This paper introduces LLM-guided MAP-Elites evolution for optimizing medical decision pipelines, improving accuracy and safety metrics significantly across tasks.

Ivan Sviridov, Artem Oskin, Ivan Panin et al.

Proposes a Constrained Dominant Set (CDS) method for multimodal long-document QA, achieving 66.99 on VisDoMBench, surpassing previous SOTA by 37.1 points.

Ambuj Mehrish, Sebatiano Vascon

SES framework optimizes symbolic solutions directly from equations without training data, successfully recovering algebraic and differential equations' explicit expressions.

Sergei Garmaev, Vinay Sharma, Olga Fink

Proposes Code2LoRA, a hypernetwork-based method generating repository-specific adapters, achieving 63.8% cross-repo and 66.2% in-repo exact match on static tasks, and 60.3% on evolving codebases.

Liliana Hotsko, Yinxi Li, Yuntian Deng et al.

Introduces RP-Regret for adaptive opponents, with algorithms achieving sublinear regret and better equilibria in repeated games.

Mingyang Liu, Asuman Ozdaglar, Tiancheng Yu et al.

PAR3D introduces part-aware 3D multimodal large language models, significantly enhancing fine-grained scene understanding via the ScenePart dataset.

Shaohui Dai, Yansong Qu, You Shen et al.

Proposes Complexity-Balanced Diffusion Splitting (CBS), using Dirichlet energy and trajectory acceleration to estimate local complexity, improving synthesis quality by ~35%.

Noam Issachar, Dani Lischinski, Raanan Fattal

Proposes Astra framework combining RL-trained VLM policy with Bagel-based world simulator for imagination-driven spatial reasoning, improving MMSI-Bench accuracy from 45.1% to 49.5%.

Chenming Zhu, Jingli Lin, Yilin Long et al.

MLEvolve is a self-evolving multi-agent framework using LLMs for end-to-end machine learning algorithm discovery, achieving 65.3% medal rate within 12 hours.

Shangheng Du, Xiangchao Yan, Jinxin Shi et al.

Proposes Polynomial Weight Preconditioning (PC) layer to regulate singular-value spectrum, accelerating LLM pretraining; achieves 2× speedup on Llama-1B with no inference overhead.

Senmiao Wang, Tiantian Fang, Haoran Zhang et al.

Using large deviation principles, the paper analyzes the distribution of generalization errors among high-dimensional interpolating classifiers, revealing that good interpolators are exceedingly rare and that algorithmic solutions outperform most interpolators.

August Y. Chen, Ahmed El Alaoui

This study uses a multi-layer pre-registered ablation framework to evaluate whether Popperian procedural content in prompts genuinely improves code correctness, finding structure outweighs content effects.

Mehmet Iscan

Proposes MedReCo, an entity-aware vision-language framework with over 690,000 images for clinical case retrieval and change description.

Tengfei Zhang, Ziheng Zhao, Lisong Dai et al.

HomeWorld introduces a hierarchical, multimodal framework trained on 300K real floorplans, using LLMs and diffusion models to generate controllable, diverse, and realistic whole-home scenes.

Wenbo Li, Xiaoliang Ju, Zipeng Qin et al.

Proposes emergent language in multi-agent RL to study consciousness-related structures without prior biases, revealing self-referential communication and echo-mismatch circuits.

Zengqing Wu, Chuan Xiao

Proposes a knowledge refinement framework using automatic rule mining and ASP-based abductive reasoning, improving scene graph generation with +4-8% F1@50 across benchmarks.

Maëlic Neau, Salim Baloch, Jakob Suchan et al.

Proposes a structure-agnostic bias correction method (SADE) that achieves the optimal rate n^{-1/2} + δ^a_μ + (δ^s_μ)^2 for semiparametric estimation with black-box models.

Yihong Gu, Qishuo Yin, Tianxi Cai et al.

GraphDETR formulates subgraph detection as set prediction, achieving 91.2 AP on molecular datasets with graphs up to 1000 nodes and 50-node substructures.

Dexiong Chen, Till Hendrik Schulz, Karsten Borgwardt

GMBFormer integrates NDVI-guided global memory with Transformer for urban green-space extraction, achieving a mean IoU of 89.25%.

Hao Lei, Xi Cheng, Chenlu Shu et al.

Proposes PhaseLock, a training-free framework that extracts motion priors from 2-step inference, improving physical consistency by 6.2 points on average.

Woojung Han, Seil Kang, Youngjun Jun et al.