L. Yao

Matthew Brady Neeley, Jorge Botas, Johnathan Jia et al.

Pediatric electronic health records capture developmentally structured clinical trajectories, yet their potential for generative healthcare foundation models remains largely unexplored. Here we present TEDDY (Temporal Event Decoder for Disease in Youth), a 1.84-million-parameter decoder transformer trained on approximately 73 million ICD-10 diagnoses from 1.6 million children at a single pediatric institution. TEDDY models longitudinal diagnosis trajectories and visit timing. Predictions were made before visit codes were revealed, limited to first occurrences, and evaluated against sex- and age-matched controls. Across 797 disease-onset prediction tasks spanning 16 ICD-10 chapters, TEDDY achieved a median AUC of 72.0%, outperforming same-data DenseNet (50.0%), CNN (57.2%), RNN (60.1%), and LSTM (62.7%) baselines on 96-99% of tasks. Performance held across sex and age and was strongest among lower-prevalence diagnoses; 202 of the 225 rarest conditions (90%) had 95% confidence intervals above chance. Predictive signal remained detectable more than two years before first recorded diagnosis, with median AUCs of 59.7% in the unrestricted analysis and 64.4% in a fixed-cohort sensitivity analysis. In asthma and attention-deficit/hyperactivity disorder benchmarks, AUCs were 79.3% and 84.7%, compared with 62.7% and 71.7% for the strongest comparators, including a general-purpose language model three orders of magnitude larger. Visit-timing predictions had a 3.0-day mean absolute restricted mean survival-time error over 365 days, although median and long-tail return intervals remained miscalibrated. Together, these results establish pediatric diagnostic histories as a substrate for compact generative models supporting broad, rare-disease, and long-horizon risk forecasting without population-scale data or billion-parameter models.

Xiangdong Zhang, Xiaohan Qin, Sunan Zou et al.

Hyper-Connections (HC) expand the residual stream of Transformers into $N$ parallel streams, providing a form of memory scaling beyond model width and depth. Manifold-Constrained HC (mHC) stabilizes this formulation at scale. The large gains from $N{=}1$ to $N{=}4$ suggest residual-stream expansion as a promising scaling axis. However, existing HC-family methods typically stop at $N{=}4$. Our experiments reveal why: scaling mHC beyond this point yields diminishing performance gains and rapidly increasing training cost. We attribute this limitation to two bottlenecks: insufficient write-back information for an expanding number of streams and residual-mixing generation whose cost scales cubically with $N$. To address both bottlenecks, we propose xHC (Expanded Hyper-Connections), the first HC-family method to achieve meaningful expansion beyond $N{=}4$. xHC combines temporal feature augmentation for richer write-back with a sparse residual-stream architecture that updates only $k=4$ of the $N=16$ streams while retaining dense access to the full residual state. Across 18B and 28B MoE models, xHC delivers strong and consistent downstream improvements. On an 18B MoE model, xHC improves the average downstream score by 4.0 points over mHC, while adding only modest training FLOPs over the vanilla baseline. Scaling-law experiments show that the vanilla and mHC require $1.50\times$ and $1.19\times$ the compute of xHC, respectively, to reach the same loss. Practical large-$N$ training also requires controlling memory traffic from the expanded residual state. We therefore introduce xHC-Flash, which reduces the per-sublayer memory traffic from $73.5C$ to $40C$, comparable to the $34C$ required by mHC at $N{=}4$, while retaining the gains of full xHC. Together, xHC and xHC-Flash make large-$N$ residual-stream expansion effective and practical for LLM pre-training.

Hao Cong, Huizu Lin, Zihan Wang et al.

Large language model (LLM)-based agentic recommender systems show promise in modeling user preferences through natural-language reasoning, yet they remain limited by text-centric inputs and coarse-grained memory updates, making agents prone to missing visual evidence, semantic noise, and preference drift. To address these limitations, we propose MMEACR, a Multimodal Memory-Enhanced Agent Collaboration framework for recommendation. MMEACR introduces a dual-track memory architecture that separates interpretable agent reasoning from fine-grained multimodal matching. In the reasoning track, collaborative User and Item Memory Agents maintain persistent multimodal memories and update them through an attribute-guided reinforcement-and-reflection mechanism. In the matching track, a decoupled multi-modal embedding memory is built from raw interaction narratives and item images to preserve detailed cross-modal signals beyond structured memory updates. The two tracks are integrated through weighted Reciprocal Rank Fusion to produce robust and interpretable rankings. Experiments on three real-world domains show that MMEACR achieves strong overall performance against competitive LLM-based and agent-based baselines, with notable gains in visually grounded recommendation scenarios.

Lin Yao

Diffusion large language models (dLLMs) commit multiple tokens per denoising step by decoding each selected position independently from the shared context; when those positions are dependent, the resulting factorization error is captured by conditional total correlation, which confidence-based selection cannot observe from marginals alone. We propose CoCommit, a marker-gated coordination pass that briefly defers commitment: after the usual bundle selection, a learned marker announces the commit set and the backbone's last-$n$ layers are re-applied so marked positions coordinate -- approximating joint-mode decoding -- before greedy argmax writes tokens. The method reuses existing weights with one extra partial forward pass and no auxiliary model. On LLaDA2.1-mini with LoRA adapters and matched greedy inference, joint commitment improves accuracy on all six benchmarks we evaluate, with the largest gains on reasoning and exact-answer tasks.

Lin Yao

Token-to-token (T2T) editing lets LLaDA2.1 revise committed tokens during block-diffusion decoding. The released recipe trains this editor on random vocabulary corruptions, but at inference the editor sees the model's own fluent, high-confidence draft errors instead. We study this training-inference mismatch and propose self-generated T2T, which performs a no-gradient draft pass, fills masked positions with predicted tokens, and supervises recovery in a second pass under these self-generated corruptions. We implement the update as a short LoRA continued-pretraining pass on LLaDA2.1-mini and evaluate on several benchmarks under the official Q-Mode T2T procedure with unchanged inference parameters. The method generally improves accuracy while reducing T2T edit intensity, mitigating failure modes such as final-digit transcription errors after otherwise correct reasoning and excessive self-correction before short factual answers.

Liuyang Yao, Zhouyu Li, Junguang He et al.

AI systems are increasingly deployed for credit assessment and investment advisory in global financial markets, yet the integrity of their inference pipelines remains insufficiently addressed by existing regulatory frameworks. This paper identifies and empirically validates an invisible manipulation channel operating at the sampling layer of LLM inference--a vulnerability that allows adversaries to systematically bias AI-generated financial opinions while preserving full compliance with output-based audit mechanisms, including statistical watermarking. We show that this inference-stage manipulation is statistically hard to detect: the Kullback-Leibler divergence between manipulated and normal output distributions can be made arbitrarily small, so that any output-based detection scheme requires impractically large sample sizes to achieve reliable detection power. Empirical experiments across credit rating and investment advisory scenarios show that directional bias keywords can be amplified by 1.8-1.9x under stealth-preserving (aware) manipulation while triggering zero of six black-box detectors and preserving watermark integrity. The vulnerability generalizes across three mainstream watermarking schemes and three heterogeneous model architectures, establishing it as a systemic financial infrastructure risk. Software-based defenses including cryptographically secure pseudorandom number generators are entirely ineffective, while QRNG combined with TEE hardware isolation achieves 100% attack blocking--reducing the target rate to the natural baseline--by replacing the predictable hash key with quantum-derived entropy that renders all pre-computed manipulation targets invalid. We propose four regulatory amendments centered on mandatory QRNG certification for high-risk financial AI systems under NIST SP 800-90B, inference-layer supply chain audits, and output provenance mechanisms.