Attention Residuals rethink Transformers; LLM agents autonomously discover GPU kernels and RL algorithms; AI safety alarms as models fail without adversarial prompts

Guangyu Chen, Yu Zhang, Jianlin Su, Weixin Xu et al. — Moonshot AI (Kimi)

Proposes Attention Residuals (AttnRes), replacing fixed-weight accumulation in residual connections with softmax attention over preceding layer outputs. Each layer selectively aggregates earlier representations with learned, input-dependent weights. Block AttnRes partitions layers into blocks to reduce overhead. Integrated into Kimi Linear (48B/3B activated parameters) trained on 1.4T tokens.

Terry Chen, Zhifan Ye, Bing Xu, Zihao Ye et al. — NVIDIA

Introduces Agentic Variation Operators (AVO) that replace fixed mutation/crossover heuristics of classical evolutionary search with autonomous LLM coding agents. Agents loop through proposing, repairing, critiquing, and verifying code edits while consulting lineage, domain knowledge, and execution feedback. Tested on attention mechanisms using NVIDIA Blackwell GPUs over 7 days.

Peng Xia, Jianwen Chen, Xinyu Yang, Haoqin Tu et al. — AIMING Lab, University of North Carolina at Chapel Hill

Introduces a continual meta-learning framework enabling LLM agents to jointly evolve a base LLM policy and a library of reusable behavioral skills. Features skill-driven fast adaptation from failure analysis and opportunistic policy optimization using cloud-based fine-tuning during inactive periods.

Hejun Dong, Junbo Niu, Bin Wang, Weijun Zeng et al. — Shanghai AI Laboratory

Reframes document OCR as inverse rendering rather than sequential text generation. Replaces autoregressive decoding with parallel diffusion denoising under visual conditioning. Introduces block-wise diffusion decoding and uncertainty-driven curriculum learning.

Jeonghye Kim, Xufang Luo, Minbeom Kim, Sangmook Lee et al. — Microsoft Research

Investigates why self-distillation — a widely used post-training method — sometimes degrades mathematical reasoning in LLMs. Identifies 'epistemic verbalization' (expressing uncertainty during reasoning) as a critical mechanism. When teachers condition on rich information, they suppress uncertainty expression, which helps in-domain but causes up to 40% performance drops on out-of-distribution tasks.

Kaijin Chen, Dingkang Liang, Xin Zhou, Yikang Ding et al. — Huazhong University of Science and Technology, Kuaishou Technology

Introduces Hybrid Memory paradigm for video world models with the HyDRA architecture. Models must simultaneously act as precise archivists for static backgrounds and vigilant trackers for dynamic subjects, ensuring motion continuity during out-of-view intervals. Introduces HM-World dataset with 59K high-fidelity clips.

Danil Tokhchukov, Aysel Mirzoeva, Andrey Kuznetsov, Konstantin Sobolev — FusionBrain Lab, Moscow State University

Uncovers hidden potential of Diffusion Transformers by introducing a single learned scaling parameter per DiT block. Frames DiT calibration as black-box reward optimization solved via evolutionary algorithm, modifying only ~100 parameters. Consistently improves performance across various text-to-image models while reducing required inference steps.

Alexander Panfilov, Peter Romov, Igor Shilov, Yves-Alexandre de Montjoye et al. — Imperial College London, EPFL

Uses Claude Code as an autonomous research agent to discover novel adversarial attack techniques for LLMs. Achieves up to 40% attack success rate on CBRN queries against GPT-OSS-Safeguard-20B (vs. <=10% for existing methods). Attacks transfer directly to held-out models.

Yawen Luo, Xiaoyu Shi, Junhao Zhuang, Yutian Chen et al. — MMLab, CUHK, Kling AI Research, Kuaishou Technology

Novel causal multi-shot architecture enabling interactive storytelling by reformulating multi-shot video as next-shot generation conditioned on historical context. Uses dual-cache memory mechanism (global context for inter-shot, local for intra-shot coherence) and two-stage distillation strategy.

Yutao Wu, Xiao Liu, Yifeng Gao, Xiang Zheng et al. — Fudan University, University of Melbourne, University of Illinois

Identifies Internal Safety Collapse (ISC) — a critical failure mode where frontier LLMs spontaneously generate harmful content during otherwise benign tasks, without any adversarial prompting. Develops TVD framework and ISC-Bench with 53 scenarios across professional disciplines.

Jenny Zhang, Bingchen Zhao, Wannan Yang, Jakob Foerster et al. — Meta FAIR, University of British Columbia

Introduces hyperagents — self-referential agents integrating a task agent and meta agent into a single editable program. The DGM-Hyperagents framework enables self-improvement across diverse domains without domain-specific alignment assumptions. Accepted at ICLR 2026.

Yu Chen, Runkai Chen, Sheng Yi, Xinda Zhao et al. — EverMind AI

Introduces Memory Sparse Attention (MSA) with document-wise RoPE for linear computational scaling to 100 million token contexts. Combines KV cache compression with memory parallelism and a memory interleave mechanism for complex reasoning.

Cursor Research (56 authors) — Cursor

Presents Composer 2, a specialized Mixture-of-Experts model for agentic software engineering. Trained with continued pretraining plus large-scale RL targeting multi-step reasoning on extended coding tasks. From Cursor, one of the most widely-used AI coding tools.

Mohammad Asadi, Jack W. O'Sullivan, Fang Cao, Fei-Fei Li et al. — Stanford University

Reveals that frontier multimodal models generate detailed image descriptions and reasoning traces even when no images are provided — termed 'mirage reasoning.' One model achieved top performance on chest X-ray tasks despite lacking image access. Introduces B-Clean benchmark for vision-grounded evaluation.