The paper proposes a new approach to large language model (LLM) reasoning that moves beyond naive repeated sampling. Instead of generating many independent solutions, it first samples problem‑specific concepts, hints, or strategies and conditions answer generation on them, producing a single trajectory of diverse concepts. A small concept generator is then trained via reinforcement learning to maximize downstream success, leading to significant improvements in pass@k on hard mathematical reasoning tasks compared to both naive sampling and concepts from larger untuned models, and the trained generator transfers to unseen answer generators, including those from different model families.
By Ismail Labiad, Matthieu Kowalski, Marc Schoenauer, R\'emi Munos, Julia Kempe
Agensh is a new multi‑agent harness that eliminates a central orchestrator by letting workers self‑organize through a continuous cooperation loop. The system uses a shared workspace, message interface, and shared context to coordinate tasks, verify results, and merge progress asynchronously. Experiments on ProgramBench and pandoc show that scaling from 1 to 1,024 agents improves test‑pass rates by up to 49% relative, demonstrating that agent count is a viable scaling dimension for complex tasks.
By Zhihao Zhan, Ting Song, Li Dong, Shaohan Huang, Jianxun Lian, Yan Xia, Furu Wei
Flash-dLLM is a training‑free inference acceleration framework that improves the speed and memory efficiency of Diffusion Large Language Models (dLLMs). It tackles GPU memory I/O bottlenecks by introducing an I/O‑aware fused KV‑cache kernel and then employs a draft‑and‑verify decoding strategy that uses the dLLM itself as both drafter and verifier. Experiments on mathematical reasoning and code‑generation tasks show Flash‑dLLM outperforms existing acceleration methods, achieving up to 11.0× speedups over the Elastic‑Cache baseline.
By Quan Nguyen-Tri, Mukul Ranjan, Zhiqiang Shen
The article surveys how large language models (LLMs) are being applied to mental health, outlining a three‑phase evolution: Phase I uses LLMs as passive information tools and pattern recognizers for assessment; Phase II employs them as empathetic conversationalists for stateless, in‑the‑moment interactions; Phase III aims to create longitudinal, personalized companions that act as stateful cognitive agents. It systematically reviews core technologies, agent architectures (Profile, Memory, Reasoning, Planning), and the datasets and benchmarks that support this progression, offering a coherent narrative and roadmap for future research. The survey also provides a curated resource list at https://github.com/Emo-gml/Awesome-Mental-Health-LLMs.
By He Hu, Yucheng Zhou, Qianning Wang, Yingjian Zou, Chiyuan Ma, Juzheng Si, Jianzhuang Liu, Zitong Yu, Laizhong Cui, Fei Ma, Qi Tian
WatchPoint is a simulated‑user system that generates and runs diagnostic scripts against a live web application, producing structured observations to guide coding model retries. Unlike prior methods that rely on screenshots or non‑executable metrics, WatchPoint operates on Web‑Bench—a benchmark of 50 multi‑file web projects with 1,000 sequential tasks verified by deterministic end‑to‑end tests. It recovers 57.6% of diagnosed tasks, matching a human tester’s 54.5% recovery rate, and identifies when such simulated feedback is beneficial or should be withheld.
By Guanqun Yang, Wei Yang, Xueqing Liu
The paper introduces DA-Cramming, a cost‑effective pretraining method that incorporates dependency agreement information into BERT‑style language models. It builds on the Cramming technique to enable training with a single GPU in a day, using a dual‑stage workflow and four submodels to embed chunk‑level dependency agreements. Experiments show that this approach outperforms prior methods on a range of tasks.
By Martin Kuo, Jianyi Zhang, Dongting Li, Yiran Chen
The paper presents a geometric framework for quantifying uncertainty in large language models (LLMs) at both the prompt and answer levels. By modeling a prompt-conditioned semantic distribution in answer embedding space and using archetypal analysis on multiple sampled answers, the method estimates distribution entropy for prompt-level uncertainty and atypicality for individual answer reliability. Experiments demonstrate comparable or superior performance to existing techniques on short-form QA datasets and notably better results on medical datasets where hallucinations pose critical risks.
By Edward Phillips, Sean Wu, Soheila Molaei, Danielle Belgrave, Anshul Thakur, David Clifton
CausalEmbed is an auto‑regressive method for generating compact multi‑vector embeddings in visual document retrieval. By using iterative margin loss during contrastive training, it reduces the number of visual tokens needed by 30‑155× while keeping performance competitive across different backbones and benchmarks. The approach offers efficient training, scalable test‑time performance, and a flexible scaling strategy for multi‑vector representations.
By Jiahao Huo, Yu Huang, Yibo Yan, Ye Pan, Kening Zheng, Wei-Chieh Huang, Yi Cao, Mingdong Ou, Philip S. Yu, Xuming Hu
The paper introduces ConRad, a reinforcement learning framework that fine‑tunes large vision‑language models to generate calibrated verbalized confidence estimates for radiology reports. ConRad offers both a single report‑level confidence score and a sentence‑level variant, trained with the GRPO algorithm and logarithmic scoring rewards to encourage truthful self‑assessment. Experiments show significant calibration improvements over existing methods, and clinical evaluation indicates that report‑level scores align well with clinicians’ judgments, enabling targeted review of low‑confidence statements.
By David Bani-Harouni, Chantal Pellegrini, Julian L\"uers, Su Hwan Kim, Markus Baalmann, Benedikt Wiestler, Rickmer Braren, Nassir Navab, Matthias Keicher
The paper introduces DeToxR, a reinforcement‑learning‑enhanced large language model designed to support decision making in acute toxicology cases. It fuses unstructured narratives from paramedics and patients with structured vital‑sign data to predict co‑ingested substances across 14 classes. In preliminary validation, DeToxR outperforms baseline models, achieving higher micro‑F1 and recall scores for poison identification.
By Nico Oberl\"ander, David Bani-Harouni, Tobias Zellner, Nassir Navab, Florian Eyer, Matthias Keicher
The paper investigates how machine unlearning for large language models (LLMs) can unintentionally erase related knowledge, even in distant domains. By analyzing the propagation of unlearning effects before any model updates, the authors discover a consistent decay pattern where collateral damage is strongest near the targeted forget set and diminishes with semantic distance but never fully disappears at domain boundaries. They propose a pre-unlearning prediction task—forget-set auditing—to identify potential collateral damage early, finding that interaction features between the forget set and evaluation set are the most predictive signals. This approach offers an early warning system for risky unlearning runs and guides the design of more reliable unlearning procedures.
By Bo Su, Ankit Shah, Thai Le
Hy‑MultiTurn is a Chinese benchmark designed to evaluate deep multi‑turn dialogue understanding over long interactions. It introduces six controlled evaluation modes—constraint memory, precise execution, constraint synthesis, object localization, action suppression, and reference resolution—across 209 tasks ranging from 12 to 76 turns, incorporating dialogue length, irrelevant distractions, and colloquial phrasing. Testing 22 state‑of‑the‑art models shows the benchmark is highly challenging, with even the best model meeting all criteria only 41.1% of the time and no model excelling in every mode.
By Eileen Ye, Jiawen Tao, Yaoming Li, Chenxu Liu, Wenhan Yu, Yaxin Fan, Xiaokun Yuan, Mengzhou Wu, Yanbing Jiang, Maxm Pan
RGSQ introduces a Riemannian geometry‑aware post‑training quantization method for large vision‑language models, treating quantization as a reconstruction problem under a Fisher‑Riemannian metric. It identifies modality‑specific sensitive directions via manifold mappings and applies geometry‑aligned rotations and whitening to steer low‑bit perturbations toward loss‑insensitive axes. Experiments on diverse VLM benchmarks show RGSQ delivers the best accuracy and stability in extremely low‑bit settings, outperforming existing VLM‑aware baselines by up to 5.9% and single‑modality methods by up to 8.6%.
By Zhiping Wu, Dongdong Ren, Yangchengyu Zhou, Zhengjie Zhang, Wenbin Li, Hongbing Pan, Yang Gao
The paper introduces STD, a hierarchical token pruning framework for Large Vision‑Language Models that aligns pruning strategies with the functional roles of different network stages. By using high‑frequency spectral analysis in shallow layers, Gaussian‑smoothed attention in intermediate layers, and a stability‑adaptive trigger in deep layers, STD preserves essential visual information while aggressively reducing token counts. Experiments demonstrate that STD outperforms existing pruning methods, achieving up to 94.4% token reduction and a 3.9× speed‑up on LLaVA‑NeXT‑7B.
By Shuo Zhang, Jintao Tong, Yixiong Zou, Yuhua Li, Ruixuan Li
Fysiverse-3D-Vision is a unified vision‑language‑geometry framework that reconstructs executable 3D scenes from a single image. It separates spatial layout reasoning from asset synthesis, using a shared representation where spatial reasoning and geometric reconstruction reinforce each other. The model employs a Transformer that integrates textual supervision, semantic visual cues, and geometric representations, and includes an object‑conditioned layout module to predict object translation, rotation, and scale while maintaining physical consistency through collision‑aware optimization.
By Dingkang Yang, Yizhou Liu, Wendong Cheng, Zizhi Chen, Shunli Wang, Yang Liu, Hongsheng Li, Lihua Zhang
Vision‑language models (VLMs) can lose accuracy when images are resized, even with minimal changes. The study shows that such small visual configuration changes—like tiling or token arrangement—cause more correctness flips across multiple checkpoints and benchmarks. Interestingly, in many cases the models still read the correct answer but fail to use it, and attention interventions reveal that configuration shifts weaken the use of readable information. By guiding models with field cues and their own transcriptions, the authors correct 97.2% of these errors.
By Dingyang Lin, Yingfeng Luo, Chenglong Wang, Chenwei Zhu, Anxiang Ma, Jingbo Zhu, Tong Xiao
The paper introduces RealFit, a virtual try‑on framework that addresses texture degradation and structural drift caused by symmetric bidirectional attention in Diffusion Transformers. By employing Unidirectional Information Flow to isolate garment conditions from stochastic noise and Decoupled Timestep Modulation to strengthen the conditional signal, RealFit achieves high‑fidelity garment rendering. The method also enables a time‑invariant condition branch and a conditional KV cache, cutting inference time by about 75%.
By Zishu Qin, Zhiyu Jin, Pipei Huang, Hao Zhou
The paper introduces BS‑O2G, a plug‑in that constructs a sparse prediction‑aware graph from decoded features, boxes, and class distributions to enable one‑to‑graph query collaboration in Detection Transformers. It uses One‑to‑Graph (O2G) calibration to propagate messages forward and Backward Sharing (BS) to route gradients backward, preserving the original one‑to‑one matcher and positive labels. Experiments on various DETR models, backbones, COCO, and CrowdHuman datasets demonstrate consistent performance gains, faster convergence, and minimal additional parameters or FLOPs.
By Wenxiao Fan, Jingling Fu, Luohang Liu, Lichen Ma, Yu He, Zhiyang Yu, Weishan Bi, Junshi Huang, Yan Li, Gu Simiu, Kan Li
The paper presents BDSLI, a hybrid CNN‑Transformer architecture designed for Bengali Sign Language Recognition (SLR). It introduces a new model combination for Bengali SLR, trained on a custom video dataset of 62 sign words (250 samples per class) and evaluated against baseline models. The hybrid model achieved high accuracy—99.58% training, 99.48% validation, and 98.65% test—before being deployed in a web application for real‑world testing.
By Abir Bin Yousuf, Muhammad Iqbal Hossain
The paper introduces a test‑time reinforcement learning framework for anomalous video understanding, addressing challenges such as unreliable pseudo‑labels, inadequate reward design, and collapsed group‑relative advantages. It proposes dual‑query consistency filtering, an entropy‑aware consensus reward, and a virtual negative anchor mechanism to improve sample reliability, reward quality, and policy‑gradient signals. Experiments on VAU‑Bench demonstrate significant performance gains, especially on the ECVA subset where accuracy rises from 75.81% to 90.00%.
By Huining Li, Yuxiang Duan, Jiyang Tan, Qian Li, MingCai Chen, Jian Zhang, Xingdong Sheng, Yuntao Du