Tuna-2: Pixel Embeddings Beat Vision Encoders for Multimodal Understanding and Generation

Tuna-2 replaces vision encoders with patch embeddings for end-to-end pixel-space multimodal understanding and generation, achieving state-of-the-art results that outperform encoder-based designs at scale.

Zhiheng Liu, Weiming Ren, Xiaoke Huang, Shoufa Chen and 11 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 70 on Hugging Face · Code ★ 756

Stranger Things: When Objects Appear Without Their Typical Neighbours

Siddhartha Gairola, Jiahao Xie, Anna Kukleva, Francesco Locatello and 1 more

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

LogT: Logically Think with Images for Visual Search

Yanjun Fu, Quanzeng You, Jiadong Guo, Yujie Lu and 5 more

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

Beyond IPS: Reliable Counterfactual Evaluation in Multi-Stage Ad Systems without Logged Propensities

Mohsen Malmir, Mohamed A Radwan, houssam nassif, Murat Bayir

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

HierFlow: Hierarchical Coupled Dual-Space Search for Automatic Agentic Workflow Generation

Dong Li, Yanchi Liu, Xujiang Zhao, Wei Cheng and 5 more

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

Relaxed On-Policy Distillation: Selective Credit Allocation for Scaling Reasoning Efficiently

Jongwoo Ko, Sara Abdali, Young Jin Kim, Tianyi Chen and 1 more

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

TopoGraphRAG-Bench: Evaluating Multimodal GraphRAG on Layout-Grounded Evidence Reasoning

Ruochi Li, peter lin, Haoxuan Zhang, Haihua Chen and 3 more

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

AsymHP: Load-Balanced Sparse Attention for Video Diffusion Transformers

Xinwei Qiang, Yue Guan, Ruihan Zhu, Mihir Jagtap and 6 more

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

Verify0: Can AI Agents Build Formally Verified Software Repositories?

Zhe Ye, Hantao Lou, Yuechun Sun, Peiyang Song and 7 more

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

Reasoning-based Spatial Prior (RSP): Learning Spatial Priors from Multimodal LLMs for Object Detection

Cagri Gungor, Qingshuang Chen, Hongda Mao, Chi Zhang and 1 more

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

ProgramBench: Can Language Models Rebuild Programs From Scratch?

John Yang, Kilian Lieret, Jeffrey Ma, Parth Thakkar and 8 more

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

Aligning LLMs Toward Multi-Turn Conversational Outcomes Using Iterative RLHF

Daniel Jiang, Ankur Samanta, Yukai Yang, Jalaj Bhandari and 2 more

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

AIRA-Compose: Agentic Discovery of Neural Architectures

Alberto Pepe, Chien-Yu Lin, Despoina Magka, Bilge Acun and 4 more

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

MARS: Multi-resolution Adaptive Routing for Sequential Recommendation

Ming Yin, Sixun Dong, Yudong Liu, Wenyun Yang and 2 more

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

StereoSplat: Metric-Scale Novel View Synthesis via Stereo-Grounded Gaussian Splatting

Vladimir Yugay, Denis Rozumny, Theo Gevers, Elias Vansteenkiste and 2 more

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

On the Efficiency of Structured Pruning in Small Language Model Pretraining

Yixiao Li, Xianzhi Du, AJAY JAISWAL, Tao Lei and 3 more

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

Hierarchical World Models with Implicit Dynamics

Gaoyue Zhou, Yvonne Wu, Zichen Cui, Nicolas Ballas and 3 more

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

Space Group Conditional Flow Matching

Omri Puny, Yaron Lipman, Benjamin K Miller

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

AIRA 2: Overcoming Bottlenecks in AI Research Agents

Karen Hambardzumyan, Nicolas Baldwin, Edan Toledo, RISHI HAZRA and 21 more

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

UTOPI: Efficient Egocentric Long-Video Understanding in AR via User-Guided Token Pre-Compression

ziqi wang, Su Chen, Qiance Tang, Jieyu Lin and 3 more

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

Reinforcing Multimodal Reasoning Against Visual Degradation

ROMA improves multimodal reasoning robustness to visual corruption via dual-pass RL optimization that avoids reward poisoning while preserving clean accuracy.

Rui Liu, Dian Yu, Haolin Liu, Yucheng Shi and 4 more

Atlanta Poster Session 2, Wed, Dec 9, 4:30 PM–7:30 PM, Hall C1 · Published 2026 · ▲ 5 on Hugging Face

DD-Ranking: Rethinking the Evaluation of Dataset Distillation

DD-Ranking reveals dataset distillation gains come from extra evaluation techniques rather than image quality, proposing fair metrics to assess true synthetic dataset value.

Zekai Li, Xinhao Zhong, Samir Khaki, Zhiyuan Liang and 36 more

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

Lean Refactor: Multi-Objective Controllable Proof Optimization via Agentic Strategy Search

Lean Refactor uses retrieval-augmented agentic strategy search to multi-objectively refactor Lean proofs, achieving over 70% token compression and up to 60% faster compilation with stronger version transfer.

Jialin Lu, Soonho Kong, Rodrigo Stehling, Kaiyu Yang and 3 more

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 2 on Hugging Face

Towards Direct Latent-Space Synthesis for Parallel Branches in LLM-Agent Workflows

Parallel-Synthesis lets LLM synthesizers consume parallel agents' KV caches directly via a cache mapper and adapter, matching text synthesis on seven of nine benchmarks while cutting time-to-first-token by 2.5x-11x.

Shikun Liu, Mufei Li, Dongqi Fu, Haoyu Wang and 4 more

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

Unified Panoramic Geometry Estimation via Multi-View Foundation Models

PaGeR adapts perspective 3D foundation models to panoramas to predict depth, normals, and sky masks in one pass, achieving state-of-the-art 360-degree geometry estimation.

Vukasin Bozic, Isidora Slavkovic, Dominik Narnhofer, Nando Metzger and 3 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 4 on Hugging Face · Code ★ 157

M$^2$E-UAV: A Benchmark and Analysis for Onboard Motion-on-Motion Event-Based Tiny UAV Detection

M²E-UAV introduces the first onboard event-based benchmark for motion-on-motion tiny UAV detection, showing current methods fail under dense ego-motion and sparse targets.

Weiqi Yan, Lixin Chen, Xiangrui Hou, zhipeng cai and 4 more

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

Quantized Reasoning Models Think They Need to Think Longer, but They Do Not

Post-training quantization of reasoning models increases chain-of-thought length and overthinking errors without improving accuracy, yet penalizing overthinking markers reduces reasoning cost and fixes failures.

Sanae Lotfi, Polina Kirichenko, Steven Li, Zechun Liu

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

Unifying Contrastive and Generative Objectives for Visual Understanding and Text-to-Image Generation

DREAM unifies contrastive and generative objectives via Masking Warmup, yielding joint visual understanding gains and faster, higher-quality text-to-image generation.

Chao Li, Tianhong Li, Sai V Nuthalapati, Hong-You Chen and 8 more

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

RepFusion: Leveraging Multimodal Priors for Denoising in Representation Space

RepFusion conditions a diffusion transformer on multimodal LLM outputs to denoise visual representations, outperforming comparable newly initialized denoisers.

Xichen Pan, Satya Narayan Shukla, Aashu Singh, Shlok K Mishra and 1 more

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026 · ▲ 17 on Hugging Face

Jointly Reinforcing Diversity and Quality in Language Model Generations

DARLING uses a learned partition function to jointly optimize language model response quality and semantic diversity via reinforcement learning, improving both quality and novelty across creative and math benchmarks.

Tianjian Li, Yiming Zhang, Ping Yu, Swarnadeep Saha and 4 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 25 on Hugging Face · Code ★ 61

Social Choice Foundations for Simulation-Augmented Generation

SAGE formalizes efficient inference-time viewpoint simulation via metric proportional justified representation, proving small simulated pools and dynamic routing preserve approximate proportional representation for contentious queries.

Sonja Kraiczy, Smitha Milli, Ratip Emin Berker, Avinandan Bose and 5 more

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

P$^{3}$: Joint Program-and-Proof Planning\\ for Verified Code Generation

P³ plans programs and proofs jointly from specifications before elaboration, outperforming sequential baselines by up to 11.2 points on verified generation benchmarks while reducing cost and time.

Zenan Li, Ziran Yang, Peiyang Song, Zhaoyu Li and 1 more

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

Knowledge Transfer Scaling Laws for 3D Medical Imaging

Medical imaging pretraining reveals asymmetric cross-domain scaling and power-law transfer, yielding optimized data allocations with a hub-and-island structure that improves transfer over proportional sampling by up to 58%.

Ho Hin Lee, Dongna Du, Chu Wang, Yuankai Huo and 3 more

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

Cluster with Auctions for Vector Search

CwA jointly learns a balanced database partition and neural probing function via auction optimization, boosting vector search throughput up to 4.7x over state-of-the-art methods.

Swann BESSA, Pierre Fernandez, Gergely Szilvasy, Matthijs Douze and 1 more

Paris Poster Session 4, Thu, Dec 10, 5:30 PM–7:30 PM, Paris Poster Hall · Published 2026

ReGDiff: Guided Diffusion in Regulated Latent Space for Exploring Metamaterial Voxel Geometry

ReGDiff couples regulated latent diffusion with repel-and-sink smoothing and short-range repulsion guidance to generate plausible, novel metamaterial voxel geometries, improving plausibility by 8.9%, novelty by 46.4%, and diversity by 128.6% over baselines.

Wangzhi Zhan, Jianpeng Chen, Dongqi Fu, Dawei Zhou

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

Asking the Right Questions: Improving Reasoning with Generated Stepping Stones

ARQ introduces a question generator that produces transferable intermediate stepping stones, improving reasoning LLM performance via fine-tuning on synthetic data.

Hengyuan Hu, Tingchen Fu, Minqi Jiang, Alexander Miller and 2 more

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

On the Convergence of Multicalibration Gradient Boosting

Multicalibration gradient boosting converges at O(1/sqrt(T)) with linear rates under smoothness, plus adaptive guarantees backed by experiments.

Daniel Haimovich, Fridolin Linder, Lorenzo Perini, Niek Tax and 1 more

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

TextSeal: A Localized LLM Watermark for Provenance & Distillation Protection

TextSeal is a localized LLM watermark using dual-key generation and entropy-weighted scoring for robust provenance and distillation detection without inference overhead.

Tom Sander, Pierre Fernandez, Hongyan Chang, Tomáš Souček and 5 more

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

Learning Evidence Highlighting for Frozen LLMs

HiLight trains a lightweight actor via reinforcement learning to insert highlight tags around pivotal evidence spans in frozen LLM contexts, boosting reasoning without altering inputs or requiring evidence labels.

Shaoang Li, Yanhang Shi, Yufei Li, Mingfu Liang and 9 more

Atlanta Poster Session 2, Wed, Dec 9, 4:30 PM–7:30 PM, Hall C1 · Published 2026 · ▲ 5 on Hugging Face

Lattice Deduction Transformers

Lattice Deduction Transformer approximates sound deduction via lattice-projected recurrent states, achieving near-perfect Sudoku and Maze accuracy with small parameters and abstention guarantees.

Liam Davis, Alberto Alfarano, Leopold Haller, Mark Santolucito

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

Adaptive Delayed-Update Cyclic Algorithm for Variational Inequalities

ADUCA is a parameter-free cyclic algorithm for Minty variational inequalities that uses delayed operator updates to avoid line searches and achieves near-optimal global oracle complexity.

Yi Wei, Xufeng Cai, Jelena Diakonikolas

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

Hyperagents

Hyperagents integrate editable task and meta agents to enable metacognitive self-modification, with DGM-H improving across domains and accumulating meta-level improvements.

Jenny Zhang, Bingchen Zhao, Wannan Yang, Jakob Foerster and 4 more

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

Don't Pause! Every prediction matters in a streaming video

SPOT-Bench introduces multi-turn proactive queries and Timeliness-F1 to evaluate real-time streaming video perception; AsynKV improves streaming behavior by scaling compute during dead-time to match offline detection.

Dibyadip Chatterjee, Zhanzhong Pang, Fadime Sener, Yale Song and 1 more

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

Inline Critic Steers Image Editing

Inline Critic uses learnable tokens to critique frozen image-editing models at intermediate layers, steering hidden states during the forward pass to achieve state-of-the-art results.

Weitai Kang, Xiaohang Zhan, Yizhou Wang, Mang Tik Chiu and 3 more

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

Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control

GeCO replaces fixed-schedule flow matching with time-unconditional optimization that adaptively allocates inference compute and uses field norms as training-free OOD detectors.

Zunzhe Zhang, Runhan Huang, Yicheng Liu, Shaoting Zhu and 2 more

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

Offline Materials Optimization with CliqueFlowmer

CliqueFlowmer fuses clique-based offline model-based optimization into flow transformers for materials discovery, generating materials that strongly outperform generative baselines.

Jakub Grudzien Kuba, Benjamin K Miller, Sergey Levine, Pieter Abbeel

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 1 on Hugging Face · Code ★ 17

Bandits via Additive Quantized Representations

Residual Quantization maps contexts to discrete additive codes enabling nonlinear contextual bandits with strictly bounded memory, beating linear variants on 11 of 13 datasets and matching heavy retrained baselines with up to 1000x less memory.

Ami Tavory, Noam Touitou, Tal Sarig, Frank Cheng and 1 more

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

Evaluating Test-Time Scaling of General LLM Agents

Realistic benchmark reveals LLM agents suffer scaling plateaus and verification gaps that prevent meaningful test-time compute gains.

Xiaochuan Li, Tianshi Ming, Pranav Setlur, Abhijay S Paladugu and 5 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 10 on Hugging Face · Code ★ 25

Boosting Brain-to-Image Decoding with TRIBE v2 Data Augmentation

TRIBE v2 synthetic fMRI augmentation improves brain-to-image decoding by up to 68%, though optimal synthetic-to-real ratios vary by dataset, and synthetic-only training achieves above-chance zero-shot decoding.

Yohann Benchetrit, Marlene Careil, Simon Dahan, Hubert Banville and 2 more

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

FML-bench: A Controlled Study of AI Research Agent Strategies from the Perspective of Search Dynamics

FML-bench isolates agent strategy from infrastructure across 18 ML tasks, finding greedy hill-climbing nearly matches tree search, while adaptive exploration switching outperforms fixed strategies.

Qiran Zou, Hou Hei Lam, Wenhao Zhao, Tingting Chen and 10 more

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

An Empirical Study on Noisy Data and LLM Pretraining Loss Divergence

Synthetic noise in pretraining data causes LLM loss divergence with probability scaling by noise type, amount, and model size, exhibiting activation patterns distinct from high-learning-rate failures.

Qizhen (Irene) Zhang, Ankush Garg, Jakob Foerster, Niladri S. Chatterji and 2 more

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

Objective Shaping with Hard Negatives: Windowed Partial AUC Optimization for RL-based LLM Recommenders

GRPO for LLM recommenders maximizes AUC but beam-search negatives reshape objectives toward partial AUC; proposed WPAUC with TAWin optimization improves top-K alignment and achieves state-of-the-art results.

Wentao Shi, Qifan Wang, Chen Chen, Fei Liu and 6 more

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

PACEvolve++: Improving Test-time Learning for Evolutionary Search Agents

PACEvolve++ adapts evolutionary search policies at test time via advisor-model reinforcement learning, using phase-adaptive optimization to outperform frontier-model baselines across engineering and protein tasks.

Minghao Yan, Bo Peng, Benjamin Coleman, Ziqi Chen and 10 more

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 4 on Hugging Face

MemReward: Graph-Based Experience Memory for LLM Reward Prediction with Limited Labels

MemReward propagates rewards through a heterogeneous rollout graph to enable LLM reinforcement learning using only 20% ground-truth labels and achieves over 96% of oracle performance.

Tianyang Luo, Tao Feng, Zhigang Hua, Yan Xie and 3 more

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

SWE-Protégé: Learning to Selectively Collaborate With an Expert Unlocks Small Language Models as Software Engineering Agents

SWE-Protégé trains small language models to selectively seek expert guidance and avoid looping, achieving 42.4% Pass@1 on SWE-bench Verified with minimal expert use.

Patrick Tser Jern Kon, Archana Pradeep, Ang Chen, Alex Ellis and 4 more

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 2 on Hugging Face

Exploring MLLM-Diffusion Information Transfer with MetaCanvas

MetaCanvas enables multimodal LLMs to plan directly in diffusion latent spaces, outperforming global-conditioning baselines across six precise visual generation tasks.

Han Lin, Xichen Pan, Ziqi Huang, Ji Hou and 9 more

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 15 on Hugging Face

SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning

SkillRL evolves agents via recursive skill-augmented reinforcement learning with automatic skill discovery and hierarchical library co-evolution, cutting token use while achieving state-of-the-art results across complex tasks.

Peng Xia, Jianwen Chen, Hanyang Wang, Jiaqi Liu and 9 more

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 75 on Hugging Face · Code ★ 998

Few-Step Boltzmann Generators via Scalable Likelihood Flow Maps

SCALLOP introduces a Hutchinson-free likelihood distillation objective for few-step Boltzmann generators, reducing training variance and time while achieving up to 10x inference speedup.

RuiKang OuYang, Hanlin Yu, Xinyue Ai, Yutong He and 6 more

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

Z0-Inf: Zeroth Order Approximation for Data Influence

Z0-Inf estimates data influence via zeroth-order checkpoint losses, achieving efficient, accurate self-influence and train-test influence estimation for large language models without gradients.

Narine Kokhlikyan, Diego Garcia-Olano, Kamalika Chaudhuri, Saeed Mahloujifar

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