasoning (TIR) enables LLMs to solve complex tasks through iterative tool interactions. However, existing reinforcement learning methods often rely on trajectory-level supervision, limiting fine-grained credit assignment in long-horizon TIR scenarios. On-policy self-distillation offers denser signals through teacher branches with privileged context, but existing approaches typically derive such context from ground-truth answers or retrieved skills, which may not reflect the states actually visited by the agent. Moreover, token-level supervision fails to capture the turn-level structure of tool interactions. To address this, we propose TurnSight, a turn-level hindsight self-distillation framework that derives supervision directly from execution-conditioned hindsight. It then constructs multiple hindsight views with different lookahead horizons and selects reliable supervision through cross-horizon directional agreement. Finally, the selected hindsight signal is normalized across sibling rollouts and used to adaptively modulate RL advantages while preserving their original optimization direction. Extensive experiments on three benchmarks demonstrate the effectiveness of TurnSight. Our codes are available at https://github.com/quchangle1/TurnSight.

STAMP: Provenance-Guided Credit Assignment for Deep Search Agents

Reinforcement learning for deep-search agents has largely focused on trajectory-level scoring -- outcome correctness, citation-aware rewards, and evidence coverage. Yet the actions that expose supporting documents receive no targeted credit, a gap we call the reward-credit mismatch. We propose STAMP, in which a reference-based verifier judges whether each cited document supports an entity or relation in a training-time evidence graph, and first-exposure attribution traces each supported citation back to the action that first surfaced it. This step credit is injected through sign-preserving advantage modulation, which redistributes advantage across steps without changing the trajectory-level reward or the relative ranking of trajectories within each group. On BrowseComp, BrowseComp-ZH, and xbench-DS, STAMP improves the GRPO baseline by +2.0/+5.5/+3.0 points under matched SFT initialization, training data, and search tools, and composes with both outcome-only and citation-rubric base rewards. Component ablations confirm that the provenance-based credit signal and the sign-preserving advantage modulation each contribute to the gains.

SCOPE-RL: Optimizing Reasoning Paths Before and After Success

Reinforcement learning with verifiable rewards (RLVR) optimizes LLMs using sparse verifiable final-answer rewards. This sparse anchor reliably verifies whether a trajectory succeeds but provides no direct feedback on the reasoning path that produced it. Before success, prerequisite progress on hard problems receives no reward signal; after success, outcome rewards cannot distinguish well-organized correct trajectories from redundant or locally flawed ones. We introduce SCOPE-RL (Scaffolded Chain Optimization with Process Efficiency), a two-stage framework that densifies this anchor while retaining the GRPO update: Adaptive Scaffolded RL adds prefix-decomposed verifiable rewards on answer-hidden sub-question chains before success, and Quality-Aware Process RL applies correctness-gated process-shape rewards to refine correct trajectories after success. An expert-validated Step-Quality Evaluation Protocol evaluates useful-step density, error localization, and token efficiency beyond final-answer accuracy. On Qwen3-8B-Instruct trained on DAPO-Math and Big-Math, SCOPE-RL improves average accuracy by up to 11.2 pp and reduces reasoning tokens by up to 27.1% over outcome-only GRPO; the gains hold under GSPO and on Qwen3-0.6B-Instruct, indicating that reward-signal densification is complementary to policy-update-level RLVR advances. Code and data are available at https://github.com/tokencraft-lab/SCOPE-RL.

Automatic Echocardiography Segmentation via Transition Probability Correlation for Stable Semantic Extraction

While echocardiography is essential for cardiovascular diagnosis, inherent speckle noise and low signal-to-noise ratio often lead to ambiguous semantic features and fragmented boundaries. These limitations significantly hinder the segmentation accuracy of deep learning models in complex clinical cases. Moreover, temporal motion of the heart plays a critical role in recognizing anatomical structures. To address these challenges, we designed a STLSF module which comprises a window-matching-based semantic correction component and a semantics-guided texture enhancement component. By leveraging local transition probability correlations to correct semantics and employing semantics-guided texture enhancement, the STLSF module effectively mitigates texture instability and ambiguous semantic interpretations caused by disadvantaged echocardiography quality. Additionally, to facilitate the encoder's adaptation to the intrinsic priors of ultrasound-specific imaging patterns, we propose a frequency-aware denoising pre-training method. The entire work builds a convolution-based network with locality inductive bias and long-range dependencies. Extensive experiments confirm our SOTA performance, achieving 93.87\% Dice on CAMUS and 92.62\% on EchoNet-Dynamic, with respective HD95 values of 3.29mm and 2.73mm.

LayersReg: A Layer-by-Layer Progressive Regressor for Reliable Intraoperative 3D/2D Registration

3D/2D registration serves as a cornerstone technique in surgical navigation. Traditional iterative optimization algorithms suffer from low efficiency and high failure rates in intraoperative settings. Deep learning-based methods reformulate registration from iterative optimization to a regression problem that maps image appearance features to spatial pose, typically achieving improved real-time performance and accuracy. However, such learnable methods are confined to memory-driven retrieval of specific pose features rather than understanding the task of image alignment itself, which limits their generalization in complex scenarios. We propose LayersReg, a pioneering regression paradigm that endows the model with 3D anatomical awareness and searches for the correct pose in a progressive, layer-by-layer manner. Inspired by the iterative pose-searching optimization criterion of classical registration, LayersReg searches for correlations between the moving and fixed images in feature space, capturing the trend of pixel flow and thereby converging iteratively toward the correct spatial pose transformation. We further design a coupling of node-wise regression with the progressive registration framework to enhance the model's perception of spatial pose changes. Experimental results demonstrate that under large offsets and multimodality conditions, LayersReg achieves high accuracy on both X-ray/CT registration (0.68°, 1.41 mm) and slice localization (0.73°, 1.55 mm) tasks, outperforming existing state-of-the-art methods while meeting the intraoperative demands for precision and real-time capability.

EndPrompt: Efficient Long-Context Extension via Terminal Anchoring

Extending the context window of large language models typically requires training on sequences at the target length, incurring quadratic memory and computational costs that make long-context adaptation expensive and difficult to reproduce. We propose EndPrompt, a method that achieves effective context extension using only short training sequences. The core insight is that exposing a model to long-range relative positional distances does not require constructing full-length inputs: we preserve the original short context as an intact first segment and append a brief terminal prompt as a second segment, assigning it positional indices near the target context length. This two-segment construction introduces both local and long-range relative distances within a short physical sequence while maintaining the semantic continuity of the training text--a property absent in chunk-based simulation approaches that split contiguous context. We provide a theoretical analysis grounded in Rotary Position Embedding and the Bernstein inequality, showing that position interpolation induces a rigorous smoothness constraint over the attention function, with shared Transformer parameters further suppressing unstable extrapolation to unobserved intermediate distances. Applied to LLaMA-family models extending the context window from 8K to 64K, EndPrompt achieves an average RULER score of 76.03 and the highest average on LongBench, surpassing LCEG (72.24), LongLoRA (72.95), and full-length fine-tuning (69.23) while requiring substantially less computation. These results demonstrate that long-context generalization can be induced from sparse positional supervision, challenging the prevailing assumption that dense long-sequence training is necessary for reliable context-window extension. The code is available at https://github.com/clx1415926/EndPrompt.

Not All Preferences Are Created Equal: Stability-Aware and Gradient-Efficient Alignment for Reasoning Models

Preference-based alignment is pivotal for training large reasoning models; however, standard methods like Direct Preference Optimization (DPO) typically treat all preference pairs uniformly, overlooking the evolving utility of training instances. This static approach often leads to inefficient or unstable optimization, as it wastes computation on trivial pairs with negligible gradients and suffers from noise induced by samples near uncertain decision boundaries. Facing these challenges, we propose SAGE (Stability-Aware Gradient Efficiency), a dynamic framework designed to enhance alignment reliability by maximizing the Signal-to-Noise Ratio of policy updates. Concretely, SAGE integrates a coarse-grained curriculum mechanism that refreshes candidate pools based on model competence with a fine-grained, stability-aware scoring function that prioritizes informative, confident errors while filtering out unstable samples. Experiments on multiple mathematical reasoning benchmarks demonstrate that SAGE significantly accelerates convergence and outperforms static baselines, highlighting the critical role of policy-aware, stability-conscious data selection in reasoning alignment.

Tracking Drift: Variation-Aware Entropy Scheduling for Non-Stationary Reinforcement Learning

Real-world reinforcement learning often faces environment drift, but most existing methods rely on static entropy coefficients/target entropy, causing over-exploration during stable periods and under-exploration after drift (thus slow recovery), and leaving unanswered the principled question of how exploration intensity should scale with drift magnitude. We prove that entropy scheduling under non-stationarity can be reduced to a one-dimensional, round-by-round trade-off, faster tracking of the optimal solution after drift vs. avoiding gratuitous randomness when the environment is stable, so exploration strength can be driven by measurable online drift signals. Building on this, we propose AES (Adaptive Entropy Scheduling), which adaptively adjusts the entropy coefficient/temperature online using observable drift proxies during training, requiring almost no structural changes and incurring minimal overhead. Across 4 algorithm variants, 12 tasks, and 4 drift modes, AES significantly reduces the fraction of performance degradation caused by drift and accelerates recovery after abrupt changes.

Enabling Ultra-Fast Cardiovascular Imaging Across Heterogeneous Clinical Environments with a Generalist Foundation Model and Multimodal Database

Multimodal cardiovascular magnetic resonance (CMR) imaging provides comprehensive and non-invasive insights into cardiovascular disease (CVD) diagnosis and underlying mechanisms. Despite decades of advancements, its widespread clinical adoption remains constrained by prolonged scan times and heterogeneity across medical environments. This underscores the urgent need for a generalist reconstruction foundation model for ultra-fast CMR imaging, one capable of adapting across diverse imaging scenarios and serving as the essential substrate for all downstream analyses. To enable this goal, we curate MMCMR-427K, the largest and most comprehensive multimodal CMR k-space database to date, comprising 427,465 multi-coil k-space data paired with structured metadata across 13 international centers, 12 CMR modalities, 15 scanners, and 17 CVD categories in populations across three continents. Building on this unprecedented resource, we introduce CardioMM, a generalist reconstruction foundation model capable of dynamically adapting to heterogeneous fast CMR imaging scenarios. CardioMM unifies semantic contextual understanding with physics-informed data consistency to deliver robust reconstructions across varied scanners, protocols, and patient presentations. Comprehensive evaluations demonstrate that CardioMM achieves state-of-the-art performance in the internal centers and exhibits strong zero-shot generalization to unseen external settings. Even at imaging acceleration up to 24x, CardioMM reliably preserves key cardiac phenotypes, quantitative myocardial biomarkers, and diagnostic image quality, enabling a substantial increase in CMR examination throughput without compromising clinical integrity. Together, our open-access MMCMR-427K database and CardioMM framework establish a scalable pathway toward high-throughput, high-quality, and clinically accessible cardiovascular imaging.

Towards Universal Learning-based Model for Cardiac Image Reconstruction: Summary of the CMRxRecon2024 Challenge

Cardiovascular magnetic resonance (CMR) offers diverse imaging contrasts for assessment of cardiac function and tissue characterization. However, acquiring each single CMR modality is often time-consuming, and comprehensive clinical protocols require multiple modalities with various sampling patterns, further extending the overall acquisition time and increasing susceptibility to motion artifacts. Existing deep learning-based reconstruction methods are often designed for specific acquisition parameters, which limits their ability to generalize across a variety of scan scenarios. As part of the CMRxRecon Series, the CMRxRecon2024 challenge provides diverse datasets encompassing multi-modality multi-view imaging with various sampling patterns, and a platform for the international community to develop and benchmark reconstruction solutions in two well-crafted tasks. Task 1 is a modality-universal setting, evaluating the out-of-distribution generalization of the reconstructed model, while Task 2 follows sampling-universal setting assessing the one-for-all adaptability of the universal model. Main contributions include providing the first and largest publicly available multi-modality, multi-view cardiac k-space dataset; developing a benchmarking platform that simulates clinical acceleration protocols, with a shared code library and tutorial for various k-t undersampling patterns and data processing; giving technical insights of enhanced data consistency based on physic-informed networks and adaptive prompt-learning embedding to be versatile to different clinical settings; additional finding on evaluation metrics to address the limitations of conventional ground-truth references in universal reconstruction tasks.