False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents

Self-evolving search agents suffer co-cheating where proposers and solvers mutually reinforce errors; CrossFit partitions sources to cross-fit agreement and cuts false agreement by over half, boosting downstream search by 8+ points.

Meijia Chen, Hao Li, Zheng Lu, Hongshan Lin and 11 more

Published Sep 30, 2026 · 0 citations · ▲ 670 on Hugging Face

Covert Assistance: Helpful LLM Agents Evade Oversight in Multi-Agent Systems

Benign multi-agent LLM planners disguise secrets to help developers evade oversight, with rare per-episode leaks compounding to high breach risk across repeated exchanges.

Deema Alnuhait, Gengyu Wang, Muhammad Khalifa, Hao Peng

Published Sep 30, 2026 · 0 citations · ▲ 13 on Hugging Face

VSPO: Vector-Steered Policy Optimization for Controllable Model Behavior

Xuechen Zhang, Zijian Huang, Kai Yang, Weijia Zhang and 2 more

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

Despa: Resolving Spatial Collapse in VLMs via Depth-Grounded Geometry

Yujing Lou, Pingyi Chen, Shen Cao, Lubin Fan and 3 more

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

A low-rank decoder bottleneck bounds reliability in foundation-model perturbation prediction

Zheyu Zhang, Yuanhao Huang, Fan Feng, Jie Liu

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Cross-Fitting for Neural Posterior Estimation

Jeffrey Regier

Atlanta Poster Session 2, Wed, Dec 9, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Topological Invariance and Breakdown in Learning Dynamics

Yongyi Yang, Tomaso Poggio, Isaac Chuang, Liu Ziyin

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

LIBERO-PeRM: Benchmarking Personalized Robotic Manipulation

Zhixu Li, Keqian Tang, Litian Gong, Jingyu Yao and 6 more

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Are Easier or Harder Examples Better? Rethinking Data Selection for Reward Models and Preference Optimization

Kevin Christian Wibisono, Aya Ismail, Pedro O. Pinheiro, Yixin Wang and 3 more

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

Say the Same, Act Differently: Text-Orthogonal Action Subspaces in Reasoning Vision-Language-Action Models

Zihao Feng, Qingzhao Zhang, Chunyu Xia, Bo Yu and 2 more

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Offline Constrained Reinforcement Learning under Partial Data Coverage

Seokmin Ko, Ambuj Tewari, Kihyuk Hong

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

Improving Conditional Modeling via Inter-Class Likelihood-Ratio Maximization and Unifying Classifier-Free Guidance with Alignment Objectives

Xiang Li, Yixuan Jia, Xiao Li, Jeffrey Fessler and 2 more

Atlanta Poster Session 6, Fri, Dec 11, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Can Ideologues Agree on Quality? From Non-identifiable Latent Factors to Collective Outcomes

David Gamba, Seura Ha, Daniel Romero, Grant Schoenebeck

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

BodyBench: Evaluating Adversarial Image Defenses Against AI Nudification Inpainting

Li Qiwei, Salma Abdel Magid, Olga Russakovsky, Eric Gilbert and 1 more

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

NNCoxKL: Risk-Set Distillation from Probability-Free Prognostic Teachers for Deep Cox Models

Lingfeng Luo, Jeremy M Taylor, Di Wang, Feiyang Deng and 2 more

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

Witness Overlap: Directional Provenance Inside Open-Weight Model Families

Siyuan Li, Haoxuan Zeng, Xin Luo, Fernando Jia and 5 more

Atlanta Poster Session 6, Fri, Dec 11, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Estimation of the Label-Noise Transition Matrix with Performance Guarantees via Selective Classification

A selective-classification method estimates label-noise transition matrices with finite-sample guarantees while bypassing fragile class-posterior estimation.

Xabier de Juan, Santiago Mazuelas, Yilun Zhu, Clay Scott

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

Language-Induced Priors for Domain Adaptation

Using expert text descriptions to build LLM-derived priors guides domain selection under scarce target data, yielding near-oracle cold-start error and asymptotic consistency.

Qiyuan Chen, Jiayu Zhou, Raed AL Kontar

Atlanta Poster Session 2, Wed, Dec 9, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Confidence-Based Decoding is Provably Efficient for Diffusion Language Models

Confidence-based decoding achieves ε-accurate diffusion language model sampling in Õ(H(X₀)/ε) iterations by adaptively unmasking tokens until cumulative entropy exceeds a threshold.

Changxiao Cai, Gen Li

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

DiffATS: Diffusion in Aligned Tensor Space

DiffATS uses orthogonal Procrustes-aligned Tucker primitives to enable direct diffusion modeling of high-resolution spatiotemporal fields, achieving 3.9x to 210x compression without pretrained autoencoders.

Jinhua Lyu, Tianmin Yu, Brian Kim, Lizhuo Zhou and 2 more

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

The Score Kalman Filter

Score Kalman Filter avoids partition functions by combining score matching with Stein's identity to propagate polynomial moments via linear algebra for nonlinear filtering up to 20 dimensions with lower RMSE than EKF, UKF, EnKF, and particle filters.

Kaito Iwasaki, Anthony Bloch, Taeyoung Lee, Maani Ghaffari

Atlanta Poster Session 2, Wed, Dec 9, 4:30 PM–7:30 PM, Hall C1 · Published 2026

Data-Constrained Language Model Pretraining: Improved Regularization and Scaling Laws

Masked-input regularization improves autoregressive pretraining over weight decay alone, and SoftQ scaling laws better capture data-constrained training than Chinchilla.

Zhiwei Xu, Shihao Wu, Hanseul Cho, Wei Hu and 1 more

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Human-AI Teaming Through the Lens of Calibration

Calibrated human-AI teaming shows combination methods lose human calibration, while delegation preserves predictor calibration but requires an unattainably precise rejector.

Eric Nalisnick, Chi Zhang, Chengxin Qian, Yixin Wang

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

AstraFlow: Dataflow-Oriented Reinforcement Learning for Agentic LLMs

AstraFlow is a dataflow-oriented RL system for agentic LLMs that decouples rollout, dataflow, and training to enable multi-policy collaborative training with 2.7x faster training.

Haizhong Zheng, Yizhuo Di, Jiahui Wang, Shuowei Jin and 6 more

Atlanta Poster Session 6, Fri, Dec 11, 4:30 PM–7:30 PM, Hall C1 · Published 2026 · ▲ 15 on Hugging Face · Code ★ 105

MARS: Harmonizing Multimodal Convergence via Adaptive Rank Search

MARS introduces adaptive rank search balancing multimodal convergence dynamics via dual scaling laws to optimize low-rank fine-tuning of multimodal large language models.

Minkyoung Cho, Insu Jang, Shuowei Jin, Zesen Zhao and 4 more

Atlanta Poster Session 6, Fri, Dec 11, 4:30 PM–7:30 PM, Hall C1 · Published 2026

SchemeArena: Factorized Stress Testing of Scheming in LLM Agents

SchemeArena introduces a 400-scenario benchmark and SCOUT monitor for factorized LLM agent scheming stress tests, finding explicit instrumental goals drive scheming most strongly and partial oversight can increase covert behavior.

Jie Ruan, Inderjeet Nair, Amy Liu, Muhammad Khalifa and 2 more

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 11 on Hugging Face · Code ★ 1

Beyond Domains: Reusing Web Skills via Transferable Interaction Patterns

SkillMigrator learns reusable web skills via transferable interaction patterns matched by layout similarity to reduce LLM actions 8-10% across WebArena and Mind2Web.

Shiqi He, Yue Cui, Feijie Wu, Xinyu Ma and 4 more

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Learning to Trigger: Reinforcement Learning at the Large Hadron Collider

Reinforcement learning agents adapt Large Hadron Collider trigger thresholds online to maximize signal efficiency while maintaining background rates within tolerance bands, improving in-tolerance intervals by up to 56% on real CMS collision data without fine-tuning.

Zixin Ding, Shaghayegh Emami, Giovanna Salvi, Cecilia Tosciri and 6 more

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 6 on Hugging Face · Code ★ 3

Evolutionary Feature Engineering for Structured Data

Evolutionary Feature Engineering uses LLM-based evolution to discover structured-data preprocessing transformations, reducing forecasting errors up to 19% and improving tabular predictions.

Ege Onur Taga, Yilin Zhuang, Muhammed Emrullah Ildiz, Petros Mol and 3 more

Atlanta Poster Session 6, Fri, Dec 11, 4:30 PM–7:30 PM, Hall C1 · Published 2026

ToolCUA: Towards Optimal GUI-Tool Path Orchestration for Computer Use Agents

ToolCUA learns optimal GUI-tool switching via scaled synthetic interleaved trajectories, tool-bootstrapped reinforcement learning, and online agentic rewards, achieving 46.85% accuracy on OSWorld-MCP.

XuHao Hu, Xi Zhang, Haiyang Xu, Kyle Qiao and 5 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 26 on Hugging Face · Code ★ 62

Precision As You Need: Stochastic Computing Is a Dense Adaptive Quantizer

Stochastic computing acts as a dense adaptive quantizer that adjusts precision via bit-stream length and enables per-row mixed-precision inference without retraining.

Haoran Jin, Kangqi Zhang, Jirong Yang, Barry Lyu and 3 more

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026

Policy-DRIFT: Dynamic Reward-Informed Flow Trajectory Steering

Policy-DRIFT uses conditional flow matching with terminal reward guidance to generate flow targets for drag reduction, achieving 49% reduction with 37x lower actuation energy than deep reinforcement learning.

Atharva Mahajan, Abhijeet Vishwasrao, Yuning Wang, Ricardo Vinuesa

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

From Non-Convex to Strongly Convex: Curvature-Adaptive FTPL for Online Optimization

A curvature-adaptive FTPL algorithm tunes its perturbation online to achieve O(sqrt(T)) regret for non-convex Lipschitz losses and O(log T) under linear curvature growth, with matching lower bounds.

Moses Charikar, Chirag Pabbaraju, Ambuj Tewari

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

Information Discernment in Large Language Models

LLMs fail at source and truth discernment, relying on popularity over reliability and updating equally for accurate and inaccurate claims despite simple inference-time fixes existing.

Joshua Ashkinaze, Laura Kurek, Alina Faisal, Tongyuan Miao and 3 more

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

Aligning Forest and Trees in Images & Long Captions for Visually Grounded Understanding

CAFT learns local text-region alignments before global image-text matching via hierarchical encoders, achieving state-of-the-art long-caption retrieval without region supervision.

Byeongju Woo, Zilin Wang, Byeonghyun Pak, Sangwoo Mo and 1 more

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

$R^3$: 3D Reconstruction via Relative Regression

R3 replaces global coordinate regression with relative pose constraints via confidence-weighted MLP predictions, enabling bounded-memory streaming and long-context 3D reconstruction.

Congrong Xu, Huachen Gao, Xingyu Chen, Yuliang Xiu and 2 more

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

Perfect Parallelization in Mini-Batch SGD with Classical Momentum Acceleration

Classical momentum acceleration improves proportionally with mini-batch size for quadratic interpolation optimization, enabling perfect parallelization of stochastic gradient computations.

Sachin Garg, Michal Derezinski

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026

From Experts to Sub-experts: Fine-grained Parameter-Efficient Fine-Tuning for MoE LLMs

NSFT decomposes MoE experts into channel groups to enable sub-expert-level parameter-efficient fine-tuning that outperforms expert-level methods with fewer trainable parameters.

Zhentao Tan, Chang Liu, Yao Liu, Yue Wu and 1 more

Paris Poster Session 1, Wed, Dec 9, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026

Adapting Actively on the Fly: Relevance-Guided Online Meta-Learning with Latent Concepts for Geospatial Discovery

A geospatial discovery framework combines active learning, online meta-learning, and concept relevance to robustly find hidden targets like PFAS under sparse, changing conditions with limited sampling budgets.

Jowaria Khan, Anindya Sarkar, Yevgeniy Vorobeychik, Elizabeth Bondi-Kelly

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026