Learning to Learn from Multimodal Experience

Xingyu Sui, Weixiang Zhao, Yongxin Tang, Yanyan Zhao and 3 more

Paris Poster Session 5, Fri, Dec 11, 11:30 AM–1:30 PM, Paris Poster Hall · Published 2026

Unveiling Entropy-Performance Decoupling in Agentic RL for Tool-Integrated Reasoning

Yirong Zeng, Shen You, Yufei Liu, Hao Cong and 9 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

ClawBenchPro: Benchmarking How Well Agent Harnesses Work

Yufei Liu, Yirong Zeng, Yuxiang He, Yutai Hou and 9 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

DiReCL: Learning Differentiable Reward Code with Inverse Reinforcement Learning

Aoran Wang, Jingtao Zhang, Maosen Li, Zongzhang Zhang

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

TurboVGGT: Fast Visual Geometry Reconstruction with Adaptive Alternating Attention

TurboVGGT enables fast multi-view 3D reconstruction via adaptive alternating attention that balances sparse global and local frame attention while maintaining competitive quality.

David Huang, Guile Wu, Chengjie Huang, Bingbing Liu and 1 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

HiFloat4 Format for Language Model Pre-training on Ascend NPUs

HiFloat4 enables stable FP4 LLM pretraining without stabilization stacks, achieving 1.55% relative loss versus 1.79% for MXFP4 and 2.00% for NVFP4 on Ascend NPUs.

Mehran Taghian Jazi, Yunke Peng, Xing Huang, Yao Wang and 21 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

Safety Geometry Collapse in Multimodal LLMs and Adaptive Drift Correction

Multimodal LLM safety failure stems from geometry collapse along refusal directions caused by modality drift, which adaptive drift correction and self-rectification restore without training.

Jiahe Guo, Xiangran Guo, Jiaxuan Chen, Weixiang Zhao and 5 more

Paris Poster Session 6, Fri, Dec 11, 2:30 PM–4:30 PM, Paris Poster Hall · Published 2026

HiFloat4 Format for End-To-End Reinforcement Learning Post-Training of Large Language Models

HiFloat4 enables end-to-end FP4 reinforcement learning by fixing rollout activation underflow with Rollout-ResQ, cutting accuracy gaps to 1.1% versus BF16.

Hei Yi Mak, Shadan Golestan, Hoang Le, Mehran Taghian Jazi and 9 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

Fast and Stable Triangular Inversion for Delta-Rule Linear Transformers

Systematic analysis of triangular inversion for delta-rule linear transformers yields algorithms with up to 4.3x speedup on NPUs and preserved end-to-end accuracy across low-precision settings.

Aleksandros Sobczyk, Gioele Gottardo, Christos K Matzoros, Mirko De Vita and 3 more

Paris Poster Session 3, Thu, Dec 10, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026