Explores unsupervised visual pretraining on visually rich documents to improve language-model intelligence; shows visual-pretrained models outperform text-only counterparts on the same corpora. Key aspects: direct use of images/layouts (no OCR-only pipeline), scalable across backbones and benchmarks.
Proposes Riemannian Isometric Policy Optimization (RIPO) to fix exploration collapse in PPO-style RL for LLMs by aligning policy updates with the policy manifold's Riemannian geometry, improving exploration–exploitation balance and optimization stability across competition benchmarks.
Builds structured knowledge graphs for retrieval-augmented generation via a multi-step GraphRAG pipeline that separates extraction from consolidation. Key features include typed two-stage extraction, DBSCAN-backed deduplication, LLM summarization, Leiden community detection, and a compact 7B extractor model (Meno-Lite-0.1).
Specialized LLM for clinical workflows trained via a human-gated self-evolution loop to improve patient consultation, multimodal clinical reasoning, interactive diagnosis, and EHR tool use. Iteratively refines targeted synthetic and curated data based on benchmark failures to raise specific capabilities without broad regressions.
Enables RL post-training with million-token prompts under a fixed GPU budget by evaluating shared prompt state without autograd, retaining only minimal model state, and replaying short response branches; instantiated as GRPO and demonstrated on Qwen3.6-27B and GLM-5.2 up to multi-million token execution.
Expands a Transformer’s residual stream into many parallel streams and introduces xHC to scale Hyper-Connections beyond N=4. Combines temporal feature augmentation with sparse residual updates (update k=4 of N=16) and xHC-Flash memory optimizations to raise downstream scores while cutting effective compute and memory traffic.
Quantifies active visual observation in multimodal LLMs with ActiveVision, a 17-task benchmark that forces repeated perception rather than one-shot description. Finds frontier MLLMs fail badly (top model 10.6% vs humans 96.1%) and that model-generated vision code does not close the gap.
A looped-Transformer LLM series using Mixture-of-Experts (20B with 2B active; 6B with 0.6B active) that trades extra pretraining compute for repeated looping. Shows superior compute-efficiency versus matched-compute vanilla baselines and attains gold-medal performance on 2025 IMO and IPhO after a post-training pipeline.
AI Video Papers2026
Predicts variable-cardinality sets of evidence intervals in videos to temporally ground queries using multimodal large language models. Combines caption-derived multi-span supervision, a temporal Wasserstein matching-free reward, and temporal IoU, yielding strong mIoU gains across multiple benchmarks.
AI Agent Papers2026
Prunes tool-output lines inside a coding LLM agent by turning the agent's own internal representations into per-line keep-or-prune labels. Implements a small classification head plus a length-aware embedding, saving up to 39% of tokens across benchmarks while preserving task quality.
Studies train-time knowledge injection via hypernetworks that generate fixed LoRA adapters from large fact corpora, empirically characterizing power-law scaling across hypernetwork depth, width, and target model size and reporting improved OOD generalization.
Performs full-parameter post-training of trillion-parameter MoE DeepSeek-V4 models on an Ascend NPU SuperPOD, using a hierarchical optimization of model parallelism, communication orchestration, and kernel execution to increase Model FLOPs Utilization. Also builds CPT/SFT pipelines with solver-verified synthetic data for Operations Research, reporting strong zero-shot Pass@1 results.