The paper tackles two main issues in multi-subject video generation—uncontrollable fidelity strength and semantic drift—by studying Diffusion Transformers (DiTs). It discovers that certain attention blocks naturally create an Intrinsic Spatial Grounding Map (ISGM) that accurately locates reference subjects. Leveraging this insight, the authors introduce Dual-phase Intrinsic Attention Leveraging (DIAL), which uses ISGM during low-noise stages to control fidelity strength without retraining and during high-noise stages to generate preference pairs for reinforcement learning, thereby anchoring attention and reducing semantic drift. Experiments on the OpenS2V-Eval benchmark show that DIAL outperforms baseline models, improving identity consistency and enabling controllable fidelity strength.
The paper investigates how to best combine synthetic and real historical Manchu word images for low‑resource OCR. Using 60,000 synthetic and 20,306 real images, the authors compare three pretrained vision‑language models and a compact CRNN across synthetic‑only, real‑only, joint, and sequential training regimes. Adding real data boosts word accuracy to 95–96%, and ensembling the best recognizers raises it to 98.27% without extra training.
FreeFlow is a hierarchical transformer for optical flow estimation that eliminates traditional flow-specific inductive biases such as correlation volumes and feature warping. It relies on a single feed-forward encoder–decoder architecture that incorporates window attention, shifted-window attention, and reduced-resolution global attention. This design allows the model to scale with capacity while achieving state‑of‑the‑art accuracy on Sintel, KITTI‑2015, and Spring benchmarks, all while remaining memory efficient at 1080p inference.
The study investigates whether cross‑lingual clinical annotation can be treated as a constrained text‑generation task that preserves the original text while inserting entity tags. Using a workflow that embeds tags directly into immutable target‑language text and then validates them deterministically, the authors compare this approach to supervised candidate‑span projection and hybrid ML‑LLM refinement across six languages. Results show that direct LLM projection, particularly with GLM 5.2 and Gemma4:31B, achieves the highest strict F1 scores, surpassing previous state‑of‑the‑art by up to 0.15 and producing over 55,000 grounded mentions with accurate offsets.
The paper introduces TASCO, a test‑time adaptation method that improves Large Language Model reasoning by optimizing stability‑aware confidence. It keeps the LLM frozen and uses a lightweight task‑level prefix, applying Random Perturbation for distributional stability and Sharpness‑Aware Perturbation for worst‑case sensitivity. Experiments show enhanced reasoning accuracy and token efficiency across various LLMs and benchmarks, while preserving stable confidence without over‑concentrating predictions.
Cloudera and Mistral have partnered to deliver specialized, sovereign AI intelligence for enterprise data. The collaboration aims to enable regulated industries to innovate while maintaining control over their data. By combining Cloudera’s data platform with Mistral’s AI capabilities, the partnership focuses on secure, compliant solutions for enterprise use.
The paper introduces DF-LLM, a Dynamic Fusion Large Language Model designed for traffic flow prediction. It combines a spatiotemporal embedding module, a fusion module that uses graph convolution to capture spatial topology and dynamic dependencies, and an LLM backbone with differentiated parameter adaptation and context aggregation attention. Experiments on four datasets show that DF-LLM outperforms existing methods in predictive accuracy.
CEM‑TUDASR is a lightweight, unsupervised Transformer-based super‑resolution framework designed to enhance low‑resolution images from Wireless Capsule Endoscopy (WCE). It uses a domain‑adaptive degradation network to generate realistic WCE‑like low‑resolution images from high‑resolution conventional endoscopy data, enabling effective unpaired learning. The model incorporates Deep Attention Blocks and a Fusion Attention Block to capture both global context and fine local details, achieving superior performance on WCE datasets and demonstrating cross‑domain adaptability to retinal images, all while keeping the parameter count and computational load low.
The paper introduces LogiMed‑RoB, a benchmark that tests large language models (LLMs) on hierarchical logical consistency in medical risk‑of‑bias assessments, using 860 randomized controlled trials and 14,820 queries based on Cochrane RoB 2.0 expert logic. It evaluates models across four dimensions—Atomic Consistency, Domain Consistency, Aggregation Consistency, and Evidential Faithfulness—revealing a severe Error Compounding Effect where high atomic accuracy does not translate to end‑to‑end consistency. Experiments on ten state‑of‑the‑art LLMs show that even top performers can collapse to 45.13% overall consistency, with some models nearly failing entirely, and that many models struggle to deduce correct outcomes from retrieved evidence.
OpenAI introduces ChatGPT for Financial Services, a specialized version that integrates built‑in financial data and GPT‑6 Astra. The tool is designed for research, modeling, and creating client‑ready materials within the financial sector.
The paper evaluates NVIDIA’s Max‑Q inference profile on a disaggregated B200 GPU system for large language model (LLM) serving, finding modest gains (+8.6% tokens/J) but increased latency (+5.2%). It proposes a phase‑decoupled, model‑calibrated power controller that sets a latency‑guaranteed SM‑clock window for prefill and a calibrated power cap for decode, achieving a 20.4% tokens/J improvement with only a 3.5% latency increase on an 8‑node Qwen3‑Coder‑480B deployment. The approach outperforms vendor profiles on both energy and latency, and demonstrates significant long‑term electricity savings in MoE‑based serving.
The paper introduces a new corpus of 1,101 human simultaneous interpreting segments, each scored on meaning transfer, delivery quality, and perceived latency. It demonstrates that standard large‑language‑model prompting and scalar supervision fail to align with these rubric dimensions, producing near‑zero correlation with human ratings. By adding dual regression heads to a LoRA‑adapted COMET‑KIWI encoder, the authors achieve modest Pearson correlations (0.388 for meaning transfer and 0.301 for delivery quality) on a held‑out test set, improving over the frozen baseline.
Large language models used as judges for social bias are affected by noisy text, such as typos and broken punctuation. In experiments with 3,822 stereotype-related responses, noise more often turns neutral judgments into biased ones than the reverse, with up to a 120‑fold difference. The effect is strongest at mild realistic noise levels and leads to systematic overestimation of bias, especially in fairness‑critical categories.
The paper introduces belief‑shift branching, a method for placing forks in tree‑structured reinforcement learning rollouts by identifying points where a model’s answer belief changes most. Unlike traditional structural or entropy‑based approaches, belief‑shift uses a probe, logit‑lens depth profile, or learned activation direction to locate pivots in the value curve, incurring minimal computational overhead. Experiments across multiple models and benchmarks show that belief‑shift forking consistently outperforms baseline methods, yielding significant gains in mathematics and code tasks.
The paper introduces a token dependency selection strategy for Transformer-based time series forecasting. By jointly applying an attention entropy constraint and a prediction error constraint, the method identifies fewer but more critical inter-token dependencies, reducing the influence of redundant dependencies that can hurt generalization. Experiments on multiple datasets show that this approach improves forecasting performance across various Transformer models.
By Jianqi Zhang, Yuchan Liu, Zeen Song, Yuefei Li, Fanjiang Xu
The paper establishes rigorous trainability results for multi-headed attention layers and Low Rank Adaptation (LoRA) models under stochastic training methods. By proving that the empirical regression loss induces a Poincaré inequality with constants independent of data dimension for LoRA and independent of head dimensions for multi-head attention, the authors show that a stochastic differential equation mimicking SGD converges to the loss minima. These results hold without assumptions on data or model size, providing the first theoretical guarantees for training such architectures.
By Zhengkai Sun, Dibyakanti Kumar, Alejandro F Frangi, Anirbit Mukherjee, Mingfei Sun
The paper reports a six‑month, population‑scale measurement of autonomous language‑model trading agents operating in two production fleets: DX Terminal Pro, with 3,505 user‑funded vaults trading real ETH in Base memecoin markets, and the DXAP live alpha fleet, with 500–599 user‑created agents trading Hyperliquid perpetuals. Across roughly 7.5 million single‑model invocations and 231,638 multi‑tool turns, the study finds that operating layer design, risk sliders, and leaderboard boundaries drive behavior more than strategy text; agents are volatility‑blind in sizing, capture little upside, and show no directional edge compared to a retail benchmark. The analysis includes regression discontinuity, permutation nulls, and a 17‑rule methodology canon to validate the findings.
By T. J. Barton, Chris Constantakis, Patti Hauseman, Annie Mous, Alaska Hoffman, Brian Bergeron, Hunter Goodreau
TreeThink is an open‑source Python library that provides modular, fully asynchronous tree search for neural theorem proving. It integrates established tree‑search methods with vLLM inference pipelines and supports a range of node evaluation techniques, from lightweight heuristics to neural evaluators. The library connects directly to the REPL servers of Lean 4, Rocq, and Isabelle/HOL, enabling real‑time verification and proof‑state extraction, and it has been evaluated on miniF2F and MATH500, achieving up to an 8.0× wall‑clock speedup from asynchronous execution.
By Burak S. Akbudak, Zeynel A. Ulu\c{s}an, Can S. Erer, G\"ozde G\"ul \c{S}ahin
The paper investigates whether large language model (LLM) chatbots can emulate human legal judgments of reasonableness. By comparing responses from 26 LLMs to those of human participants across 25 legal scenarios, the study finds that chatbots generally track human answers but tend to produce more homogeneous, government‑ and corporation‑friendly responses and align more closely with white, male, older, and more educated respondents. The authors note that these patterns warrant further systematic research.
By Nirav Patel, Emily Wenger, Christopher Buccafusco
TTGBench is a new benchmark for temporal graph learning that evaluates both structural evolution and semantic drift in text‑attributed graphs. It includes six real‑world, text‑rich datasets with dual volatility and supports multi‑class and multi‑label temporal node classification, addressing gaps left by existing benchmarks. A comprehensive evaluation of 17 state‑of‑the‑art methods shows a clear divide: TGNNs excel at structural prediction but struggle with semantic tracking, while LLM‑based models perform better on semantic tasks but lag in structural prediction.
By Longfei Ma, Zemin Liu, Fei Wu